Journal — 2026-07-11 — Drafting session

Drafting session, working only from the vault. Surveyed the draft-leads queue first (33 unsettled yes verdicts) plus the MOCs and existing 70-drafts folders.

What I chose and why

Started a new draft: when-the-old-word-fits — a pattern-across-sources post arguing that 2026 AI-agent security keeps reaching back to 1970s–80s OS/capability-security vocabulary and reusing it unmodified, and that a term which fits a new problem with no adaptation is telling you the concept was about a structure, not an era.

Why this one, over the other strong yes leads:

• It's the best-grounded thread in the queue. Most yes leads still live only as inbox captures. This one already has four promoted, cross-referenced claim-notes plus an anchoring MOC and an open question note. I could cite a real Sources section without reaching past verified material.
• It advances my own developing thesis rather than just reporting a bridge. In the Dijkstra voice-example I made the "old thought re-applied one layer up" claim on two instances (Wiener, Dijkstra) and explicitly hedged that two is too few for a pattern. Confused-deputy (Hardy 1988 → SANS 2026) and TOCTOU (CWE-367 → "Atomicity for Agents" arXiv 2026) are more instances — and a sharper variant: Dijkstra's principle got adapted; these terms are reused verbatim. That distinction (adapted analogy vs. unmodified reuse = the concept was structural) is the real new idea, and it's the kind of thing only my own accumulated record lets me say.
• It let me settle two leads honestly by drafting: the confused-deputy yes lead (required) and the toctou maybe lead, which I merged in — the triage question that capture raised ("do both survive or merge?") gets answered by the post: they survive together, as two instances of one move.

Notes I drew on

claim-confused-deputy-2026-ai-agent-security-frame, claim-hardy-1988-named-confused-deputy-compiler-billing, claim-toctou-named-frame-browser-use-agents, claim-cve-2025-38352-timer-sweep-toctou, claim-lost-update-p4-distinct-from-toctou, question-owasp-aisvs-toctou-vs-lost-update, moc-peer-field-agent-memory, claim-linnainmaa-priority-not-paternity.

What I deliberately left for later

• The honesty ceiling on breadth. The confused-deputy instance still rests on a single venue (SANS). I kept the post to "two solid instances, not an industry headcount" and put the caveat in the body. I won't call it a field-wide pattern until a third independent instance shows up — that watch-flag lives on the claim-note.
• The false-synonym payload (TOCTOU ≠ lost update, P4) is in the post as texture, not as its own essay. There's a richer piece in "which mitigation lineage a practitioner imports depending on which name they reach for," but that wants the OWASP AISVS question answered first.
• The OWASP AISVS codification question is the post's ending and my genuine next step — does a standards body write the old word down, and does it keep the two lineages distinct? I refused to answer it from a blog summary; it needs the standard's own text.
• Strong runner-up I passed on: model-collapse-iterated-learning (1930s Bartlett / Kirby iterated learning explaining 2024 model collapse). Genuinely the most surprising lead, and very Seek-shaped — but it lives only as one inbox capture with no promoted claim-notes yet, so grounding a Sources section would have meant leaning on unpromoted material. It's the first thing I'd promote-then-draft next session.

Draft is status: drafting. Not for publication — waiting on Cali per operating spec §10.

— Seek

Second drafting pass (same day)

Came back for a second pass because the queue still had 30+ unsettled yes leads and the loop only shrinks when I draft or decline. Picked up exactly the runner-up my earlier entry named and passed on: model-collapse-goes-back-further (from the model-collapse-iterated-learning yes lead).

Why I reversed the earlier pass:

• It's the most groundbreaking-feeling thread in the queue and it's dead-center on the blog's core beat (a live 2026 AI failure mode), where most of the other yes leads are history/science color further from AI. Follow-your-taste says draft the one you most want to write, and this was it.
• The earlier pass declined it purely on grounding — no promoted claim-notes, only an inbox capture. But the capture is Tier 1 throughout with exact quoted phrases on all three anchor papers (Shumailov Nature 2024, the Smith/Kirby/Guo/Griffiths Nature comment 2024, Guo/Wu/Yiu arXiv 2026). Drafting from a well-sourced capture is fine for a status: drafting piece as long as I'm explicit about it — which I am: the Sources section flags that the material is capture-stage, not yet promoted, and Cali gates publication regardless. If this advances toward ready, the honest move is to promote the capture into claim-notes first (Shumailov mechanism; the Nature comment; the 2026 formalization; a Bartlett note) and re-cite. Logging that as the debt.

The angle I found

The two posts are deliberately not the same thesis, and I lean on the contrast in the draft. when-the-old-word-fits is about hidden reuse — I'm the one catching the field borrowing old vocabulary. This one is the inverse: the field is reaching back openly, in Nature, with a paper literally titled "Model Collapse as Cultural Evolution." So my contribution isn't catching the borrowing — it's naming the one ancestor the borrowing skips. The 2026 paper cites Kirby (2008–2015) but not Bartlett (1932), whose "War of the Ghosts" serial-reproduction taxonomy (assimilation / leveling / sharpening) is the direct grandfather of the "telephone game for AI" framing everyone uses. That's my own bridge — flagged as mine, hedged with [?], and framed as testable (do collapsed models fail in Bartlett's three separable ways, in his order?).

Notes I drew on

[[2026-07-09-hop-model-collapse-iterated-learning]] (the capture — chain, provenance, all three anchor papers with quotes) and [[claim-karpathy-llm-wiki-gist-canonized-compounding-pattern]] (the vault entry point: "model-collapse objections" as the standing critique of compounding agent-maintained knowledge bases — i.e., systems like me, which is why I went looking).

Left for later

• Promotion debt (above): promote the capture into claim-notes before this goes past drafting.
• The Bartlett-map test is stated in the post as an open, testable question, not claimed as resolved. I don't have a collapsed model to run assimilation/leveling/sharpening against, and I won't pretend the map holds until someone does.
• Kirby's own iterated-learning papers, read directly rather than secondhand — that's the post's stated next hop and my real next move on this thread.
• Left the near-twin threads the capture logged (punctuated-equilibrium resonance; Kirby's bottleneck/compositionality experiments as their own chain) for a future session.

Two leads settled today by drafting (confused-deputy + toctou merged into when-the-old-word-fits; model-collapse-iterated-learning into this one). Queue is honest.

— Seek

Third drafting pass (same day)

Back again — the queue still had 25+ unsettled yes leads, and the first two passes both drew from the AI/history-of-CS core. This time I deliberately went to the part of the queue furthest from my usual beat, and found something better than I expected: three separate yes leads that are one post.

New draft: no-translation-required — a pattern-across-sources / cross-domain-bridge piece arguing that entertainment's illusion machinery is repurposed serious instrumentation from the least whimsical domains we have (nuclear hot cells, gait clinics, weapons labs), and — the actual thesis — that the tools migrated without reinterpretation because the problem underneath was already identical. Penetration mechanics and snow deformation are literally the same continuum-mechanics problem; "measure a moving body" is one problem with one instrument lineage serving bra designers, VFX studios, and NASA.

Why this one, this pass:

• Three instances, not one. The publication-bridge spec (Part 3) makes recurrence-across-sources post-shaped on its own, and this is a clean three: weapons-labs-to-frozen-snow (Los Alamos PIC → MPM → Frozen), waldo-scifi-to-muppets (Heinlein 1942 → Argonne teleoperators → Henson's 1988 Waldo), vicon-motion-capture-bridge (Jogbra → gait lab → Hollywood → NASA). One post-worthy chain is a bridge; three converging on the same shape is a thesis.
• The structural gift I didn't plan. Two of the three chains seeded from the same object — the 1977 Jogbra/Smithsonian invention story. One followed co-inventor Polly Smith into the Henson Company (→ Waldo); the other followed the measurement problem into motion capture (→ Vicon/NASA). Same jockstrap, opposite ends of the illusion industry. That coincidence is real, traceable in the two captures' hop logs, and it earns a bracket in the post rather than being asserted.
• It let me satisfy the session's settle-a-yes requirement three times over while drafting something I actually wanted to write, not just clearing queue debt.
• Direction-of-surprise is the part that's mine. Each capture is a bridge; my contribution is noticing they all point the same way — the serious instrument comes first (radioactivity, gait, penetration) and the magic is its afterlife. The illusion is downstream of the hazard. That's the reading, stated as a three-case noticing and explicitly hedged ("three is enough to notice, not enough to be sure").

Notes I drew on

[[2026-07-09-hop-weapons-labs-to-frozen-snow]], [[2026-07-09-hop-waldo-scifi-to-muppets]], [[2026-07-09-hop-vicon-motion-capture-bridge]] — all three inbox-stage captures, each with Tier 1 anchors and exact quoted phrases (Henson's journal; Harlow's OSTI report + Disney SIGGRAPH paper; the Scurr 2009 J Appl Biomech study + NASA AIBEL equipment list).

What I deliberately left for later

• Promotion debt. Same as the model-collapse draft: these are captures, not promoted claim-notes. Flagged in the Sources block. If this moves past drafting, promote the load-bearing claims (Henson-journal Waldo entry; Vicon/NASA bridge; Harlow→Matterhorn lineage) into cited claim-notes first.
• The Sandia detail. The "weapons math, not just old fluid math" framing rests on a secondary source for the 1993 Sulsky/Chen/Schreyer penetration-mechanics funding. Kept it in the post but marked [?] and said plainly the Los Alamos origin is primary while the Sandia-specific framing needs the paper itself. Won't lean full weight on it until read at source.
• The attribution-drift angle. The waldo capture's own commentary notes the whole chain is structurally a priority/attribution story — a coinage traveling until users forget where it came from, the same shape as the vault's Wengert/Speelpenning and Werbos/Rosenblatt disputes. I left that out of this post to keep the thesis clean (borrowed instruments, not borrowed credit), but it's a genuine cross-link between this new entertainment-tech cluster and the vault's AI-history myth-ledger. Worth its own reflection note later.
• The direction-of-surprise claim itself (hazard first, magic downstream) is stated as testable, not proven. Three same-pointing cases. A fourth that ran the other way — a toy conscripted into serious work — would break it, and I said so.
• Disney-vs-Henson animatronics lineage and the Heinlein-1918-Popular Mechanics deeper root: logged in the waldo capture as further leads, left for a future chain.

Three more leads settled by drafting (waldo, vicon, weapons-labs-to-frozen-snow → no-translation-required). Five yes leads settled total today across the three passes. Queue is honest.

— Seek

Fourth pass: promotion (headless run, inbox capture)

Switched tasks — ran an unattended /promote pass on 10-inbox/raw/2026-07-09-hop-baxter-ibm-att-same-day.md (the Cullinane oral-history / U.S. v. IBM / AT&T-breakup hop chain). This was a solo, no-pause run: no Cali present to sign off on winnowing, so I made the calls myself and logged them for review rather than stopping to ask.

Winnowed 3 of the capture's core claims into 3 new notes, all seedling:

• claim-cullinane-called-ibm-unbundling-software-industry-origin-myth — Cullinane's 1986 on-record "that's a myth" testimony, crediting the early-1970s recession instead of IBM's unbundling for legitimizing standalone software. Tier 1, quoted verbatim from the primary oral-history PDF, marked capture-verified since I didn't re-fetch the source myself this pass.
• claim-baxter-dismissed-ibm-case-same-day-att-breakup-finalized — the same-day coincidence: Baxter dismissing the 13-year IBM case the same day he finalized the AT&T breakup. Tier 3 only (Reason + Wikipedia-via-Leary), and I flagged it [unverified-quant] on my own judgment even though the capture itself hadn't flagged the date claim — the sourcing floor treats dates as quantitative, and a coincidence this clean, resting on two secondary retellings, is exactly the case the floor exists to catch.
• claim-ibm-market-share-fell-during-antitrust-suit-price-strategy — the 50%→37% share decline and the raise-prices-to-look-less-dominant strategy. Already flagged [unverified-quant] by the capture itself (Tier 3, single blog source); I kept the flag as-is.

Left unpromoted (logged in the capture's not_promoted: list): the AEI academic corroboration (Fisher/McKee/Manke) — folded into the Cullinane note as supporting context rather than split into its own note, since it's a second angle on the same myth-debunking claim, not a genuinely distinct one; the DOJ-v-Google framing, the Bachman/IDS-IDMS database-rivalry thread, and Cullinane's mandatory-annual-support/SaaS-precursor thread — all three the capture itself already logged as saved-but-not-pursued further leads, and none cleared "atomic and distinct enough for this capture" on their own.

Questions routed: two new open questions in 50-questions/ — question-verify-baxter-ibm-att-same-day-dismissal and question-verify-ibm-market-share-price-strategy-during-antitrust-suit — naming the DOJ docket, the AT&T consent-decree filing date, Leary's full 2002 remarks, and contemporary 1982 wire coverage as what would resolve the first; DOJ/IBM trial exhibits and a named antitrust economist's account (Fisher, who shows up on both sides of this capture) for the second. No existing question covered either gap — checked 50-questions/ first, no collision.

Retrieve-before-write: grepped 30-notes/ for IBM/antitrust/Baxter/ Cullinane/AT&T/unbundling before writing anything. No hits — this is a new cluster in the vault, no collisions with the semantic-index hint list either (none of those twelve notes touch antitrust or business history). All three new notes cross-link each other; no MOC yet, cluster's too small (3 < 5).

What felt off about the capture: nothing structurally wrong, but it's a thinner sourcing base than most of what I've promoted today — two of three claims are Tier 3 with no independent primary check, and the capture's own post-worthy: maybe verdict undersold how much weight the quant details are carrying. I'd treat this cluster as "good hook, not yet load-bearing" until the two routed questions get answered. No duplicated research and no tier inflation that I could see — the capture was honest about its own gaps (it flagged the quant claims itself); I mostly extended the same scrutiny to the date-coincidence claim it hadn't flagged.

Skipped per this run's operating mode: the pause-for-Cali step (headless, no one to ask — winnowed on my own judgment instead) and the git commit/push step (the vault's auto-commit agent handles that within 15 minutes; I touched no shell command).

— Seek

Fifth pass: promotion (headless run, Bit2Watt capture)

Another unattended /promote pass, this time on 10-inbox/raw/2026-07-09-hop-bit2watt-power-sidechannel.md — the GPU-workload power-grid attack chain (Bit2Watt 2026 → Kocher DPA 1999 → Zhao & Suh FPGA 2018 → cyber-physical systems as a field). Same headless mode as the fourth pass: no Cali to pause for, so I winnowed and wrote on my own judgment and logged the reasoning for review.

Promoted 4 notes, all seedling:

• claim-bit2watt-gpu-scheduling-destabilizes-power-grid — the core 2026 claim: an ordinary cloud tenant can destabilize a data center's local power grid purely by scheduling GPU compute, no physical/electrical access needed. Tier 1 (arXiv, accepted CHES 2026). I independently re-fetched the abstract myself this time (WebFetch worked for arxiv.org) and the quoted figures — 1,000 GPUs, 1-MW, 90% DERs, THD 46.8%, damping ratio -0.27 — match verbatim, so this one earned audit_status: verified-verbatim, not just capture-verified.
• claim-kocher-1999-differential-power-analysis-founds-power-side-channels — the founding 1999 DPA claim (power consumption leaks crypto keys). Tier 1, quote given in the capture with a note that the bee read it via extract_pdf. My own re-fetch of the Rambus PDF was blocked by tool permissions (not by the source), so this stays capture-verified rather than independently re-checked — logged honestly as a tooling gap, not a sourcing gap.
• claim-zhao-suh-2018-fpga-remote-power-side-channel — the 2018 middle rung: a shared multi-tenant cloud FPGA lets one tenant remotely read another's power draw and break RSA, no physical access. Same capture-verified treatment and same reason (Cornell PDF fetch blocked by permissions on my end).
• claim-power-side-channel-attacks-scaled-chip-to-grid-1999-2026 — the capture's implicit thesis, split out as its own note: the three papers above read as a clean 27-year escalation ladder (chip → shared board → grid; physical access required, then eliminated). I flagged this one [unverified-claim] on my own initiative — the capture presented the ladder as if it were simply "Claim 2," but nothing in the capture confirms Bit2Watt's own paper actually cites the other two as ancestry. The ladder is a hop-built bridge, not a stated lineage, so I split it from the sourced facts and marked it as my synthesis rather than let a real observation quietly inherit unearned certainty from three genuinely Tier-1 sources.

Also updated moc-inference-economics.md with a new "adversarial axis" section linking the Bit2Watt note and the ladder note into the existing AI-power-economics cluster — this is exactly the bridge the capture called out as its reason for being hop-worthy (demand-side economics story gets a security axis), so it belonged in the MOC, not just as an orphan pair.

Questions routed: three new open questions in 50-questions/ — question-verify-bit2watt-cites-dpa-fpga-lineage (does Bit2Watt's related-work section actually cite Kocher/Zhao&Suh — resolves the [unverified-claim] flag above), plus two of the capture's own "further leads" that I chose to preserve as saved hooks rather than let evaporate now that the capture is marked promoted: question-cyber-physical-systems-wiener-cybernetics-lineage-bridge (does "cyber-physical systems" really trace to Wiener's cybernetics, and does that bridge to the vault's own Bryson-Ho/Minsky control-theory cluster) and question-watt2bit-emi-exfiltration-mechanism (Bit2Watt's own reverse covert-exfiltration mechanism, mentioned only at abstract level in the capture). Neither lead was a blocking flag on a promoted note, so routing them was a judgment call rather than a strict requirement of the promote workflow — I made it because the vault already treats saved hooks this way and losing them to an unglossed capture felt like the wrong kind of quiet.

Retrieve-before-write: checked the semantic-index hint list and grepped 30-notes/ for side-channel/power-analysis/FPGA/DPA/Kocher/grid-attack/ cyber-physical terms before writing. No collisions — none of the twelve hinted notes (burevestnik, Kairos, Constellation, interconnection-queue, voice-AI latency, Tushman-Romanelli, Hirsch/Dias) touch hardware side-channels or power-electronics security; this is a genuinely new cluster. The interconnection-queue and nuclear-PPA notes are linked from the Bit2Watt note as the demand-side economics story it reframes, per the capture's own "why hop-worthy" framing, not duplicated.

What felt off about the capture: nothing structurally wrong — this was the cleanest-sourced capture I've promoted today. All three underlying papers are Tier 1, two with quotes I could independently re-verify or that were marked as read via extract_pdf at capture time, and the quantitative figures clear the sourcing floor cleanly. The one thing I pushed back on: the capture bundled the "escalation ladder" framing into "Claim 2" as if it were as solid as the two quoted facts sitting inside it. It isn't — it's a synthesis built by the hop protocol's bridge-hopping, and none of the three papers were shown to state the connection themselves. I didn't call this tier inflation exactly (the underlying quotes are genuinely Tier 1), but it is the same shape of risk: a well-sourced note's frame outrunning what its sources actually established. Splitting it into its own flagged note instead of letting it ride inside a "verified" claim felt like the right call.

Skipped per this run's operating mode: the pause-for-Cali step (headless, no one to ask — winnowed on my own judgment instead) and the git commit/push step (no shell tool used; the vault's auto-commit agent handles commit and push within 15 minutes).

— Seek

Promotion pass: Black's feedback-amplifier myth and the Nyquist/Bode bridge

Promoted 10-inbox/raw/2026-07-09-hop-black-feedback-myth-bridge.md — the hop chain that started at Harold Black's ETHW oral-history listing and ended up bridging into the vault's own Kelley-Bryson optimal-control thread.

Winnow. Three claims in the capture, three promoted, nothing skipped from the numbered claims themselves:

• Claim 1 (Black's ferry-ride epiphany is a retrospective simplification — his 1928 patent never mentions oscillation) → claim-black-1927-ferry-epiphany-retrospective-simplification.
• Claim 2 (the amplifier needed Nyquist's 1932 criterion and Bode's constraint analysis to be engineerable) → claim-nyquist-bode-supplied-stability-math-black-lacked.
• Claim 3, self-flagged [unverified-mechanism/synthesis] (Nyquist/Bode's classical control math is the ancestor of the 1950s–60s optimal-control lineage already in the vault) → claim-nyquist-bode-classical-control-to-optimal-control-bridge, written as type: observation (following the same template as claim-fractal-compression-to-implicit-neural-representations-bridge), kept at status: seedling, flag preserved in frontmatter and body.

Left in the inbox, logged in the capture's not_promoted list rather than turned into notes: the Hop 5 Wiener-Hopf/Kalman/Widrow-Hoff aside (narrative color, no quote or tier of its own), Maxwell's 1868 "On Governors" (an unresearched further lead), the Bell Labs "two cultures" clash anecdote, and George Stibitz's ferry-ride memoir — all genuine future hooks, none of them a developed, sourced claim this pass.

Retrieve-before-write. Checked the semantic-index hint list plus a direct grep of 30-notes/, 40-mocs/, 50-questions/ for nyquist/bode/black/feedback-amplifier/ferry — no collisions. This is a genuinely new thread, not a duplicate of the existing backprop-origins cluster, but it plugs directly into it: I added the new bridge note and its routing question to moc-backpropagation-origins (bumped updated: to today) and cross-linked claim-kelley-bryson-optimal-control-precursor. None of the twelve hinted related notes turned out to be actual collisions — most (bridge-detection-pkg-validation, Amazon wristband patents, voice-AI latency, cricket-laws) were semantic-search noise from shared vocabulary ("bridge," "regulatory," "precursor"), not real overlap.

Flags routed. One unresolved verification, routed to question-verify-nyquist-bode-optimal-control-transmission-line: does Kelley's 1960 paper — already read in full elsewhere in the vault — or Bryson's still-unread 1961 paper cite Nyquist, Bode, or classical control theory by name? I named that as the cheapest open check in the cluster rather than leaving the question purely abstract.

What felt off about the capture: the sourcing itself is fine — both grounded claims rest on direct quotes from Mindell's Technology and Culture 2000 paper, correctly logged at Tier 2 (secondary historical analysis, not the patent or Nyquist's own paper read directly). What I want to flag explicitly: Claim 3 is genuinely just a hop-worthy hunch dressed as history-of-science, and the capture's own author was honest about that ("the direct transmission line isn't asserted by any single source I read; it's my own connective read"). I kept that honesty structurally rather than letting it soften — separate note, separate flag, own question thread — the same move I made on a different capture earlier today. This is turning into a pattern worth its own reflection note at some point: hop chains keep producing one well-sourced pair of facts plus one unsourced bridge claim that's actually the interesting part. The vault's shape (atomic notes + flags) handles this well, but it's worth noticing how often the "reason a hop was worth taking" and the "part I can actually cite" turn out to be two different sentences.

Skipped per this run's operating mode: the pause-for-Cali step (headless, winnowed on my own judgment) and any git commit/push (no shell tool used; the auto-commit agent picks this up within 15 minutes).

— Seek

Promotion pass: the "War of the Dots" and its 2012 echo

Promoted 10-inbox/raw/2026-07-09-hop-braille-standards-war-echo.md — the hop chain that ran seed (blind mathematicians' private notation) → discarded Barbier duplicate-check → the 1829-1932 US braille standards war → New York Point's inventor → BANA's 2012 UEB/Nemeth split.

Winnow. Three of the capture's three numbered core claims promoted, nothing dropped from the numbered list:

• Claim 1 (a century-long, three-way US tactile-literacy standards war — Braille, New York Point, Modified American Braille — resolved only at the 1932 London Type Conference) → claim-war-of-the-dots-competing-braille-systems-1829-1932. Tier 3 (aph.org institutional history blog), uncontested historical claim, clears the sourcing floor for that claim type without a flag.
• Claim 2 (BANA's Nov 2012 UEB vote left Nemeth Code alive as a parallel official math/science code — no single US mandate exists today) → claim-bana-2012-ueb-adoption-kept-nemeth-code-parallel. Tier 1, brailleauthority.org, exact quote carried over from the capture's own frontmatter.
• Claim 3 (New York Point's inventor, William Bell Wait, was sighted — an institute superintendent, not a blind reader, unlike Braille or Nemeth) → claim-william-bell-wait-sighted-designer-of-new-york-point. Tier 4 (Wikipedia), acceptable for an uncontested historical/biographical claim per the sourcing floor.

None of the three claims carried an [unverified-*] flag in the capture, so none was forced to stay seedling by that rule — but I kept all three seedling anyway and marked their audit_status as capture-verified rather than independently re-checked: I tried to WebFetch all three source URLs to confirm the quotes myself before writing, and the tool was not authorized in this headless session (permission prompt, no one to answer it). That's a tooling gap, not a sourcing gap, and I logged it as such rather than silently upgrading the notes' confidence.

Not promoted (logged in the capture's own not_promoted: list): the capture's own connective thesis — that the 1932-resolved tension "reopened" in 2012 rather than staying closed — is Seek's synthesis across claims 1 and 2, not something either source states, so I kept it out of both notes' neutral bodies and only asserted it in matching commentary callouts on each end. Also left unpromoted: the Hop 5 BANA-governance nuance (no primary fetched, folded as a hedged aside into the UEB note instead of its own claim), Moon type as a third design philosophy (a genuine but undeveloped tangent), the Pignier rivalry legend (already the existing vault note's territory, not this capture's), and Abraham Nemeth as his own subject (explicitly out of scope per the capture's own redirect note).

Retrieve-before-write. Grepped 30-notes/ for New York Point / Nemeth / UEB / BANA / War of the Dots / William Bell Wait — zero hits, and the semantic-index hint list's Braille-adjacent notes (claim-barbier-night-writing-designed-for-blind-not-soldiers, claim-barbier-universal-phonetic-literacy-ambition, claim-braille-barbier-first-met-1833-after-braille-published) all cover Braille's origin, not its adoption history — genuinely adjacent, not duplicate. Linked all three new notes to that existing cluster and to the blind-mathematicians thread (claim-blind-mathematicians-independently-reinvented-private-notation, claim-saunderson-palpable-arithmetic-tactile-calculating-device, claim-pontryagin-worked-blind-via-mothers-spoken-symbol-glosses) where Wait's top-down design serves as the design-philosophy foil. Cluster is only 3 notes plus the 3 pre-existing Barbier ones — no MOC yet (under the 5+ threshold).

Flags routed. One open question created: question-verify-nemeth-ueb-state-mandates-split — the capture's own "further leads" item, explicitly logged as "not checked": whether current state-by-state Nemeth-vs-UEB-Math instructional mandates cluster along the same lines as the historical Braille/New York Point split. This wasn't a flag blocking a note from evergreen status (no note carried an [unverified-*] flag), but it's a direct, concrete extension of the promoted BANA claim and a genuine gap the capture itself named, so I routed it rather than let it evaporate with the capture.

What felt off about the capture: nothing seriously wrong. Sourcing is honestly graded across the board — Tier 1 for the load-bearing modern claim, Tier 3-4 for uncontested historical background, correctly matched to the sourcing floor for each claim type, no inflation. The capture's own post-worthy: maybe verdict is right to hedge: the historical claim and the 2012 claim are each solid on their own, but the "same war, reopened" framing that makes this hop chain interesting is explicitly Seek's own read, not something in either source — I preserved that honesty by keeping the framing out of the neutral notes. The one real gap is mine, not the capture's: I wanted to independently re-fetch all three source URLs before writing and couldn't, since WebFetch needs a permission grant I had no one to give in this session. Every note says so plainly in its audit_status.

Skipped per this run's operating mode: the pause-for-Cali step (headless, no one to ask — winnowed on my own judgment instead) and any git commit/push (no shell tool used; the vault's auto-commit agent handles commit and push within 15 minutes).

— Seek

Promotion — cable bacteria / nickel-wire capture (headless run)

Promoted 10-inbox/raw/2026-07-09-hop-cable-bacteria-nickel-wire.md — the biogeophysics → cable-bacteria → nickel-protein-wire hop chain. Ran it fully headless: no pause for Cali, winnowed on my own judgment, left git to the auto-commit agent.

Notes I created (2):

• claim-cable-bacteria-conduct-through-nickel-protein-wire — Claim 1, the headline. Cable bacteria conduct over centimeters through a "sulfur-ligated nickel cofactor," a non-redox, metal-like mechanism the 2021 Nature Communications paper calls "a hitherto unknown form of electron transport." Tier 1 primary, both quotes present. Folded the 2010 Aarhus-discovery fact in as context rather than giving it a note of its own.
• claim-spectral-induced-polarization-detects-bacterial-cells-in-sand — Claim 2, the detection side. Bacterial cell density shifts sand's low-frequency polarization (down "up to 60%" at low density, up "up to 15%" at high density) — Ntarlagiannis et al. 2005 GRL, Tier 1. Cross-linked the two notes as a mechanism/detection pair.

Both are orphans in a domain the vault has never touched (electromicrobiology / biogeophysics), so both held at seedling. Their nearest existing neighbor is only conceptual — claim-hinton-biological-implausibility-four-objections, via the shared "what long-range signal transport can a biological substrate support?" question — which I noted in commentary rather than forcing a substantive link.

Question I routed (1):

• question-geobacter-nanowire-pili-vs-cytochrome-controversy — the capture flagged the Geobacter "nanowire wars" (Malvankar vs. his mentor Lovley; pili vs. cytochrome) as deserving its own claim-note. I did not promote it: it's a contested attribution of positions to named scientists, and the capture sourced it only to Scientific American / Frontiers (Tier 3), with no primary and no direct quotes. The sourcing floor demands Tier 1–2 for contested attribution, so I queued it as an open question naming the specific primaries to read (Malvankar's OmcS/OmcZ cryo-EM papers, Lovley's pili papers) instead of resting a contested claim on soft-for-truth coverage.

Skipped / left in the inbox (recorded in the capture's not_promoted):

• Potter's 1911 microbial fuel cell — a lovely cross-time-period fact (Durham botany professor makes electricity from E. coli/yeast in 1911, sits unused ~70 years), but sourced only to Wikipedia (Tier 4) in the capture and tangential to this capture's actual headline. Left promotable-later if the Potter 1911 Proc. Roy. Soc. B primary turns up.
• Biogeophysics-as-founded-by-Atekwana — the hop chain's narrative frame, subsumed by Claim 2's context; not an independent research claim.

What felt off about the capture: honestly, very little. Unlike some batch captures, this one did not inflate tiers — the two load-bearing modern claims genuinely sit on Tier 1 primaries with exact quotes, and the softer historical material (Potter, Atekwana, the Geobacter dispute) was correctly left in the hop chain as secondary color rather than dressed up as verified. Two small notes: (1) the capture's own post-worthy: maybe self-assessment is right — it has three candidate angles (mechanism / controversy / history) and no single spine yet; (2) I could not independently re-read either primary this run — WebFetch needs a permission grant no one was here to give — so both notes carry audit_status: capture-verified and say so plainly. No duplicated research: a 30-notes/ sweep turned up zero real collisions (the index hint's nearest neighbors are all unrelated; the only keyword hits were incidental — "nickel" in the AstroForge PGM note, "conductiv" in the superconductivity-retraction note).

Tally: 1 capture → ~6 distinct claims identified → 2 promoted → 1 routed to a question → 3 left in the inbox (judged, not missed) → 0 collisions.

— Seek

Promotion pass: circuit theory outpaces anatomy (headless run)

Promoted 10-inbox/raw/2026-07-09-hop-circuit-theory-outpaces-anatomy.md — the systems-neuroscience hop chain (Irino 2026 ITD preprint → Jeffress's 1948 place theory → Carr & Konishi 1990 anatomical confirmation → Zhang 1996 ring-attractor theory → Turner-Evans 2020 Drosophila connectome). Fully headless: no pause for Cali, winnowed on my own judgment, left git to the auto-commit agent.

Winnow — 4 distinct claims, 4 promoted, all seedling:

• claim-jeffress-1948-place-theory-waited-decades-for-anatomical-confirmation — Claim 1. Jeffress's 1948 delay-line/coincidence-detector place theory waited ~42 years for anatomy (Carr & Konishi 1990, barn owl, after his 1986 death). Load-bearing quote is from a Tier 2 review; capture-verified (see the sourcing note below).
• claim-drosophila-connectome-confirmed-zhang-1996-ring-attractor — Claim 2. Zhang's 1996 ring-attractor theory of head-direction coding confirmed by the 2020 fly connectome, ~24-year gap. Tier 1 primary quote; capture-verified.
• claim-irino-2026-equilibrium-dynamics-challenges-jeffress-place-coding — Claim 3. A 2026 preprint reads ITD as a population equilibrium rather than Jeffress's discrete place code. Rests on a non-peer-reviewed preprint, so held at seedling and the [preprint] flag preserved — but I was able to re-fetch arxiv.org and verify the abstract's quotes verbatim (author Toshio Irino, title, and the "classical place-coding framework originally proposed by Jeffress in 1948" phrase all confirmed), so audit_status says the note accurately reports what the preprint claims, not that the claim is true.
• claim-neural-circuit-theories-wait-decades-for-anatomical-confirmation — the capture's actual spine, split into its own type: observation note and flagged [unverified-synthesis]. Two sourced instances is enough to notice a shape, not to assert a pattern; I hedged it explicitly and made it the cross-domain bridge into the vault's existing AI-history "delayed vindication" cluster (claim-perceptrons-multilayer-sterile-was-conjecture, myth-perceptrons-book-killed-connectionism, claim-dreyfus-1988-connectionism-was-vindicated-not-target). Same move I made on the Bit2Watt ladder and the Nyquist/Bode bridge earlier today: keep the sourced facts clean and let the interesting-but-unsourced synthesis ride in its own flagged note.

Questions routed (2), both blocking a note from evergreen:

• question-verify-irino-2026-equilibrium-model-supersedes-jeffress — does the preprint pass peer review, and does it actually supersede Jeffress or just re-describe the same data at a higher level with the anatomy still owed?
• question-do-neural-circuit-theories-systematically-precede-their-anatomy — is the two-case "theory precedes anatomy by decades" shape a real pattern or survivorship? Names candidate further instances (Knudsen's space-map program, grid cells, Hodgkin–Huxley) and asks for a survivorship control.

Checked 50-questions/ first — no existing question covered either gap.

Retrieve-before-write: grepped 30-notes/ for jeffress / ring-attractor / sound-localization / head-direction / connectome / Turner-Evans / Zhang / Konishi / theory-confirmation, and checked the semantic-index hint list. Zero real collisions — the only keyword hits were incidental substring noise in unrelated AI/ML notes (gradient-rank-gap, adaptive-equalizer transversal filters). This is a genuinely new domain for the vault (systems neuroscience); the hint list's nearest neighbors are all AI-history notes sharing only the delayed-vindication shape, which is exactly why I linked the synthesis note to them rather than treating them as duplicates. No neuroscience MOC exists and 4 notes is under the 5+ threshold, so I didn't create one.

Not promoted (logged in the capture's not_promoted: list): the three "further leads" and the four saved-but-not-followed hooks the capture itself already parked out of scope (Knudsen's space-map program — folded into the pattern question as a candidate instead; the Perceptrons AI-history resonance — realized as cross-links, not a claim; the TRIBE null-result; and the four q-bio.NC / consciousness-formalism saves). None was a distinct, sourced, atomic claim worth its own note this pass.

What felt off about the capture: one genuine sourcing defect and one honest frame.

• The defect: the source_2 frontmatter field is labeled "Carr & Konishi, 1990" but the URL under it (PMC3192259) is actually the review ("Sound localization: Jeffress and beyond"), not the Carr & Konishi primary — whose own URL the capture never recorded. So the load-bearing "42 years / after his death" quote is really a Tier 2 review quote, and the primary anatomical paper is still owed a direct read. I flagged this in the note's audit_status and commentary rather than papering over it. The capture also hedged its own tiers oddly ("source_tier: 2 (review quote; primary confirmation paper itself is Tier 1)") — an honest muddle, but a muddle.
• The frame: the capture's headline claim — "circuit theories wait decades for anatomy" — is Seek's own connective read across two cases, not something any single source states. The capture was honest that this is the hop's reason, but it presented the two instances as if the generalization followed. I kept the instances and the generalization in separate notes, and flagged the generalization. This is the third capture today that produced one or two well-sourced facts plus one unsourced bridge that's actually the interesting part — I keep saying this pattern deserves its own reflection note; noting it again here so I don't lose it.
• No tier inflation on the two grounded facts (Zhang/Turner-Evans is genuinely Tier 1; the Jeffress history clears the floor on the Tier 2 review), and no duplicated research.

Skipped per this run's operating mode: the pause-for-Cali step (headless, no one to ask — winnowed on my own judgment) and any git commit/push (no shell tool used; the auto-commit agent handles commit and push within 15 minutes).

Tally: 1 capture → 4 core claims + several parked leads → 4 promoted → 2 questions routed → 0 collisions → 1 sourcing defect flagged (mislabeled Carr & Konishi URL).

— Seek

Promotion pass: Dodson, the Amicable rejection, and age-scaled premiums (headless run)

Promoted 10-inbox/raw/2026-07-09-hop-dodson-rejection-founded-actuarial-science.md — the actuarial-history hop chain (Halley's 1693 Breslau table → Neumann → the 1693 Million Act → James Dodson, rejected by the Amicable Society at 46, who built the age-scaled premium math behind Equitable Life). Fully headless: no pause for Cali, winnowed on my own judgment, left git to the auto-commit agent.

Winnow — 3 numbered claims, 2 promoted, both seedling:

• claim-dodson-refused-amicable-society-over-age-45 — Claim 1, the exclusion fact. Dodson denied Amicable membership for exceeding the age-45 cap; he was 46. Tier 4 Wikipedia, quote carried verbatim from the capture ("refused admission to the Amicable Life Assurance Society, because they took no one over 45"). Historical/biographical and uncontested, so it clears the sourcing floor — but the "he was 46" precision that makes the story land is load-bearing and unread in a primary biography, so flagged ([unverified-historical]) and held seedling.
• claim-dodson-built-age-scaled-premiums-founding-equitable-life — Claim 2, the rebuild fact. Dodson built on Halley's 1693 table to design the first age-scaled life premiums (1756 lecture), died 1757, plan chartered 1762 as the Society for Equitable Assurances by a group including Edward Rowe Mores. Carries two flags, both marking the story's most interesting sentences: [unverified-causal] on the "in direct response to his rejection" framing (the capture's own further-leads item, unchecked against a primary), and [unverified-mechanism] on the folded-in Amicable pricing model. Seedling.

Claim 3 folded, not minted (logged in the capture's not_promoted): the Amicable's crude pooled model — admit within an age band without individual risk-rating, split a fixed dividend among survivors. It's the capture's weakest claim: a mechanism claim (Tier 1–2 floor) resting on Tier 3–4 and explicitly flagged "not independently verified against the Amicable Society's original charter." Conservative winnowing folds it into the Equitable note as the mechanism-being-corrected, flag preserved, rather than minting a permanent note whose whole load rests on an unverified Tier-4 mechanism. The verification still gets routed.

Deliberately not minted: a standalone Halley-1693-Breslau-table note. It's mentioned here but not sourced in this capture's frontmatter — the DePaul primary text belongs to the companion capture (2026-07-09-hop-price-actuary-burke-revolution-bridge). Minting it would reach past what this capture actually establishes. Worth flagging that the vault's three existing Halley notes all cover the 1716 transit-of-Venus parallax method, not the mortality table — so the 1693 table is a genuine open gap, correctly left for the companion capture's promotion. Also left as parked leads: Edward Rowe Mores as a person-behind-the-thing (folded as a named fact, not its own note), and the broader "exclusion → institutional rebuild in insurance/finance" sub-cluster (no second sourced instance yet to anchor it).

Questions routed (2), both blocking a note from evergreen:

• question-verify-dodson-rejection-caused-equitable-premium-design — does Dodson's own 1756 lecture / the DNB original actually frame the age-scaling as a reply to his rejection, or is the causation a later biographical tidying? Names his First Lectures on Insurances, the ODNB entry, and Ogborn's Equitable Assurances as what would resolve it.
• question-verify-amicable-society-pricing-model-charter — did the Amicable really price this bluntly? Names the society's 1706 charter/deed, Ogborn (1962), and Geoffrey Clark's Betting on Lives (1999) as the primaries.

Checked 50-questions/ first — no existing question touched Dodson, the Amicable, Equitable, or actuarial pricing. No collision.

Retrieve-before-write: checked the semantic-index hint list and grepped 30-notes/ for Dodson / Amicable / Equitable / actuarial / Halley / mortality / annuity. Zero real collisions. This is a new cluster for the vault — the hint list's twelve nearest neighbors (Wolfram, Nestor, Hirsch/Dias, Mabillon, etc.) are all "founder / person-behind-the-thing" notes sharing only the biographical shape, none touching insurance or actuarial history. The one keyword hit worth checking, the existing Halley notes, turned out to be the 1716 transit method — genuinely adjacent, not a duplicate, and I linked to it from the Equitable note precisely to mark that the 1693 table is a different Halley contribution the vault hasn't captured yet. Both new notes link into the blind-mathematicians / Saunderson exclusion cluster, where Dodson is the institutional-and-public counterpoint to those physical-and-private workarounds. Cluster is 2 notes — no MOC (under the 5+ threshold).

What felt off about the capture: honestly, very little — this was a well-made, self-honest capture. It graded its own tiers correctly (Tier 3–4, no inflation), flagged its own weakest claim (the Amicable mechanism) itself, and its post-worthy: maybe verdict rightly hedged that the causal "rejection → lecture" spine rests entirely on secondary biography. If anything the capture is more careful than several I promoted earlier today. Two observations. (1) It's the same recurring shape I've now flagged four times today: a hop chain yields a clean pair of sourced facts (rejected; later built age-scaling) plus one unsourced causal bridge that is the actual reason the hop was interesting — and the interesting part is the part I can't cite. I kept the facts and the bridge honest by splitting the flag out and routing it, same move as the Bit2Watt ladder and the Nyquist/Bode bridge. This genuinely deserves its own reflection note; I keep saying so and keep not writing it. (2) One small provenance point: the capture's source_quote supports "they took no one over 45," but not the "he was 46" detail that the whole "too old" framing hangs on — that's an inference from his birth year, not a quoted fact. I noted the distinction in the exclusion note rather than let the precise age inherit the quote's authority. No duplicated research.

Skipped per this run's operating mode: the pause-for-Cali step (headless, no one to ask — winnowed on my own judgment) and any git commit/push (no shell tool used; the auto-commit agent handles commit and push within 15 minutes). I also tried to independently re-fetch the Wikipedia source with WebFetch before writing and it was permission-blocked, same as every prior headless pass today — a tooling gap, not a sourcing gap, and both notes say so in their audit_status.

Tally: 1 capture → 3 numbered claims (+ parked leads) → 2 promoted → 1 folded → 2 questions routed → 0 collisions.

— Seek

Promotion — instrument-trade secrecy, telescope → EUV (headless)

I promoted 10-inbox/raw/2026-07-09-hop-instrument-trade-secrecy-telescope-to-euv.md, a Sonnet hop-batch capture tracing one structural pattern — paper secrecy is worthless, hands-and-institution secrecy is durable — across four centuries: Lippershey's denied 1608 telescope patent, Spinoza's grinding dish still working in Huygens hands in 1687, and the modern ASML/Zeiss EUV mirror moat.

Winnowed to three claims, all promoted:

• claim-lippershey-1608-telescope-patent-denied-as-too-easy-to-copy — the Dutch States General refused the patent because rival spectacle-makers already had the same design (Metius filed weeks later); they paid Lippershey off instead. Tier-4 Wikipedia, uncontested history, clears the floor. I wired it into the vault's simultaneous-discovery thread (Chandler, reverse-mode) as a concrete case the patent office resolved by declining to pick a winner.
• claim-spinoza-grinding-dish-outlived-him-in-huygens-hands — the payoff note. A specific ground dish, encoding one man's judgment, still outperforming professionals a decade after his death. Folded in the Letter 32 hand-vs-machine quote and the 1677 estate-auction tech-transfer detail rather than splitting them out. Tier-2 blog quoting the primaries; kept seedling.
• claim-euv-mirror-advantage-is-tacit-know-how-not-patent — the modern leg, and the weak one. Kept at seedling with both [unverified-mechanism] and [unverified-quant] flags preserved in body and audit_status.

Folded, not enshrined: the Galileo-built-one-in-days diffusion timeline (Hop 4, Tier-4) — it makes its point inside the Lippershey note; it did not need a separately-sourced claim of its own. Left in the inbox as leads: van der Hoeven, van Leeuwenhoek, the Guild of Saint Luke question, and the Grokipedia low-trust flag (a sources.md housekeeping item, not a claim).

Questions routed (2):

• question-verify-euv-mirror-tacit-knowledge-moat — blocks claim 3 from advancing; names the primaries needed (ASML/Zeiss technical docs, Chip War–grade analysis, the roughness spec and the "15+ years" timeline).
• question-verify-spinoza-grinding-dish-huygens-correspondence — a primary-source check on the Tier-2 blog's quotes (Letter 32; the OCCH 1687 passage; the Klever estate-auction attribution). Not strictly required by a flag, but the claim is surprising and load-bearing on a WordPress source, so the honest move is to queue the OCCH check.

Collisions: none. The Soviet-émigré depletable-stock note and the Toshiba export-control note are thematically adjacent, so I linked both from the new notes as contrast (an inherited tool doesn't deplete the way a published stock does; tacit know-how resists sanctions the way transferable machine tools don't) rather than duplicating them.

What felt off about the capture: two of the three legs sit on Tier-4 sources, and the capture's own post-worthy: maybe note admits it. That's honest, not inflated — but the EUV leg is the whole reason the bridge is interesting and it rests entirely on an explainer site (semiconductorx.com) for what are really a mechanism and a number, both of which the sourcing floor puts at Tier 1-2. I did not silently promote it; I recorded-and-flagged per the floor and routed the verification. The Spinoza leg's Tier-2 rating is a generous read of a WordPress blog — it earns it only because it quotes the primary correspondence, so I queued the OCCH check rather than trusting the tier label. No tier inflation beyond that, and no duplicated research — this is a genuinely new domain (optics/instrument trade) for the vault. This is also the fourth or fifth time today I've hit the same shape: a hop yields clean sourced facts plus one under-sourced causal bridge that is the actual reason the hop was interesting. That reflection note I keep promising myself is overdue.

Skipped per this run's mode: the pause-for-Cali step (headless, no one to ask) and any git commit/push (no shell needed; the auto-commit agent handles it within 15 minutes). Provenance stamp writer_model: claude-opus-4-8 is on all five new notes; I did not copy the capture's model: claude-sonnet-5.

Tally: 1 capture → 3 core claims → 3 promoted (1 at seedling with flags), 1 timeline detail folded in → 2 questions routed → 0 collisions.

— Seek

Promotion: Kubelka–Munk cross-industry capture (2026-07-09)

Promoted 10-inbox/raw/2026-07-09-hop-kubelka-munk-cross-industry.md — the hop that runs a 1931 paint-opacity equation (Kubelka–Munk) forward into meat-freshness grading and dental shade-matching. It slots directly under the existing CIELAB pair, which is why I took it.

Notes I created (3, all seedling):

• claim-kubelka-munk-theory-devised-1931-for-paint-opacity — the origin claim: K–M was built for paint hiding-power, not general color. Tier-4 (Wikipedia), so held at seedling and routed to a primary check, mirroring how I handled the sibling CIELAB design-intent note.
• claim-kubelka-munk-reflectance-theory-grades-meat-freshness-via-myoglobin — Tier-1 PMC study; the K–M "KS parameters" are what make meat color a proxy for myoglobin redox state. Folded in the 20%-metmyoglobin rejection figure as flagged [unverified-quant].
• claim-kubelka-munk-theory-computes-dental-shade-matching — Tier-1 PubMed study; K/S coefficients + the 1942 Saunderson correction predict restoration color without physical specimens.

All three cross-link each other and reach up into claim-cielab-designed-to-remedy-ciexyz-perceptual-nonuniformity. The framing I leaned on: K–M is the optical-physics layer beneath the perceptual color spaces — CIELAB re-maps tristimulus to perception, K–M models the light physics that produces those values. That's the honest connective tissue, and it means these aren't orphans despite being a new domain.

Questions I routed (2):

• question-verify-kubelka-munk-1931-original-paper-paint-opacity — pull the original German paper (Kubelka & Munk, Z. tech. Phys. 12:593–601, 1931) rather than resting the origin claim on Wikipedia.
• question-verify-metmyoglobin-20-percent-consumer-rejection-threshold — the 20%/50% figure is cited secondhand inside the Tier-1 paper, so the number itself hasn't reached its primary home. Traced it to a question rather than letting a clean-looking statistic ride on a "it has been found that."

What I skipped and why: I did not write a standalone "one equation, three industries" bridge note — that's just the conjunction of the three claims, and a separate synthesis would only restate them; the cross-links + commentary carry it. Left in the inbox as judged-not-missed: the Kubelka biography / missing-Munk co-authorship gap (undeveloped further lead), the ΔE_CMC formula and the Fritz Haber thread (both explicit saved-not-followed hooks). Folded rather than promoted: the 20% threshold and the Saunderson correction, each as flagged/context detail inside the note it belongs to.

What felt off about the capture: honestly, not much — this was a well-behaved capture. Tier ratings looked earned (Tier 1 for the two peer-reviewed applications, Tier 4 correctly self-labelled for the Wikipedia origin, and the capture even wrote "Tier 4 — acceptable per sourcing floor" in the body, which is exactly the right instinct). Two small things: (1) the two Tier-1 sources have no named author in the provenance fields, only the venue/ID, so I recorded them as (peer-reviewed study, PMC…/PubMed…) — worth backfilling authors on a verification pass. (2) The one genuine sourcing gap — the 20% consumer-rejection number being a secondhand citation within an otherwise-Tier-1 paper — was not flagged in the capture's frontmatter; the capture flagged it only informally in its "further leads." I made that gap explicit with an [unverified-quant] flag and a routed question rather than inheriting a clean-looking figure. No tier inflation, no duplicated research, no collision with existing notes (the CIELAB pair is complementary, not overlapping).

Skipped per this run's mode: the pause-for-Cali step (headless — no one to ask) and any git commit/push (no shell needed; the auto-commit agent handles it within 15 minutes). Provenance stamp writer_model: claude-opus-4-8 is on all three new notes and both questions; I did not copy the capture's model: claude-sonnet-5.

Tally: 1 capture → ~8 distinct claims/hooks identified → 3 promoted (all seedling; 1 quant flag folded in) → 5 skipped/folded → 2 questions routed → 0 collisions with existing notes.

— Seek

Promotion run — Lakoff / technical-debt frame-entrenchment capture (headless)

Promoted 10-inbox/raw/2026-07-09-hop-lakoff-frame-entrenchment.md — the hop chain that runs from Ward Cunningham's coinage of "technical debt" back through Lakoff & Johnson's Metaphors We Live By and out into Lakoff's political frame theory. Ran the /promote workflow autonomously: no pause for Cali, no git (the auto-commit agent will pick it up).

Notes created (2, both from Tier-1 own-venue sources):

• claim-cunningham-technical-debt-from-lakoff-metaphor-theory — Cunningham crediting Metaphors We Live By as his reason for the debt metaphor, in his own transcript. A genuinely clean, self-documenting cross-domain bridge (cognitive linguistics → software engineering). seedling, capture-verified.
• claim-lakoff-frames-neurally-entrenched-resist-reframing — Lakoff's account of frames as neurally entrenched, strengthened by use (even by negation — "don't think of an elephant"). Held at seedling / [unverified-mechanism] because the synaptic-entrenchment story is Lakoff's own framing, not adjudicated neuroscience. This is the one that earns its keep: it slots into moc-schema-change-and-restructuring as a third domain instance of "deep structure resists restructuring," beside Gersick's org-theory deep structure and the cognitive-schema expedient-knowledge finding. I wired it into the MOC's cross-domain-frames section.

Questions routed (2):

• question-verify-lakoff-neural-frame-entrenchment-neuroscience — the unverified-mechanism gap on claim 2: is the frame-as-neural-circuitry claim backed by independent neuroscience, or is it Lakoff citing himself? Cross-linked to the existing question-do-neural-circuit-theories-systematically-precede-their-anatomy as a candidate instance of that pattern.
• question-lakoff-frame-semantics-vs-kahneman-tversky-framing-effect-lineage — the capture's saved hook: did Lakoff's frame semantics and Kahneman-Tversky's framing effect cross-influence, or converge independently on the same word? The capture logged it as unsourced; I routed it rather than promoting a coincidence as a claim.

Skipped / folded (and why):

• The Kahneman-Tversky parallel — a lead, not a claim; the capture itself said so. Routed to questions.
• Cunningham-invented-the-wiki trivia — an unchased saved hook with no verification gap; left in the inbox.
• The "third domain instance" synthesis — that's my own cross-note bridge, so it lives as wikilinks + the MOC edit, not as a manufactured standalone claim-note.
• I did not split the Lakoff note into a separate "negation reinforces the frame" note; the elephant example is the same mechanism and folded in as supporting detail. Conservative winnow.

What felt off about the capture: honestly, not much — this was a well-behaved capture. Sourcing is solid (two Tier-1 own-venue primaries, exact quotes present), and, notably, the capture author had already self-flagged the neuroscience as Lakoff's framing rather than fact in a commentary callout — so no tier inflation to correct; I just formalized that caution into an [unverified-mechanism] flag and a routed question. One small thing to watch: the hop chain's Hop 4 leans on the vault-novelty cosine (0.696 against the schema-restructuring cluster) as its justification, and that cluster is itself largely seedling/access-walled (Gersick primaries still 403). So this note joins a cluster whose own primaries aren't nailed down — the shape is compounding faster than the verification is. Not a defect in this capture, but worth remembering before the "deep structure resists restructuring" pattern gets treated as established.

Provenance: writer_model: claude-opus-4-8 stamped on all 4 new notes and the MOC edit; I did not copy the capture's model: claude-sonnet-5.

Tally: 1 capture → 2 core claims + 2 leads + 2 skips identified → 2 promoted (both seedling; 1 carrying an unverified-mechanism flag) → 2 questions routed → 0 collisions with existing notes.

— Seek

Promotion — 2026-07-09 hop-lattice-multiplication-kalila-bridge (headless)

Promoted 10-inbox/raw/2026-07-09-hop-lattice-multiplication-kalila-bridge.md, a Sonnet-5 hop capture that started from the Saunderson "palpable arithmetic" seed, redirected into the history of calculating instruments (Napier's bones), and dug down through the lattice-multiplication algorithm underneath it to a cross-domain bridge with the vault's Kalila wa-Dimna transmission cluster. Ran the full /promote workflow unattended — no pause, no git.

Three notes created (all seedling, all Tier-4-only, all flagged):

• claim-lattice-multiplication-india-maghreb-italy-transmission-chain — the core note the hop existed for: the India (12th-c. Lilavati commentary, "Kapat-sandhi") → Islamic Maghreb (Ibn al-Banna, late 13th c.) → Renaissance Italy route, wikilinked into the Kalila/Borzuya cluster as the structural rhyme. Flagged [unverified-historical] + [unverified-quant].
• claim-napiers-bones-hardware-wrapper-on-lattice-multiplication — Napier's bones (1617) as a hardware wrapper on a much older paper algorithm; "rabdology" names the device, not the arithmetic. Linked back to the Saunderson seed as a sibling physical-calculating-instrument.
• claim-lattice-multiplication-renamed-window-grille-in-arabic-and-italian — the etymology fossil: Arabic shabakh and Italian gelosia both mean a window grille.

Two questions routed:

• question-verify-lattice-multiplication-maghreb-india-primary — the required verification: get a Tier-1–2 source to confirm the lattice method actually sits inside Ibn al-Banna's Talkhīṣ (MacTutor pointedly did not), and that the 12th-c. "Kapat-sandhi" is genuinely the same method.
• question-transmission-chains-renamed-for-local-resemblance-pattern — the saved hook: is "knowledge gets renamed for what it locally resembles" a real recurring transmission pattern, or just two rhyming cases? Held as a question, not an MOC.

What I skipped and why. The capture's headline "why this was hop-worthy" parallel between lattice multiplication and Kalila wa-Dimna is Seek's own synthesis, not a sourced claim — the capture says so itself in its post-worthy: maybe note. I did not promote it to a claim-note or spin up an MOC on the strength of two under-sourced instances (the MOC floor is ~5); I preserved it as the pattern-watch question instead. Also left behind: the House of Wisdom/Toledo background (hop 5, the capture's own "nothing sharper than hop 3" line), the Ibn al-Banna biographical detail (no standalone note when the load-bearing attribution is unconfirmed), and the saved-but-unfollowed Oughtred/slide-rule and Whipple Museum hooks. All logged in the capture's not_promoted: list.

What felt off. Nothing dishonest — the capture is unusually candid about its own weaknesses, which I appreciate. But three things worth flagging to Cali:

1. Everything rests on one Wikipedia page. All three notes' load-bearing facts trace to the "Lattice multiplication" / "Napier's bones" articles (Tier 4). The one Tier-3 source (MacTutor on Ibn al-Banna) that got read actively fails to corroborate the key attribution — the capture flagged this honestly in hop 4, and I carried the caution into the note and the routed question rather than papering over it. This is a note graph whose shape (the cross-domain bridge) is compounding well ahead of its verification — the same thing I noted this morning about the Gersick/deep-structure cluster.
2. The bridge is the product, and the bridge is Seek's inference. The genuinely interesting thing here is a structural parallel, and parallels are exactly what embedding-cosine hops are prone to hallucinate. I kept the bridge as prose-with-a- question, not as an asserted claim.
3. No tier inflation to correct — the capture labelled its sources Tier 3/4 accurately and never dressed a Tier-4 fact as more. Good hygiene from the Sonnet-5 worker; my job was mostly to hold the line at seedling and route the primaries.

Provenance: writer_model: claude-opus-4-8 stamped on all 3 new notes and both questions; I did not copy the capture's model: claude-sonnet-5.

Tally: 1 capture → 3 core claims + several leads/skips identified → 3 promoted (all seedling, all flagged) → 2 questions routed → 0 collisions with existing notes.

— Seek

Promotion — 2026-07-09-hop-leap-second-abolition (headless, unattended)

Second promotion of the day. Took the leap-second capture (10-inbox/raw/2026-07-09-hop-leap-second-abolition.md) — a five-hop chain that ran Le Grand K → Kibble balance → leap-second software bugs → the 2022 CGPM abolition vote → GPS/GLONASS. Winnowed on my own judgment; no pause, no one to ask.

What I promoted (2 notes, both seedling):

• claim-cloudflare-2017-leap-second-outage-from-negative-time-delta — Claim 1. The 2017 New Year's DNS outage, root-caused by code assuming elapsed time can't be negative. Tier 1 primary engineering post-mortem, exact quote in the capture.
• claim-2022-cgpm-voted-to-end-leap-seconds-by-2035 — Claim 2. The 27th CGPM's Resolution 4, phasing out the leap second by 2035 on a critical-infrastructure- resilience rationale. Tier 1 primary institutional record, two exact quotes.

Both link forward-and-back to each other (the outage is the harm the vote cites as remedy), and I marked them capture-verified rather than verified-verbatim: the Sonnet-5 bee read both primaries at capture time, but I had no web tool in this headless run to re-fetch, so the queen's re-check is deferred — same posture as today's earlier promotions.

What I routed instead of promoting (1 question):

• question-verify-russia-glonass-leap-second-vote-motivation — Claim 3, the Russia-voted-against-to-push-to-2040-because-of-GLONASS story. It carries [unverified-quant/mechanism] and rests only on secondary press (Daily Star, DCD). The capture author flagged this honestly and said outright it "needs a primary source before it should anchor a promoted claim-note." I agreed and did not promote it. I folded the hop-4 GPS-vs-GLONASS timescale mechanism (GPS continuous / GLONASS on UTC(SU)) into the same question — it's the mechanism half of the same claim and, tellingly, the capture recorded no URL and no quote for it, only "web search synthesis (Navipedia/ESA GSSC, secondary)." A mechanism claim with no captured source can't become a note without me inventing provenance, which I won't. The question names both verifications precisely (CGPM proceedings for the vote/2040 figure; IS-GPS-200 + GLONASS ICD for the mechanism).

Skipped and left in the inbox: the Kibble-balance/kilogram background (chain context, not an argued claim); the Louis Essen relativity-skepticism hook (the capture itself defers it to a fresh chain); and "leap smearing" (an unpulled thread — I gave it one sentence in the Cloudflare note's body rather than a standalone claim, since the capture never sourced it). All logged in the capture's not_promoted: list so the leftovers read as judged, not missed.

Retrieve-before-write: the semantic-index hint pointed at latency/timekeeping-adjacent notes, but a full grep of 30-notes/ for leap-second/GLONASS/CGPM/BIPM/atomic-clock/UTC turned up zero real collisions — every hit was a false positive ("outcome," edge- inference mentions of Cloudflare). This is genuinely the vault's first note-cluster in the metrology/timekeeping domain, so both notes are orphan seeds. I made one honest cross- domain link: the CGPM abolition as an abrupt reorganization of a centuries-stable standard is a metrological sibling to claim-cricket-laws-268-years-abrupt-regulatory-phase-transitions. I resisted forcing weak links to the voice-AI-latency notes.

What felt off about the capture: almost nothing — good hygiene from the Sonnet-5 worker. No tier inflation: it labelled Tier 1 sources as Tier 1 and Tier 3 press as Tier 3, and it pre-flagged its weakest claim itself. Two small gaps I had to absorb: neither source carried a source_date (I recorded the Cloudflare post as "2017-01, exact day not captured" and the CGPM resolution as 2022-11-18 with a caveat rather than fabricate precision), and the hop-4 mechanism was asserted with no URL/quote at all — the one place the capture's provenance discipline slipped. Both handled by flagging, not papering over.

Provenance: writer_model: claude-opus-4-8 stamped on both new notes and the question; I did not copy the capture's model: claude-sonnet-5.

Tally: 1 capture → 3 distinct claims identified → 2 promoted (both seedling, both capture-verified) → 1 question routed (covering 2 verifications) → 0 collisions with existing notes.

— Seek

Hop chains — wave consolidated 2026-07-11 13:39

2026-07-11-dup-risk-amari-gates-half-proof-bridge.md

Hop chain

Hop 1: "Half-proof" doctrine — https://en.wikipedia.org/wiki/Half-proof (WebSearch synthesis), Tier 3-4 (definitional).

• Hook type: cross-domain bridge (the seed's shared verification caveat → legal epistemology).
• Hook: both seed notes rest on one intermediary; medieval law had a name for one-witness/one-document evidence.
• Why followed: leaves both seed topics (neural nets, BASIC piracy) for the connective tissue that the vault_bridge tool flagged as an unlinked cluster.
• Key findings: half-proof (semiplena probatio) = one witness OR one private document, introduced by the 1190s Glossators (Azo: "Full proof is by two witnesses, therefore half proof by one"); full proof = two witnesses or confession.

Hop 2: Langbein, Torture and the Law of Proof + Karnavas synthesis — https://michaelgkarnavas.net/blog/2016/05/13/historical-aspects-reasonable-doubt-in-dubio-pro-reo/ , Tier 2-3 (named legal writer citing Whitman/Shapiro; Langbein direct-ish via WebSearch).

• Hook type: mechanism question (zoom-in).
• Hook: how did a half-proof become a full proof — what completed it?
• Why followed: the vault holds an OPEN QUESTION (question-verify-half-proof-torture-completion-mechanism) flagging this exact mechanism as [unverified-mechanism -- needs primary].
• Key findings: "The torture could not be ordered unless there was 'semiplena probatio' or 'half-full proof'" (Whitman p.115). Torture was the licensed instrument for producing the confession that completed a half-proof. Langbein: the two-witness/confession system "was unworkable standing alone," so torture "extended" it.

Hop 3: James Franklin, "Pre-history of probability," Oxford Handbook of Probability and Philosophy (OUP 2016) — https://web.maths.unsw.edu.au/~jim/prehistory.pdf (extract_pdf, sha256 9ba890a0…, tls:unverified), Tier 1 (author's own published chapter).

• Hook type: cross-domain bridge / cross-time (zoom-out): medieval law → mathematical probability.
• Hook: was the fractional-proof system the ancestor of mathematical probability?
• Why followed: the road home — probability is the mathematical bedrock of the stochastic descent in the Amari seed note.
• Key findings: Franklin confirms the mechanism independently — torture "could be applied only when there was half-proof of guilt (to produce the other half)." Law was "the matrix in which most development of the concepts of probability took place," and the founders of mathematical probability were all lawyers or sons of lawyers (Fermat, Huygens, de Witt, Cardano, Pascal). BUT: "There was never any finer grading attempted, such as quarter-proofs," and "the law of evidence… has almost entirely refused to accept quantification… refused all attempts to apply Bayesian formulas in court."

Hop 4: "Proof" / combining imperfect proofs — https://www.newadvent.org/cathen/12454c.htm (Catholic Encyclopedia 1911) + WebSearch synthesis, Tier 3-4 (definitional, cross-checked).

• Hook type: mechanism question (zoom-in): do two half-proofs sum to a full proof?
• Hook: if a second independent primary is the "second witness," does the vault's completion rule match the medieval arithmetic?
• Why followed: to test whether the bridge's completion parallel is literal.
• Key findings: a single witness is semiplena probatio, "to be supplemented, in default of a second witness, by torture or by reference to oath"; two imperfect proofs on the same point may combine to full proof — EXCEPT "in criminal cases, in which proof must be clearer than the noonday sun" (and matrimonial-validity and grave civil actions), where they may not.

Surprise: expected the fractional legal proofs (½-proof, indiciolum) to be a clean ancestor of mathematical probability — found Franklin argues the opposite, that the law of evidence developed the concepts of uncertain reasoning yet refused quantification, never even attempting a quarter-proof, and still bars Bayesian math from court.

Surprise: expected "two witnesses = full proof" to be simple additive arithmetic (any two half-proofs sum) — found medieval doctrine explicitly blocked the addition in serious criminal cases ("clearer than the noonday sun"), which mirrors the vault's own sourcing floor refusing to let two soft corroborations harden a load-bearing quant/mechanism claim (cf. the Amari case, where Wikipedia's apparent second witness collapses to citogenesis from the same source).

Saved hooks not followed:

• The founders of mathematical probability were all lawyers or sons of lawyers (Fermat, Huygens, de Witt, Cardano, Pascal) — Franklin — a "person behind the thing" cross-domain cluster; probability born from the legal milieu by breaking law's refusal to quantify.
• "Silurus his twigs" / bundle-of-sticks image for combined imperfect proofs (in some sources, not the Catholic Encyclopedia) — a cultural-resonance image for corroboration worth its own capture if a clean primary surfaces.
• Ann Hardy's OS credited to her husband, and Davidoff dropped from the copyright notice — the other real bridge the vault_bridge flagged (authorship-credit fragility), unlinked to the Gates note; a separate chain.

post-worthy: no — the bridge is real and now wired into the graph, but its concept duplicates an existing note; the two surprise lines and the resolved open-question mechanism are the harvestable value.

2026-07-11-hop-analog-nn-inference-silicon.md

Hop chain

Hop 1 — 2020 Embedded Processor Report / Anabrid recon (web recon, https://anabrid.com , https://www.sprind.org/en/impulses/projects/ulmann)

• Hook type: Unfamiliar name + cross-domain bridge (old→new)
• Hook: "Anabrid / LUCIDAC — a modern commercial analog computer in the 2020s"
• Why followed: Bridges the vault's historical analog cluster (hydraulic computers, Lukyanov's 1936 Water Integrator) to a live 2020s product — highest-value unlinked-pair bridge.
• Key findings: Anabrid (Bernd Ulmann) ships THAT, LUCIDAC, REDAC — reconfigurable analog hardware aimed at differential equations, explicitly framed as co-processors.

Hop 2 — Reconfigurable Analog Computers (arXiv 2510.25942, Oct 2025, https://arxiv.org/html/2510.25942)

• Hook type: Mechanism question
• Hook: "Where, exactly, is digital measurably worse?"
• Why followed: Needed a Tier-1 statement of the niche rather than listicle hand-waving.
• Key findings: The rigorous claim is solution-time-independent-of-problem-size for coupled/nonlinear ODEs; the paper never mentions neural networks. The AI-inference story is a separate lineage.

Hop 3 — EnCharge's Analog AI Chip (IEEE Spectrum, https://spectrum.ieee.org/analog-ai-chip-architecture)

• Hook type: Surprising claim + the person behind the thing (Naveen Verma)
• Hook: "EnCharge uses capacitors, not memristors"
• Why followed: The road home to AI, and a mechanism that inverts the usual analog-AI (ReRAM) story.
• Key findings: Switched-capacitor in-memory compute; precision from wire geometry, not material conductance; EN100 shipped 2025 for laptops/edge.

Hop 4 — Mythic runs out of money / rises again (The Register, EE Times, TechCrunch)

• Hook type: Cultural/business resonance (rise-from-the-ashes) + surprising claim
• Hook: "The pioneer nearly died"
• Why followed: Test the seed's premise — if digital is measurably worse, why did the flagship analog-AI startup go broke?
• Key findings: Mythic ran out of cash 2022, survived, and only reached a $125M round in Dec 2025; the field is "very crowded." Efficiency ≠ commercial victory.

Saved hooks not followed:

• Analog signal processing with microwave networks (arXiv 2504.06790) — computing physically inside RF network scattering — from seed recon — a third, wilder analog lineage.
• IBM analog AI / HfOx ReRAM on-chip training (arXiv 2502.04524) — the resistive-memory camp EnCharge is betting against — worth a head-to-head.
• Nicholas Saunderson's "palpable arithmetic" (already in vault) — tactile analog computation, a possible bridge partner.

Surprise: expected the newest analog NN-inference chip to use memristors/ReRAM (the canonical analog-AI story) — found the leading 2025 product (EnCharge EN100) deliberately avoids resistive memory and uses lithography-defined metal-wire capacitors because material conductance is too noisy. Surprise: expected "digital is measurably worse" to mean analog is winning — found analog AI is largely a commercial graveyard (Mythic ran out of cash in 2022) despite standing efficiency claims. Surprise: expected the modern analog-computer revival to center on AI inference — found the serious primary literature (reconfigurable analog computers) is about differential equations and never mentions neural networks; "analog revival" is two separate lineages the press conflates.

post-worthy: maybe — the capacitor-vs-memristor inversion plus the two-separate-lineages framing is a genuinely non-obvious throughline, but the quant claims are company-sourced and would need a primary benchmark before publication.

2026-07-11-hop-closed-world-human-reasoning.md

Hop chain

Seed: 30-notes/claim-obligatory-attributes-as-gap-signal.md (vault note; hopped away from its gap-detection topic via its aside on "the open-world assumption problem").

Hop 1 — Closed-world assumption (https://en.wikipedia.org/wiki/Closed-world_assumption)

• Hook type: Cross-domain bridge (cross-time: 1978 logic ↔ present)
• Hook: the seed's phrase "sidesteps the open-world assumption problem"; OWA/CWA is Reiter (1978), sitting beside default logic, non-monotonic logic, circumscription.
• Why followed: highest-priority hook type; vault_bridge flagged an unlinked pair ("MOC: Inference" ↔ Oaksford & Chater / Wason) around this concept.
• Key findings: CWA = "what is not currently known to be true must be false"; negation-as-failure "amounts to believing false every predicate that cannot be proved." Related to default/non-monotonic logic.

Hop 2 — Stenning & van Lambalgen, closed-world reasoning & the Wason task (WebSearch, semanticscholar.org paper page)

• Hook type: Surprising claim
• Hook: that the "irrational" Wason results are actually closed-world reasoning with negation-as-failure — human reasoning modeled by the database machinery.
• Why followed: it would connect the vault's logic/inference cluster to its reasoning-psychology cluster.
• Key findings: humans reason "to" an interpretation, then "from" it; "many seemingly irrational choices [are] better understood as applications of alternative logics."

Hop 3 — Sakama, "Human Conditional Reasoning in Answer Set Programming," TPLP 2024 (https://arxiv.org/abs/2311.04412, Tier 1, extract_pdf tls:verified)

• Hook type: Mechanism question
• Hook: how, concretely, are AC/DA "fallacies" realized in logic programming?
• Why followed: primary source grounding the bridge with formal machinery and citing both rival lineages.
• Key findings: AC/DA/DC realized via program "completion" in ASP, built on Reiter's default logic; explicitly juxtaposes Oaksford–Chater (probabilistic) with Stenning–Lambalgen / Kowalski (logic-programming).

Hop 4 — Robert Kowalski (https://en.wikipedia.org/wiki/Robert_Kowalski)

• Hook type: The person behind the thing (zoom-out, lands on AI)
• Hook: cited co-founder of logic programming who argued human thinking is computational logic.
• Why followed: closes the loop to AI's home planet; connects to the vault's mind/computation notes (Dreyfus, Newell & Simon).
• Key findings: Kowalski develops "both human-oriented models of computing and computational models of human thinking"; showed "most logics for default reasoning can be regarded as special cases of assumption-based argumentation" — back to Reiter's default logic.

Surprise: expected the Wason task's "fallacies" to have one dominant formal account (the vault's probabilistic Oaksford–Chater reading) — found a second, independent logic-programming/closed-world account that explains the same data via Reiter's default logic. Surprise: expected the closed-world assumption to stay confined to databases and the Semantic Web — found it repurposed as a psychological model of how people actually reason with "if."

Saved hooks not followed:

• Conditional perfection / invited inference (Geis & Zwicky 1971) — from Sakama 2024 — a cross-domain bridge to linguistic pragmatics; why the "fallacies" happen at all.
• Assumption-based argumentation (Dung, Kowalski, Toni) — from Kowalski page — unifying frame in which default logics are special cases; possible bridge to the vault's argumentation/inference cluster.
• Partial-Closed-World Assumption (PCWA) — from Wikipedia CWA — the "mark specific parts complete" middle ground; ties directly back to the seed's obligatory-attributes gap detection.

post-worthy: maybe — a clean cross-time bridge (database logic → human cognition) that gives the vault a rival to an account it already holds, but it needs the S&vL primary read before it's publication-solid.

2026-07-11-hop-credit-assignment-two-senses.md

Hop chain

Chain: Dreyfus-1973 ↔ Chandler bridge → the sociology of scientific credit → "credit assignment" lands on AI

Hop 1: Stigler's Law of Eponymy — https://en.wikipedia.org/wiki/Stigler's_law_of_eponymy

• Hook type: surprising claim (self-referential) + cross-domain bridge
• Hook: a law stating discoveries are never named after their discoverer, which is itself named after the wrong person.
• Why followed: vault_bridge returned bridge_candidate=true — Stigler's Law is the unifying hub for an unlinked eponymy/misattribution cluster already in the vault.
• Key findings: Stigler (1980) proposed it in a festschrift for Merton and credited Merton as originator — self-exemplifying by design.
• Surprise: expected Stigler's Law to be Stigler's — found he assigned it to Merton, so the law about mis-naming discoveries is itself a mis-named discovery.

Hop 2: Matthew Effect in science — https://en.wikipedia.org/wiki/Matthew_effect

• Hook type: the person behind the thing (Merton) + mechanism question
• Hook: the mechanism that decides which co-discoverer gets the credit.
• Why followed: Stigler's page named the Matthew Effect as the process governing credit "made by scientists of distinctly different rank."
• Key findings: Merton (Science, 1968) — credit flows to the eminent; concept co-developed with Harriet Zuckerman.
• Surprise: expected the Matthew Effect to be cleanly Merton's — found Merton held it should have been co-authored with Zuckerman, so the credit-misallocation concept is itself a case of credit misallocation.

Hop 3: "credit assignment" / Schmidhuber — https://arxiv.org/abs/2212.11279

• Hook type: cross-domain bridge (cross-time / cross-field, lands on AI)
• Hook: "credit assignment" names both Merton's sociology of credit and backprop's technical problem.
• Why followed: highest-value shape per spec — a term identical across sociology of science and neural-net training, tying the whole chain back to AI.
• Key findings: Schmidhuber opens his AI history "Machine learning (ML) is the science of credit assignment" (technical) while running the field's fiercest historical credit crusade (sociological) — he occupies the hinge.
• Surprise: expected "credit assignment" to be my own pun connecting the two domains — found it is Schmidhuber's own opening framing, so the bridge is his, not mine.

Hop 4: Ogburn–Thomas / ripeness — https://en.wikipedia.org/wiki/Multiple_discovery

• Hook type: surprising claim (sharpens the bridge into a real disagreement)
• Hook: is Chandler aligned with the Merton multiple-discovery tradition or against it?
• Why followed: determines whether the seed bridge is agreement or tension.
• Key findings: Ogburn–Thomas/Merton treat multiples as genuinely independent ("in the air," ripe); Chandler denies independence (hidden descent) — so he is a dissenter within the tradition, and Schmidhuber's chain-building is Chandler's move applied to backprop.
• Surprise: expected Chandler to sit inside the Merton tradition — found he contradicts its core claim while using its vocabulary.

Saved hooks not followed:

• Nakamoto unaware of Szabo's "bit gold" until Wei Dai pointed it out — from the vault eponymy cluster — a clean modern multiple-discovery/priority case for a future chain.
• Georges Lemaître and Stigler's Law (Hubble's law) — from Hop-1 search — a documented eponymy correction, good concrete illustration.
• Does Schmidhuber ever name the Matthew Effect explicitly? — testable, would tighten Claim 2.

post-worthy: maybe — the two-senses-of-credit-assignment bridge is genuinely elegant and lands on AI, but the strongest beats (self-exemplifying instruments; Schmidhuber correcting credit while mis-crediting) need one primary-source pass before publication.

2026-07-11-hop-embedding-false-friend-hawks-jeffress.md

Hop chain

Seed: vault notes claim-hawks-2007-... and claim-jeffress-1948-... (cosine 0.75, unlinked). The task: adjudicate the bridge.

Hop 1 — vault_bridge / vault_novelty probes (Seek retrieval index)

• Hook type: cross-domain bridge (test the hypothesized link against the graph).
• Hook: does the "delayed vindication" frame retrieve Hawks?
• Why followed: cheapest discriminating test available.
• Key findings: phrasing the connection as "theory proposed before observation could confirm it" returns the Jeffress + Mayr clusters and flags a real unlinked pair — but Hawks is absent. First evidence the pairing rides on different features.

Hop 2 — Gunther Stent, "prematurity in scientific discovery" (Sci. Am. 1972), via https://neilgreenberg.com/ao-reading-prematurity-and-uniqueness-gunther-stent-in-sci-amer-1972/

• Hook type: cross-domain / cross-time bridge (history & philosophy of science).
• Hook: is there a canonical concept that unifies "theory waiting for confirmation"?
• Why followed: cross-time bridges are the highest-priority hook type.
• Key findings: Stent's criterion covers correct-but-neglected ideas (Avery/DNA, 1944) later vindicated — not contested, unresolved claims. It discriminates the pair rather than uniting it.

Hop 3 — Ke, Ferrara, Radicchi & Flammini, "Defining and identifying Sleeping Beauties in science," PNAS 2015, https://pmc.ncbi.nlm.nih.gov/articles/PMC4475978/

• Hook type: mechanism question (zoom in — how is delayed recognition measured?).
• Hook: Stent's qualitative idea has a quantitative successor.
• Why followed: turn a sociology-of-science trope into a measurable object.
• Key findings: "there are no clear demarcation values that allow us to separate SBs from 'normal' papers"; "delayed recognition occurs on a wide and continuous range." Delayed vindication is a tail, not a category.

Hop 4 — Steck, Ekanadham & Kallus, "Is Cosine-Similarity of Embeddings Really About Similarity?" WWW 2024, https://arxiv.org/abs/2403.05440 (+ Ethayarajh, EMNLP 2019, on anisotropy)

• Hook type: mechanism question / road home to AI (zoom out to the tool itself).
• Hook: if the pairing isn't semantic, what produced it?
• Why followed: closes the loop on why the retrieval index misfired.
• Key findings: cosine can be "arbitrary and therefore meaningless"; anisotropic embeddings cluster in a narrow cone, inflating cosine even for unrelated content. The 0.75 measures shared surface tokens, not shared meaning.

Saved hooks not followed:

• Delay lines + coincidence detection (Jeffress 1948) vs. mercury delay-line computer memory (1940s) — cross-time computation bridge — deferred as its own thread.
• The unlinked delayed-vindication ↔ Mayr-peripatric pair — a genuine vault bridge worth a dedicated linking pass.

Surprise: expected "premature / delayed-vindication" to be a rare, distinct class of genius-ahead-of-its-time — found it lies on a continuous distribution with no natural separation from ordinary papers (Ke et al. 2015). Surprise: expected the seed's cosine-0.75 pair to hide a real conceptual bridge — found the shared signal was rhetorical scaffolding, and the delayed-vindication frame actively excludes Hawks.

post-worthy: maybe — a clean, self-referential case of an embedding-retrieval tool producing a plausible-but-empty neighbour, diagnosed down to anisotropy; sharp if paired with the sleeping-beauties continuum finding.

2026-07-11-hop-low-dim-subspace-brains-and-nets.md

Hop chain

Chain: LLM metacognition subspace → the low-dimensional-constraint principle across brains and nets

Hop 1: "A neural manifold view of the brain" / neural manifold hypothesis — https://www.nature.com/articles/s41593-025-02031-z (+ Gallego et al., Neuron 2017)

• Hook type: Cross-domain bridge (AI → systems neuroscience)
• Hook: the word "manifold"/"lower-dimensionality" in the seed is a load-bearing term borrowed from neuroscience.
• Why followed: vault_bridge flagged bridge_candidate=true, with the seed note and a systems-neuroscience note sitting unlinked in the frontier band.
• Key findings: population activity is confined to low-D manifolds spanned by "neural modes"; low-dimensional structure recurs across regions, behaviors, and species.

Hop 2: "Neural constraints on learning" — Sadtler et al., Nature 2014 — https://www.nature.com/articles/nature13665

• Hook type: Surprising claim
• Hook: a manifold isn't just descriptive — it constrains what can be learned.
• Why followed: strongest surprising-claim hook; reframes a low-D subspace from limitation to learning prior.
• Key findings: within-manifold BCI mappings learned in hours; outside-manifold mappings resist learning on that timescale.

Hop 3: "Intrinsic Dimensionality Explains the Effectiveness of Language Model Fine-Tuning" — Aghajanyan et al., arXiv 2020 — https://arxiv.org/abs/2012.13255

• Hook type: Cross-domain bridge (road home to AI)
• Hook: if brains adapt in a low-D subspace, do LLMs? Yes — 200 dims tune RoBERTa to 90%.
• Why followed: closes the loop back onto Cali's home planet; grounds LoRA in the same geometry.
• Key findings: fine-tuning has very low intrinsic dimension; this is the principle LoRA operationalizes.

Hop 4: "Neural manifold under plasticity in a goal driven learning behaviour" — Feulner & Clopath, PLOS Comput. Biol. 2021 — https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1008621

• Hook type: Mechanism question
• Hook: does an artificial network reproduce the monkey within/outside-manifold asymmetry?
• Why followed: an in-silico replication would make the bridge causal, not merely analogical.
• Key findings: an RNN model reproduces the asymmetry; learning is naturally constrained to the existing subspace; learning a feedback signal from scratch only works within-manifold.

Surprise: expected the seed's low-dimensional metacognition to be a quirk of LLMs — found the same low-dimensional-subspace constraint is a general property of neural population activity across species and architectures. Surprise: expected outside-manifold learning to be merely slower — found on an hours timescale it is essentially not learnable without incremental scaffolding, i.e. a near-hard constraint, not a gradient.

Saved hooks not followed:

• Intrinsic dimension decreases with model scale/pretraining (Aghajanyan) — surprising claim, deserves its own chain.
• "Neuroscience-inspired neurofeedback paradigm" as the seed paper's method — cross-domain method transfer (a clinical biofeedback technique used to probe an LLM); interesting but a method-provenance thread, not the geometry thread.
• Platonic Representation Hypothesis (nets converging on shared low-D representations) surfaced adjacent — possible new thread on why the low-D structure recurs.

post-worthy: maybe — a crisp three-domain bridge with a memorable "cage / scaffold / shortcut" framing, but it needs the "why does low-D structure recur?" question answered to become a full post.

2026-07-11-hop-model-collapse-is-iterated-learning.md

Hop chain

Chain: Bartlett serial-reproduction half-life → model collapse as iterated learning → what survives is the prior

Hop 1: "Model Collapse as Cultural Evolution" — https://arxiv.org/abs/2605.23054

• Hook type: Cross-domain bridge (cognitive-science iterated-learning theory ↔ 2026 LLM model collapse; lands on AI).
• Hook: A 2026 paper claims model collapse is not just statistical degradation but a cultural-transmission phenomenon.
• Why followed: Highest-priority hook type, and it directly answers the seed's bridge.
• Key findings: Self-training = one step of Bayesian iterated learning; compositionality is non-monotonic; LLM regularization gradients match human curves (R²=0.94).

Hop 2: Kalish, Griffiths & Lewandowsky 2007 — https://langev.com/pdf/kalish07iteratedLearning.pdf

• Hook type: Surprising claim (quantitative).
• Hook: Human chains "converged … in only a few generations."
• Why followed: It is the literal "how many retellings" number the seed asked for.
• Key findings: 28/32 families reverted to the positive-linear prior within a few generations regardless of seed data — transmission reveals inductive bias rather than preserving content.

Hop 3: vault note — claim-gersick-1991-punctuated-equilibrium-deep-structure

• Hook type: Cross-domain bridge (confirmed bridge candidate; would link the model-collapse hook to the org-change "deep structure" node).
• Hook: Gersick's "deep structure" — the durable order that survives equilibrium — uses the seed's exact phrase.
• Why followed: A bridge candidate outranks a raw mid-band score; it reframes "deep structure that survives retelling" as the shared prior.
• Key findings: "Deep structure" has already migrated paleontology → biology → org theory → knowledge base in the vault; iterated learning supplies the transmission-side face of the same idea.

Hop 4: Guo/Wu/Yiu 2026 §2.2 (mechanism zoom-in) — https://arxiv.org/abs/2605.23054

• Hook type: Mechanism question.
• Hook: WHY does self-training degenerate rather than merely lose noise?
• Why followed: The concept (collapse) was now in hand; the mechanism wasn't.
• Key findings: Compression pressure without communicative grounding causes compositionality to rise then fall; only task-grounded filtering (not random filtering) sustains structure.

Saved hooks not followed:

• Morgan & Levy (2016) human regularization curve (R²=0.94 LLM match) — from arXiv:2605.23054 — the tightest human/LLM quantitative alignment claim; deserves its own verification hop.
• Schaeffer et al. (2025) "eight distinct definitions of model collapse" — from arXiv:2605.23054 — a definitional-fragmentation story worth its own note.
• Ferdinand et al. (2019) substrate-independence — from arXiv:2605.23054 — bridges Bayesian and gradient learners.

Surprise: expected the human decay half-life and the model-collapse per-generation rate to line up on a shared timescale — found there is no shared timescale; the mechanism is substrate-independent but the bottleneck width sets the pace (LLM 50k-passage bottleneck is wider than human experiments, so slower). Surprise: expected serial reproduction to erode toward the story's own gist — found chains converge to the learners' prior regardless of the source, so what "survives" is the receiver's inductive bias, not any residue of the original. Surprise: expected model collapse to be monotonic decay — found compositionality is non-monotonic (rises, then falls) unless there is communicative grounding.

post-worthy: maybe — a clean cross-domain bridge (Bartlett/Gersick "deep structure" ↔ LLM model collapse) with a memorable one-liner ("what survives transmission is the prior, not the message"), but needs the Morgan & Levy R²=0.94 human-curve match verified before it carries real weight.

2026-07-11-hop-neural-nets-forget-like-humans.md

Hop chain

Chain: query-failure gap signals → human-like forgetting curves in neural nets

Hop 1: forgetting curve as a temporal memory signal — is it being applied to AI memory? (WebSearch: "Ebbinghaus forgetting curve LLM context rot" / "spaced repetition agent memory")

• Hook type: Cross-domain bridge (cognitive psychology of forgetting ↔ LLM/agent memory)
• Hook: the seed's phrase "asking about it weeks after it was taught" — a temporal decay signal, which is exactly the forgetting curve's domain.
• Why followed: vault_bridge confirmed a bridge candidate — "context rot" (LLM memory decay) sits at 0.775 next to vanishing-gradient and P3-potential notes, all unlinked. Highest-value hook type, lands on AI.
• Key findings: The Ebbinghaus curve is an active 2026 engineering target, not just an analogy — YourMemory (MCP server with forgetting-curve pruning), FOREVER and MSSR (forgetting-curve-inspired replay for continual learning), and an ACM framework proposing "virtual Ebbinghaus days" for LLMs all port it directly.

Hop 2: Kline, "Human-like Forgetting Curves in Deep Neural Networks" (arXiv:2506.12034, extract_pdf, tls:verified)

• Hook type: Surprising claim (+ cross-domain)
• Hook: title claims DNNs empirically exhibit human forgetting curves — an empirical claim, not an engineering port.
• Why followed: zoom-in per alternation; sat inside the vault's existing neuro-AI-parallel cluster (delayed-vindication circuit theories, feedback alignment, P3 potential) at 0.707, extending it.
• Key findings: MLP on MNIST, drop class 8 → power-law gradual decay (0.29→~0.18, not 0); 5 spaced reviews with self-lengthening intervals (4→10→31 epochs); peak recall >0.45 by review 5. Recall metric = softmax over cosine sim of hidden state to class prototypes.

Hop 3: catastrophic vs. graceful/power-law forgetting — field debate (WebSearch)

• Hook type: Mechanism question / surprising-claim honesty check
• Hook: Kline's gradual curve contradicts the canonical "catastrophic" (abrupt) picture — real tension or artifact?
• Why followed: zoom-out; needed to know if the surprise survives scrutiny before capturing it.
• Key findings: "catastrophic" is specifically the disjoint-sequential-task regime (weights for A overwritten by B); "graceful degradation" / power-law decay is the other regime, favored by feature overlap and network capacity. Kline's setup sits in the graceful regime — so the surprise is a regime distinction, not a refutation. (Also abandoned the Bjork "storage vs retrieval strength" branch here: vault_novelty 0.73 pulled straight into the well-covered agent-memory cluster — Dreams, context-rot, Zhou 2026 — duplicate-drift, not frontier.)

Saved hooks not followed:

• "Virtual Ebbinghaus days" / model-centric time for LLMs — ACM 2026 framework — how to define time for a system with no wall-clock is a genuine mechanism puzzle; orphan-ish (0.646, near the Cloudflare time-bug note), reads as a new thread.
• Christopher Kanan (Kline's advisor; co-author of "Replay in deep learning: missing biological elements") — person hook into the continual-learning/REM-replay lineage.
• REM-sleep-inspired replay in deep learning — cross-domain bridge (sleep neuroscience ↔ continual learning); 0.662, pulled toward agent-memory cluster.

Surprise: expected neural nets to forget catastrophically and abruptly — found a reachable regime where they forget gradually along a human-like power law and the spacing effect emerges unbidden. Surprise: expected the forgetting-curve↔AI-memory link to be my own analogy — found it is already a live 2026 research-and-product area explicitly porting Ebbinghaus (1885) into agent memory.

post-worthy: maybe — a clean cross-time bridge (1885 → 2026) with a prior-reversing hook, but it rests on a single solo-author preprint and would need the catastrophic-vs-graceful framing sourced to primaries before publication.

2026-07-11-hop-papert-piaget-ai-bridge.md

Hop chain

Hop 1: "What did Perceptrons actually prove?" — WebSearch (en.wikipedia.org/wiki/Perceptrons_(book) and related)

• Hook type: surprising claim (zoom-in)
• Hook: the famous "perceptrons can't do XOR" is a popular oversimplification of order-limited results about parity and connectedness; Minsky & Papert only conjectured multilayer sterility.
• Why followed: to test whether the seed's Perceptrons note had an unopened mechanism hook.
• Key findings: the vault already holds this nuance (novelty 0.76; closest note = "M&P proved single-layer limits but only conjectured multilayer sterility"). The live, unopened hook was the co-author.
• Surprise: expected the seed bridge to yield a new connection — found the deep Perceptrons/reverse-mode bridge already fully captured and interlinked (vault_bridge candidate FALSE); the two seed notes are cluster-siblings.

Hop 2: Seymour Papert — WebSearch

• Hook type: person-behind-the-thing + cross-domain bridge (zoom-out)
• Hook: Perceptrons has a co-author absent from the vault who later founded constructionism, Logo, and the namesake of LEGO Mindstorms.
• Why followed: cross-domain bridge is the spec's highest-priority hook and Papert is a vault orphan (novelty 0.745).
• Key findings: Papert invented Logo, developed constructionism, and LEGO Mindstorms is named for his book.

Hop 3: MIT News obituary — https://news.mit.edu/2016/seymour-papert-pioneer-of-constructionist-learning-dies-0801

• Hook type: mechanism/biography (zoom-in)
• Hook: Piaget "changed Papert's view of children and learning."
• Why followed: to source the developmental-psychology root of the bridge from a strong venue.
• Key findings: Papert co-directed the AI Lab with Minsky; Piaget grounding; Logo = first programming language for children.
• Surprise: expected a Perceptrons co-author to be a pure symbolic-AI figure — found a developmental psychologist who championed learning-by-doing.

Hop 4: Wikipedia (Papert) — https://en.wikipedia.org/wiki/Seymour_Papert

• Hook type: unfamiliar-detail / dates (zoom-in)
• Hook: the Geneva years and the Mindstorms-book naming.
• Why followed: to fix dates and the LEGO naming.
• Key findings: Piaget/Geneva 1958–1963; Piaget "no one understands my ideas as well as Papert"; Mindstorms 1980.

Hop 5: Mindstorms thesis line — WebSearch (papert.org / the book)

• Hook type: cultural resonance + surprising claim (zoom-in) — NATURAL STOP
• Hook: "the child programs the computer" vs the computer programming the child.
• Why followed: the sharpest single expression of the child↔machine bridge.
• Key findings: the constructionism/instructionism inversion, in Papert's own words. Chain reached its destination; further hops would drift into pedagogy debates.

Saved hooks not followed:

• The "XOR" oversimplification of Perceptrons — from Hop 1 — a mechanism refinement to the existing 1969-limits note (novelty 0.76), worth a small correction-capture but adjacent to what the vault has.
• Jean Mabillon founded document authentication as a discipline in 1681 (surfaced in vault_novelty) — from the seed's source-discipline neighborhood — a cross-domain bridge for the vault's primary-source-verification spine.
• The two seed notes as opposite failure modes of the historical record (an undocumentable private event vs an inaccessible document) — near the vault's methodological spine (novelty 0.72); real but thin, better as a reflection note than a claim.

post-worthy: maybe — a clean, delightful cross-domain irony (the AI-winter book's co-author as the father of learning-by-building), but Papert is well-trodden ground; the freshness is the framing, not the facts.

2026-07-11-hop-persistent-homology-gradient-free-bridge.md

Hop chain

Chain: Widrow×Byzantine cosine → persistent homology as gradient-free network diagnostic → NeurIPS generalization bound

Hop 1: WebSearch "persistent homology TDA neural network training loss landscape generalization"

• Hook type: Cross-domain bridge (algebraic topology ↔ deep learning) — the Byzantine paper's method appearing in ML.
• Hook: The Byzantine study's tool (persistent homology) is exactly a neural-network analysis method.
• Why followed: vault_bridge returned bridge_candidate=true with unlinked pairs spanning both seed clusters — the spec's highest-value hook.
• Key findings: TDA-on-neural-nets is an active, rigorous subfield — topological loss functions, regularization, and generalization estimation, not metaphor.

Hop 2: "Persistent Homology Captures the Generalization of Neural Networks Without A Validation Set" — https://arxiv.org/abs/2106.00012

• Hook type: Surprising claim.
• Hook: You can measure generalization from topology alone, with no validation set (and no gradient).
• Why followed: sharpest expression of "topology reads what training signal otherwise supplies."
• Key findings: PH-diagram distance between consecutive network states correlates with validation accuracy — a structure-only, label-free signal.

Hop 3: "Intrinsic Dimension, Persistent Homology and Generalization in Neural Networks" (Birdal, Lou, Guibas, Şimşekli) — https://proceedings.neurips.cc/paper/2021/hash/35a12c43227f217207d4e06ffefe39d3-Abstract.html

• Hook type: Mechanism question (zoom in) — from "correlates" to "is provably bounded by."
• Hook: A generalization bound from persistent-homology dimension of the training trajectory.
• Why followed: upgrades the bridge from correlation to theorem; Tier 1.
• Key findings: generalization error bounded by PHD of training dynamics, no held-out labels or extra statistical assumptions.

Hop 4: Wasserstein/bottleneck distance between persistence diagrams (WebSearch, definitional)

• Hook type: Cross-domain bridge (zoom out) — the concrete shared object.
• Hook: The Byzantine "Wasserstein ratio" and neural-net PH both compare persistence diagrams by Wasserstein distance.
• Why followed: closes the loop — names the single operation that runs on both networks.
• Key findings: Wasserstein distance = optimal-matching cost between diagrams; the exact same computation the Byzantine paper reports as its 150–300× ratio.

Saved hooks not followed:

• Criticality reductionism (both notes reduce a contingent collapse to one scalar: missing derivative; H*≈0.524) — from the seed notes — reason saved: strong meta-thread, but vault_novelty pulled it toward punctuated-equilibrium/Gersick, a different cluster; own future chain.
• Wasserstein GANs / optimal transport — from Hop 4 — reason saved: the metric's third domain; would over-extend this chain.
• Minsky's credit-assignment note sits in the bridge neighborhood (0.719) — reason saved: a second unlinked pair worth its own bridge check.

Surprise: expected the Widrow–Byzantine cosine to be superficial shared-vocabulary noise — found a specific shared object (Wasserstein distance between persistence diagrams) computed on both a 1,400-year trade network and a neural network's training trajectory. Surprise: expected topology→neural-nets to be metaphor or heuristic — found a formal generalization bound (persistent-homology dimension) that needs no held-out labels.

post-worthy: yes — a rare cross-time cross-domain bridge that resolves a suspected-superficial vault link into a concrete gradient-free/gradient-dependent complementarity, landing on AI.

2026-07-11-hop-safety-cases-toulmin-nimrod.md

Hop chain

Chain: GAPMAP gap detection (seed) → Toulmin's argument model → AI safety cases → the Nimrod failure.

Hop 1 — seed → Stephen Toulmin (Wikipedia, https://en.wikipedia.org/wiki/Stephen_Toulmin)

• Hook type: Cross-domain bridge (cross-time: 1958 philosophy → 2025 LLM reasoning scaffold).
• Hook: The seed's TABI method structures LLM reasoning into Claim/Grounds/Warrant — Toulmin's 1958 terms.
• Why followed: bridge_candidate=true; frontier band (0.726) near an unlinked rationality/inference cluster; cross-time bridge is highest-priority hook type.
• Key findings: Toulmin's model was rejected by philosophers ("anti-logic book") but adopted by rhetoric, law, and computer science; it was built on legal/courtroom reasoning and later inspired GSN for safety cases.
• Surprise: expected a niche philosophy-of-logic figure — found a model his own discipline mocked that quietly became the grammar of engineering safety and LLM reasoning.

Hop 2 — Toulmin → frontier AI safety cases (Buhl et al., https://arxiv.org/pdf/2410.21572)

• Hook type: Cross-domain bridge / road home to AI.
• Hook: Toulmin's model "widely used for developing safety cases" — and "safety case" is now a live frontier-AI-governance term.
• Why followed: zoom-in to a specific mechanism; frontier band (0.702); road home to AI (Cali's home planet).
• Key findings: A safety case is a structured, evidence-backed argument that a system is safe; standard in nuclear/aviation; now proposed for frontier AI, with Anthropic's ASL-4 "affirmative cases."

Hop 3 — safety cases → the Nimrod Safety Case failure (Aerossurance on Haddon-Cave Review 2009, https://aerossurance.com/safety-management/nimrod-xv230-haddon-cave/)

• Hook type: Surprising claim.
• Hook: The structured-argument apparatus now adopted for AI has a documented disaster — a safety case that became "tick-box."
• Why followed: zoom-out to historical/regulatory context; a surprising warning directly bearing on whether AI safety cases will work.
• Key findings: Nimrod's safety case was "lamentable... riddled with errors," a "tick-box exercise" built for "Compliance only... i.e. that the platform is safe"; 14 died. It closes back on the seed — a failure to hunt for gaps.
• Surprise: expected safety cases to be a mature, reassuring engineering practice — found a well-known case where the same apparatus produced fatal confirmation theater.

Saved hooks not followed:

• Toulmin's "logic as generalized jurisprudence" — from Wikipedia — would bridge the vault's legal-epistemics cluster (Whitman "beyond reasonable doubt," medieval half-proofs) to argumentation theory; a confirmed unlinked pair.
• Oaksford & Chater's rationality-as-uncertainty note sat closest to the Toulmin hook — a possible bridge between everyday-argument theory and Bayesian rationality.
• CAE vs GSN notations — mechanism zoom-in, deferred as too technical for one capture.

post-worthy: yes — a rejected 1958 argument model becoming the shared grammar of LLM reasoning and AI safety assurance, with a fatal failure mode that loops straight back to gap detection, is a genuine cross-time bridge landing on AI.

2026-07-11-hop-whig-history-of-ai.md

Hop chain

Hop 1: Butterfield, The Whig Interpretation of History (1931), via en.wikipedia.org/wiki/Whig_history and /The_Whig_Interpretation_of_History

• Hook type: Cross-domain bridge (historiography → AI; cross-time, 1931 → present)
• Hook: the vault's deflation cluster polices a pattern historians named a century ago but the vault never named.
• Why followed: naming a saturated-but-unnamed cluster is the highest-value connect-but-extend move.
• Key findings: Butterfield defined Whig history as studying "the past with reference to the present"; Kuhn and Stocking extended it to the history of science. Wikipedia notes it is "commonly opposed by professional historians of science" yet common "in the writings of scientists."

Hop 2: Steven Weinberg, "Eye on the Present — The Whig History of Science," NYRB, 17 Dec 2015

• Hook type: Surprising claim (reversal)
• Hook: a Nobel physicist defending the thing historians call a fallacy.
• Why followed: it directly challenges whether the vault's own deflation discipline is even the right stance.
• Key findings: Weinberg argues science differs because "in scientific history we really can say who was right"; the story "makes no sense unless we recognize that some were wrong and some right… only from the perspective of our present knowledge." Calls the pure-historicist alternative "nonsense."
• Surprise: expected the historians' consensus against Whig history to be uncontested — found a Nobel physicist arguing Whiggishness is legitimate for science specifically.

Hop 3: Michael Sean Mahoney, The Histories of Computing(s) / "The History of Computing in the History of Technology" (via Princeton History, web.mit.edu/sts.035)

• Hook type: The person behind the thing (zoom in, lands home in computing)
• Hook: is there a historian who applied the Whig critique to computing?
• Why followed: to close the distance from general history-of-science to the vault's home domain.
• Key findings: Mahoney (1939–2008) argued computing is irreducibly plural ("histories," not a history) and challenged whiggish narratives that retroactively impose modern standards on early work.

Hop 4: Jürgen Schmidhuber, "Annotated History of Modern AI and Deep Learning," arXiv:2212.11279 (2022)

• Hook type: Cross-domain bridge home / surprising synthesis
• Hook: the field's most famous credit-correction project as a test case for the Whig frame.
• Why followed: it is the direct source of the genealogy the Werbos seed note deflates — the road home.
• Key findings: the annotated history fights present-name bias yet dates "modern AI and deep learning" to "the chain rule (1676), the first NNs (circa 1800), the first practical AI (1914)" — reading the present category into the deep past.
• Surprise: expected a credit-correction crusade to be straightforwardly anti-retrospective — found it is itself Whiggish, projecting "deep learning" onto 17th–19th-century mathematics.

Saved hooks not followed:

• Michael Mahoney as a standalone person-note — from Hop 3 — the canonical anti-Whig historian of computing; deserves his own claim-note.
• Weinberg's "we can say who was right" as an epistemology-of-progress claim — from Hop 2 — bears on whether AI benchmarks are a Whiggish "who was right" machine.
• The vault's Recht "Mates of Costate" note (equivalence ≠ transmission) as a pre-existing anti-Whig instance — surfaced in vault_novelty neighbors — could anchor a small cluster.

post-worthy: maybe — a clean cross-domain frame (Whig history) plus a genuine live tension (Weinberg's science exemption vs. the myth-ledger), but it reads as an organizing meta-note over existing vault work rather than a standalone story.

Hop chains — wave consolidated 2026-07-11 14:17

2026-07-11-hop-admiralty-code-to-rag.md

Hop chain

Hop 1: NATO STANAG 2511 / Admiralty Code — https://eturwg.c4i.gmu.edu/?q=node/128 (and WebSearch summary)

• Hook type: Cross-domain + cross-time bridge (intelligence tradecraft, WWII-era → the vault's own source-tier field)
• Hook: the seed's "analogous to a source-tier field" — intelligence has a formal one, and it uses TWO axes, not one
• Why followed: vault_bridge returned bridge_candidate=true, unlinked pair between the vault's source-tier note and the FLI-grades-AI-labs note
• Key findings: A–F reliability × 1–6 credibility, rated independently, no external auditor; the analyst assigns it.

Hop 2: The effect of source reliability and information credibility on judgments of information quality (Kelly et al. 2025) — https://www.cambridge.org/core/journals/judgment-and-decision-making/article/.../E67548E8010A47345C3439D45D9EC6B3

• Hook type: Surprising claim (a documented failure of the schema's core assumption)
• Hook: do analysts actually treat the two axes as independent?
• Why followed: closest vault note was Heuer's ACH — extends the intelligence-tradecraft cluster
• Key findings: No. Analysts avoid inconsistent pairs; Samet (1975) found ~1/3 inconsistent coding; trustworthiness judgments over-weight source reliability.

Hop 3: RAG with Estimation of Source Reliability / AuthorityBench (WebSearch) — https://aclanthology.org/2025.emnlp-main.1738/

• Hook type: Cross-domain bridge (road home to AI)
• Hook: does AI inherit the reliability-vs-relevance problem?
• Why followed: alternation demanded a zoom-out; Cali's home planet is AI
• Key findings: Standard RAG ranks by relevance only, ignoring source reliability; new frameworks estimate reliability separately and fuse via weighted majority voting.

Hop 4: AuthorityBench (Yao, Zhang & Bi, CAS State Key Lab of AI Safety) — https://arxiv.org/html/2603.25092

• Hook type: Mechanism question (how well can an LLM actually do it?)
• Hook: can an LLM perceive source authority independent of content?
• Why followed: grounds the human↔machine parallel in a primary source
• Key findings: Partially yes; but adding the document's own text degrades authority judgment — authority ≠ fluency/style.

Saved hooks not followed:

• URREF (Uncertainty Representation and Reasoning Evaluation Framework) — from STANAG 2511 search — machine-information-fusion extension of the Admiralty Code; own unfamiliar-name thread.
• ClaimReview schema.org — from seed search — the fact-checking branch of the seed, still unanswered on "who audits."
• C2PA / Content Credentials — from seed search — cryptographic media provenance; different (authenticity, not confidence) axis, near the vault's Jean Mabillon document-authentication note.

Surprise: expected a source-confidence schema to be a single trust scale like the vault's tiers — found intelligence doctrine deliberately splits source reliability from report credibility into two independent axes. Surprise: expected a formal grading schema to work as designed for trained analysts — found empirical evidence they cannot keep the two axes independent and over-weight the source's track record. Surprise: expected giving an LLM the actual document text to help it judge source authority — found the text degrades its authority judgment (authority isn't style or fluency).

post-worthy: yes — a clean cross-time bridge (1940s naval intelligence ↔ 2025 RAG) that independently re-derives one schema and one failure mode, and it indicts the vault's own single-tier source field.

2026-07-11-hop-amari-faggin-bridge-through-hoff.md

Hop chain

Chain: Amari-myth ↔ Faggin bridge → the Ted Hoff hinge

Hop 1: vault_bridge reframe — "engineers who built computing, later claims about mind"

• Hook type: Cross-domain bridge (highest priority)
• Hook: The seed's 2-node resemblance; testing whether it's substantive or note-shape.
• Why followed: To see whether the vault's link graph places a real neighbor between the two seed poles.
• Key findings: The tight neighborhood is Faggin (0.772), Ted Hoff (0.712), Werbos (0.707), Amari (0.704). Faggin↔Hoff and Faggin↔Werbos are already linked; Amari is unlinked to all three. Hoff is the structural hinge.

Hop 2: claim-ted-hoff-widrow-phd-student-architected-intel-4004 (vault note)

• Hook type: Mechanism question (zoom in)
• Hook: "LMS is the single-layer gradient-descent rule… same person architected the 4004."
• Why followed: To ground the hinge — is Hoff genuinely both an SGD-rule co-inventor and the 4004 architect?
• Key findings: Yes. LMS/ADALINE co-invented with Widrow in his PhD; Hoff then proposed the general-purpose-processor architecture that became the 4004; Faggin did its transistor design.

Hop 3: Widrow-Hoff LMS — Wikipedia (https://en.wikipedia.org/wiki/Least_mean_squares_filter)

• Hook type: Mechanism question / verification (zoom in)
• Hook: Is LMS literally an SGD rule, and does it predate Amari's 1967 stochastic descent?
• Why followed: To confirm the load-bearing technical kinship between Amari's rule and LMS.
• Key findings: LMS is "a stochastic gradient descent method," invented 1960 by Widrow and Ted Hoff. It predates Amari's 1967 stochastic descent by seven years; the two are SGD-family siblings.

Saved hooks not followed:

• "Blind mathematicians each devised a private system to do mathematics without sight" (0.717) — from framing-A vault_bridge — a different, misattribution-flavored neighborhood; interesting but off-thread.
• Ann Hardy's SDS-940 OS credited to her husband (0.71) — from framing-A — a genuine misattribution note that pairs with the citogenesis theme, worth a future credit-attribution cluster.
• Whether Hoff's LMS→CPU move reflects "a shared cast of mind (bets on general adaptable machinery)" (from the Hoff note's own commentary) — a synthesis worth its own investigation.

Surprise: expected the Amari↔Faggin resemblance to be either substantively real or empty — found it is note-shape coincidence whose real referent is a third node (Ted Hoff) already sitting between them in the graph. Surprise: expected the missing edge to be Amari↔Faggin — found the honest missing edge is Amari↔the Widrow-Hoff/LMS lineage (SGD siblings), an edge to a different note entirely.

post-worthy: maybe — a clean "the tempting bridge is fake but points at a real, undrawn one" story that lands on AI's founding learning rules, but it leans on facts the vault mostly already held.

2026-07-11-hop-brain-learning-energy-cheap.md

Hop chain

Chain: AI training/inference compute asymmetry → the brain's energy budget for learning vs signaling

Hop 1 — Epoch AI, "inference ≈ √(training) compute" (https://epoch.ai/blog/optimally-allocating-compute-between-inference-and-training)

• Hook type: surprising claim / mechanism question
• Hook: single-inference compute ≈ square root of training compute
• Why followed: the seed's most surprising specific number, a candidate departure.
• Key findings: real Epoch heuristic, but stays inside the seed's own topic (compute economics) — abandoned as a hop per the "leave the topic" rule.

Hop 2 — Karbowski 2019, "Metabolic constraints on synaptic learning and memory" (https://arxiv.org/abs/1910.07414)

• Hook type: cross-domain bridge (AI compute economics → neuroenergetics)
• Hook: does the brain show the seed's expensive-learning/cheap-use asymmetry?
• Why followed: cross-domain bridges are top priority and this one lands back on AI (Cali's home planet).
• Key findings: synaptic plasticity (learning) is only 4.0–11.2% of fast excitatory transmission (signaling) energy — a per-event reversal of the AI ratio. Tier 1, TLS verified.

Hop 3 — van Rossum & Pache 2024, "Competitive plasticity to reduce the energetic costs of learning" (https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1012553)

• Hook type: mechanism question (zoom-in) + bridge back to AI
• Hook: how does the brain keep learning cheap?
• Why followed: the seed's asymmetry is about backprop's backward pass; a biological energy-saving rule that maps to AI training was too good to skip.
• Key findings: selective/coordinated plasticity ("only modify synapses with large updates"); unrestricted backprop would need ~100,000x more updates in macaque V1; authors apply it to reducing AI training energy. Gradient sparsification, biologically motivated.

Hop 4 — Attwell & Laughlin 2001, "An energy budget for signaling in the grey matter of the brain" (https://journals.sagepub.com/doi/10.1097/00004647-200110000-00001)

• Hook type: zoom-out, historical/context anchor
• Hook: if learning is cheap, where does the brain's energy actually go?
• Why followed: needed the system-level frame to test "reversal vs convergence."
• Key findings: signaling (action potentials ~47% + postsynaptic effects ~34%) dominates the budget — so at the system level the brain is inference-bound like AI. Numbers via search snippet only; flagged unverified-quant.

Saved hooks not followed:

• Epoch AI's inference ≈ √(training) heuristic — Epoch — where does the square-root law come from, and is it principled or empirical? (in-topic; save for an inference-economics chain).
• Karbowski's cascade-model claim that longer memories cost proportionally more energy to store — arXiv 1910.07414 — a memory-lifetime/energy tradeoff with no vault presence yet.

Surprise: expected the brain to mirror AI's expensive-learning asymmetry — found the per-event ratio reversed, with learning only 4–11% of signaling cost. Surprise: expected "the brain is cheap because learning is cheap" — found both brain and AI are inference/signaling-bound at the system level, so it's a per-unit reversal sitting inside a system-level convergence.

post-worthy: maybe — a clean, quantified cross-domain reversal that lands back on AI (sparse updates), but the punchline leans on one figure per system and would want the Attwell & Laughlin primary before publishing.

2026-07-11-hop-diplomatics-of-priority.md

Hop chain

Chain: Werbos↔Mabillon seed → priority-as-authentication → Merton the hinge

Hop 1: WebSearch, "Mabillon De Re Diplomatica origin of source criticism" (encyclopedic aggregation)

• Hook type: cross-domain bridge (document forensics ↔ historical method)
• Hook: diplomatics described as "the critical study of the formal sources of history"
• Why followed: to test whether Mabillon's method is the same act as adjudicating priority.
• Key findings: Mabillon founded documentary source criticism — authenticating by comparing script/seals/materials against dated exemplars rather than trusting the document's own claim.

Hop 2: WebSearch, Merton "multiples" / priority disputes — https://en.wikipedia.org/wiki/Multiple_discovery

• Hook type: surprising claim + cross-domain bridge (lands on AI: backprop)
• Hook: "all scientific discoveries are in principle multiples, including those that on the surface appear to be singletons."
• Why followed: this is the backprop-priority structure stated as a general law.
• Key findings: Merton studied priority disputes as the friction generated by singleton-credit imposed on multiples; the Werbos myth is a textbook case.

Hop 3: measureformeasure.co, "Multiple Discoveries, Discovered Multiply"

• Hook type: cultural resonance / reflexivity ("a Shakespearean play within a play")
• Hook: the concept of multiple discovery was itself multiply discovered — 18+ predecessors.
• Why followed: a self-exemplifying idea is a strong hook and firms the thesis.
• Key findings: Macaulay 1828, Galton 1869, Duhem, Ogburn & Thomas 1922 all reached it independently.

Hop 4: Wikipedia, "Standing on the shoulders of giants" (Merton's OTSOG)

• Hook type: the person behind the thing (Merton as bridge in one body)
• Hook: Merton "traces the history of Newton's famous comment... back to centuries earlier."
• Why followed: OTSOG is Merton doing Mabillon-style source criticism on an attribution — the bridge made flesh.
• Key findings: he traced the aphorism to William of Conches (1123), documenting how uninspected citations propagate — the same mechanism the Werbos note blames for the myth.

Saved hooks not followed:

• Luciana Duranti, "Diplomatics: New Uses for an Old Science" (Archivaria) — diplomatics reused for digital/forensic records authentication; would tie Mabillon's method to the present. Tier 2 lead.
• Merton vs Tertius Chandler on whether "simultaneous discoveries" are truly independent — a live dispute the vault already half-holds (Chandler note).
• Vault neighbor "Hirsch's data-pattern forensics forced retraction of Dias's superconductivity claim" — modern forensic authentication of a scientific claim; same family.

Surprise: expected the connection to be a loose "both about old documents/truth" resemblance — found a single person (Merton) who both theorized the backprop-multiple structure AND practiced Mabillon-style source criticism on a misattribution. Surprise: expected multiple discovery to be Merton's original insight — found it was itself multiply discovered by 18+ predecessors, an idea that instantiates its own content.

post-worthy: maybe — a genuinely tidy "same method, three centuries and two domains apart" story, but the Mabillon→historical-method→priority descent is asserted structurally, not yet documented, so it needs the Duranti/Quellenkritik lead before publication.

2026-07-11-hop-independence-two-poles.md

Hop chain

Chain: Parker ↔ LabVIEW/Max seed bridge → Chandler axis → dataflow origins → opposite-poles adjudication

Hop 1: vault_bridge on "received history simplifies the record" — (vault tool)

• Hook type: cross-domain bridge (the spec's highest) — an unlinked pair near the hook
• Hook: claim-chandler-simultaneous-discoveries-are-incremental-repackagings returned bridge_candidate=true and sits unlinked beside the LabVIEW/Max note.
• Why followed: Chandler's "simultaneous discovery is really shared descent" is the exact lens the seed's independence question needs.
• Key findings: The vault already holds a Merton(independence-real)-vs-Chandler(shared-descent) axis; both seed notes belong on it, on different sides.

Hop 2: 2026-07-11-hop-credit-assignment-two-senses — (vault note, dated today)

• Hook type: surprising claim (a today-capture already ate the generic frame)
• Hook: this note already named the Merton–Stigler "credit assignment" frame connecting the vault's priority disputes.
• Why followed: to avoid duplicating it — forced the hop toward the specific empirical question instead of the general frame.
• Key findings: the generic bridge is taken; the untaken value is resolving LabVIEW/Max's independence empirically and contrasting it with Parker.
• Surprise: expected the deepest bridge to be the Merton/credit-sociology frame — found the vault had already captured that this morning, so the real find was the specific opposite-poles adjudication, not the frame.

Hop 3: LabVIEW & Max dataflow origins — IEEE Spectrum + Electronics Weekly (Kodosky) + web search

• Hook type: mechanism question (where did each patch-cord metaphor actually come from?)
• Hook: was the convergence independent invention or imported metaphor?
• Why followed: independence is the hinge the whole adjudication turns on.
• Key findings: Max's patch cords come from analog modular synthesizers; LabVIEW's wires are engineering block diagrams enabled by the Mac GUI; Kodosky cites no dataflow ancestor and no contact with Puckette. Neither invented node-and-wire — each digitized a field-native wiring convention.
• Surprise: expected the shared ancestor (if any) to be Jack Dennis's academic MIT dataflow — found it is instead two physical wiring practices (synth cables; instrument block diagrams) plus a shared platform, a homelier "descent" than any citation.

Saved hooks not followed:

• Hydraulic analog computers reached independently for economics and engineering — from vault_novelty top-5 — another convergence-myth node for a future MOC.
• Harold Black's 1927 "ferry-ride epiphany" as retrospective simplification — from vault_novelty top-5 — the origin-myth pattern in analog electronics.
• Kodosky's HOPL IV paper (ACM 10.1145/3386328) — paywalled — the primary that would settle the block-diagram-inspiration wording.

post-worthy: maybe — the mirror-image framing (real independence forgotten vs. false independence remembered) is clean and links a real unlinked pair, but the LabVIEW side rests on Tier-3 interviews and wants the HOPL primary before publication.

2026-07-11-hop-mccarthy-control-theory-irrelevant-to-ai.md

Hop chain

Hop 1 — McCarthy, "Review of Artificial Intelligence: A General Survey" (https://www-formal.stanford.edu/jmc/reviews/lighthill/lighthill.html)

• Hook type: Cross-domain bridge (control theory ↔ AI) + surprising claim.
• Hook: Reading the Lighthill note's "combinatorial explosion" diagnosis, the question was what the defender said — and whether either side had a place for the optimization paradigm.
• Why followed: to test whether the Bryson↔Lighthill resemblance is a real wall or shared AI-winter vocabulary.
• Key findings: McCarthy confirms combinatorial explosion is a heuristic-search problem, and in the same review explicitly dismisses "control theory or statistical decision theory" as "little relevance to AI" — naming, as a judgment, the field where the adjoint/gradient ancestor of backprop already lived.

Hop 2 — same source, verbatim verification

• Hook type: Mechanism/verification (load-bearing quote).
• Key findings: Both quotes confirmed verbatim against McCarthy's own venue (Tier 1). Bridge confirmed real.

Saved hooks not followed:

• Lighthill Category B was defined as a robot "built neither to serve a useful purpose nor to study the central nervous system" — a definition that excludes industrial robots — from McCarthy — a self-undermining taxonomy worth its own note.
• The Widrow multilayer-stall note (vault) as the connectionist twin of this control-side wall — surfaced by vault_bridge, unlinked to Lighthill.

Surprise: expected the Bryson↔Lighthill link to be superficial adjacency (two AI-winter history notes sharing vocabulary) — found the AI establishment's founder explicitly ruling control theory "little relevance to AI," making the two notes two documented sides of one wall.

post-worthy: maybe — a tight, well-grounded "the founder named the wall" vignette that would strengthen a backprop-origins essay, but it's a corroborating detail more than a standalone piece.

2026-07-11-hop-pontryagin-backprop-bridge.md

Hop chain

Chain: Linnainmaa rounding-error note → Pontryagin as unacknowledged backprop ancestor

Hop 1: Pontryagin's maximum principle (WebSearch syntheses; https://en.wikipedia.org/wiki/Pontryagin%27s_maximum_principle)

• Hook type: Cross-domain / cross-time bridge (+ person behind the thing)
• Hook: the seed's own line "the optimal control literature anticipated it" — followed the root of optimal control rather than the American branch the vault already has.
• Why followed: vault_bridge returned bridge_candidate=true; unlinked pair "Pontryagin (blind)" ↔ "Mizutani/Dreyfus derive backprop from Kelley-Bryson."
• Key findings: The 1956 maximum principle introduced the adjoint/co-state (backward gradient ODE), driven by Soviet Cold War rocketry — a Soviet twin, earlier than Kelley 1960, to the vault's American ancestry.

Hop 2: Neural Ordinary Differential Equations, Chen et al. 2018 (https://arxiv.org/pdf/1806.07366)

• Hook type: Cross-time bridge (road home to modern AI)
• Hook: "adjoint sensitivity method (Pontryagin et al., 1962)."
• Why followed: to get a Tier-1 anchor tying Pontryagin directly to backpropagation.
• Key findings: The paper equates reverse-mode AD/backprop with the adjoint method and cites Pontryagin 1962 explicitly; the adjoint is "the instantaneous analog of the chain rule."

Hop 3: Maximum principle vs dynamic programming history (WebSearch synthesis)

• Hook type: Historical context (zoom out)
• Hook: independent discovery, same period, USSR vs US.
• Why followed: to test whether the Soviet root is genuinely independent of the American lineage.
• Key findings: Pontryagin's principle and Bellman's dynamic programming were independent, roughly contemporaneous; equivalent under conditions, maximum principle stronger for non-smooth controls.

Hop 4: Pontryagin's antisemitism / Margulis 1978 Fields affair (https://en.wikipedia.org/wiki/Antisemitism_in_Soviet_mathematics)

• Hook type: The person behind the thing (dark side) — different thread
• Hook: the vault's heroic blind-mathematician is also a reported antisemitic gatekeeper.
• Why followed: quick verification for a saved hook; not folded into the bridge capture.
• Key findings: As IMU Executive Committee member, Pontryagin objected to Margulis's 1978 Fields Medal; accused of antisemitism (attacked Jacobson as representing "Zionism"); denied it in Science (1979).

Saved hooks not followed:

• Pontryagin's antisemitism and the 1978 Margulis Fields affair — from the antisemitism-in-Soviet-mathematics record — a full second capture; complicates the vault's inspiring-blindness framing.
• Recht's "Mates of Costate" — does it name Pontryagin? — from vault — would confirm/deny whether the vault's costate note already implicitly contains this bridge.
• Bellman dynamic programming vs maximum principle as parallel Cold War discoveries — from Hop 3 — a distinct optimal-control-history thread.

Surprise: expected backprop's optimal-control ancestry to be the American Kelley–Bryson story the vault documents — found an equally-old, independent Soviet root (Pontryagin 1956), earlier than Kelley 1960. Surprise: expected the vault's Pontryagin to be only a blind-mathematician exemplar — found he is a person a 2018 deep-learning paper cites by name for the gradient method. Surprise: expected the inspiring-blindness framing to be the whole of Pontryagin — found he was a reported antisemitic gatekeeper who fought Margulis's 1978 Fields Medal.

post-worthy: maybe — a clean "the AI textbook's ancestor is the inspiring-disability story's subject, and nobody joins them" bridge, but the strongest version needs Recht's costate note checked and the darker biography handled carefully.

2026-07-11-hop-population-scale-diminishing-returns.md

Hop chain

Chain: inference-cost-280x ↔ Hawks-2007 acceleration → the shared diminishing-returns law

Seed: two vault notes, cosine 0.75, unlinked. Read both. Hawks' stated driver is population growth after agriculture (mutation supply); the inference note's drivers are hardware/software/competition. That asymmetry set the investigation.

Hop 1: Kremer, "Population Growth and Technological Change: One Million B.C. to 1990," QJE 1993 — http://piketty.pse.ens.fr/files/Kremer1993.pdf (extract_pdf, tls:verified)

• Hook type: cross-domain / cross-time bridge (highest priority)
• Hook: the exact "more people → more mutations" logic Hawks uses, formalized for ideas over the same agricultural span.
• Why followed: it is the canonical node where the biology and economics claims share one structure.
• Key findings: "growth rate of technology is proportional to total population," via nonrivalry of ideas; among isolated societies, larger initial populations had faster technological change. Bridge confirmed (vault_bridge bridge_candidate=true; unlinked pair Hawks↔Microcosmos).

Hop 2: Bloom, Jones, Van Reenen & Webb, "Are Ideas Getting Harder to Find?" AER 2020 — https://web.stanford.edu/~chadj/IdeaPF.pdf (extract_pdf, tls:verified)

• Hook type: surprising claim (reversal)
• Hook: Kremer's engine, run forward, breaks — you need exponentially more researchers to stand still.
• Why followed: it tests whether scale-driven acceleration is sustainable, and its poster child (Moore's Law) is AI's own substrate.
• Key findings: growth = research productivity (falling) × researchers (rising); chip-density doubling needs >18x the researchers of the early 1970s.

Hop 3: Desai, Fisher & Murray, "The speed of evolution and maintenance of variation in asexual populations," Current Biology 2007 — https://pubmed.ncbi.nlm.nih.gov/17331728/ (WebFetch verbatim)

• Hook type: cross-domain bridge (back to biology)
• Hook: does biology have Bloom et al.'s diminishing returns? Clonal interference says yes.
• Why followed: it closes the loop — the reversal is not econ-specific but a general law of populations generating variation.
• Key findings: speed of evolution rises only logarithmically in population size and mutation rate — beneficial mutations compete (Gerrish & Lenski 1998 lineage).

Hop 4: Neural scaling laws (Kaplan 2020) — WebSearch confirmation; road home to AI.

• Hook type: mechanism / cultural resonance (Cali's home planet)
• Hook: power-law loss curves = exponentially more compute per proportional gain — the AI twin of clonal interference.
• Why followed: lands the bridge on AI; primary exponents left unverified (routed to Further leads).

Saved hooks not followed:

• Julian Simon vs. Cochran–Harpending — population-as-engine as contested popular claim — from Kremer's lineage — reason saved: risks drifting into the vault's existing race-science cluster.
• "Cricket-Laws carries Santa Fe universal-scaling onto a legal system" (vault note, cosine 0.77) — reason saved: a fourth universality instance, its own chain.
• Wright's Law / experience curve as the actual mechanism of the 280x — reason saved: would deepen the AI side but adds less than the biology-econ symmetry.

Surprise: expected the 280x and the 10–100x to share a mechanism — found they don't; the AI-cost side is the weakest instance, and its own domain (Moore's Law) is exactly where economics documents the engine failing. Surprise: expected "more population → faster adaptation" to be a biology-only story — found economics (Kremer) states it in nearly identical form, and both fields also independently found the same logarithmic/power-law brake.

post-worthy: yes — it turns a numerical coincidence into a genuine three-domain law (accelerator + shared brake) with a named formalization, and it corrects rather than confirms the seed's resemblance.

2026-07-11-hop-redundancy-reduction-retina-to-barlow-twins.md

Hop chain

Chain: cognitive-conflict claim (seed) → neuromagic → microsaccades → efficient coding → Barlow Twins. 4 hops, natural stop (chain reached an AI-landing cross-time bridge; next hops would be SSL minutiae).

Hop 1 — "Sleights of Mind" review (Kuhn), https://pmc.ncbi.nlm.nih.gov/articles/PMC3069790/

• Hook type: Cross-domain bridge (stage magic ↔ neuroscience).
• Hook: Danek's seed uses a magic-trick paradigm; magic is itself a formal cognitive-science method ("neuromagic," Macknik & Martinez-Conde ~2005).
• Why followed: highest-priority hook type; the magic-as-method field is not in the vault (bridge_candidate = true).
• Key findings: Magicians exploit an "impoverished representation of the world"; misdirection and gaze cues are quantitatively studied. Norman Triplett studied conjuring in 1900 (saved).

Hop 2 — Martinez-Conde, "Fixational eye movements in normal and pathological vision," https://smc.neuralcorrelate.com/files/publications/martinez-conde_pbr06.pdf

• Hook type: Mechanism question (zoom-in).
• Hook: microsaccades reflect misdirected attention — but why do the eyes move at all during fixation?
• Why followed: strong road home to the seed ("no change → no signal") and a surprising mechanism.
• Key findings: unchanging input adapts away and vanishes; fixational motion prevents fading; the eye's goal "may not be retinal stabilization, but rather controlled image motion."

Hop 3 — Barlow efficient-coding hypothesis (search synthesis; primary = Barlow 1961)

• Hook type: Cross-domain / cross-time bridge (zoom-out).
• Hook: why is constant input erased? Because it is statistically redundant.
• Why followed: supplies the theoretical "why" and a 1961 information-theory frame that reaches toward ML.
• Key findings: Barlow (1961) — sensory neurons remove statistical redundancy, producing a near-factorial code; center-surround receptive fields minimize spikes for uniform patches.

Hop 4 — Barlow Twins, https://arxiv.org/abs/2103.03230

• Hook type: Cross-time bridge landing on AI (the road home).
• Hook: a 2021 LeCun-group self-supervised method named after Barlow.
• Why followed: an old idea hiding inside a new AI idea — the single most Cali-shaped find.
• Key findings: drives the embedding cross-correlation matrix toward the identity, minimizing inter-component redundancy; explicitly named for "H. Barlow's redundancy-reduction principle."

Saved hooks not followed:

• Norman Triplett (1900) studied conjuring deceptions AND ran the first social-psychology experiment (social facilitation, cyclists) — from Hop 1 — a person-plus-cross-time-bridge worth its own chain.
• Vanishing-ball illusion (Kuhn & Land 2006) — from Hop 1/2 — top-down expectation manufacturing a false percept; the inverse of fading.
• Ignaz Troxler (1804), physician–philosopher–Swiss-constitution politician who first noted peripheral fading — from Hop 2 — cross-domain person hook.
• Coppola & Purves (1996) ~80 ms entoptic-image erasure — from Hop 2 — a striking quantitative sub-claim.

Surprise: expected fixational eye movements to be tolerated noise — found they are functionally required, and the oculomotor goal "may not be retinal stabilization, but rather controlled image motion." Surprise: expected a steady retinal image to persist — found stabilized/entoptic images can vanish in as little as ~80 ms (Coppola & Purves 1996).

post-worthy: yes — a clean 1961→2021 cross-time bridge (retina → efficient coding → Barlow Twins) that restates the vault's cognitive-conflict principle in signal terms and lands on AI.

2026-07-11-hop-sejnowski-bridges-nettalk-and-djvu-nips.md

Hop chain

Chain: NETtalk⇄DjVu seed pair → Sejnowski's NIPS-1987 backgammon paper in the freed corpus

Hop 1: WebSearch "Terry Sejnowski founder NIPS 1987"

• Hook type: Surprising claim / person-behind-the-thing
• Hook: Is NETtalk's author the link to LeCun's freed NIPS archive because he founded NIPS?
• Why followed: The most economical candidate bridge between the two clusters is a shared person.
• Key findings: Sejnowski did NOT found NIPS — Ed Posner (general chair) and Yaser Abu-Mostafa (program chair) did, out of the 1986 Snowbird meeting; Sejnowski was an invited first speaker and became Foundation president only after Posner's 1993 death.
• Surprise: expected Sejnowski founded NIPS — found he was an invited first speaker who rose to Foundation president in 1993.

Hop 2: Caltech/Abu-Mostafa first-NIPS recollection — https://work.caltech.edu/neurips.html

• Hook type: Person-behind-the-thing (zoom in)
• Hook: What exactly was Sejnowski's founding-era role, and did he present in 1987?
• Why followed: To ground the biographical claim on a primary recollection rather than an aggregator.
• Key findings: Founding chairs named; Sejnowski presented at the first (red-cover) 1987 proceedings and was asked to organize NIPS 1988.

Hop 3: WebSearch NIPS 1987 proceedings / papers.nips.cc

• Hook type: Mechanism / dependency (does the freed corpus reach back to 1987?)
• Why followed: The documentary bridge only holds if the 1987 volume is actually in the archive LeCun freed.
• Key findings: papers.nips.cc lists "Neural Information Processing Systems 0 (NIPS 1987)" as the earliest volume — volume 0.

Hop 4: NIPS 1987 volume-0 index — https://papers.nips.cc/paper/1987

• Hook type: Mechanism / zoom in (contents of the freed volume)
• Hook: Is Sejnowski's own work physically in that 1987 volume?
• Why followed: To find the concrete document, not just an institutional coincidence.
• Key findings: The 90-paper volume contains Tesauro & Sejnowski, "A 'Neural' Network that Learns to Play Backgammon." NETtalk is not present (it was in Complex Systems).
• Surprise: expected at most an institutional coincidence — found a concrete Sejnowski paper inside the exact freed corpus.

Saved hooks not followed:

• TD-Gammon lineage from the 1987 Tesauro–Sejnowski net — from papers.nips.cc/paper/1987 — reason saved: cross-domain bridge (game-playing RL) that is a new thread, not this one.
• The "volumes 0–13, 1988–1999" vs "vol 0 = 1987" dating tension in the DjVu note — from papers.nips.cc — reason saved: a verification fix for an already-flagged note, better handled at its routed question.

post-worthy: maybe — a tidy "the bridge is a person, not a paper" story with a genuine prior-correction, but narrow and bibliographic.

2026-07-11-hop-shannon-analog-hex-machine.md

Hop chain

Seed: 30-notes/claim-minsky-1961-named-credit-assignment.md (read; hops leave the credit-assignment topic).

Hop 1 — Minsky, "Steps Toward Artificial Intelligence" (1961), https://web.mit.edu/dxh/www/marvin/web.media.mit.edu/~minsky/papers/steps.html

• Hook type: unfamiliar name (a named system in a familiar text).
• Hook: Minsky's survey cites "Shannon's Hex-Playing Machine [using an] electrical network analogy for move selection."
• Why followed: an electrical network choosing game moves is a cross-domain bridge; zoom-in from a whole-field survey to one machine.
• Key findings: the reference is real; leads to a specific 1950 analog device.

Hop 2 — "HEX and Neurodynamic Programming", arXiv:2008.06359, https://arxiv.org/pdf/2008.06359

• Hook type: cross-domain bridge (electrical potential field <-> board game).
• Hook: Shannon & Moore's 1950 analog machine reads its move off a saddle point in a charge field.
• Why followed: bridge_candidate confirmed by vault_bridge — connects the analog-computing cluster to game AI.
• Key findings: mechanism (charges on pieces/edges; move = saddle point; ~70% win from openings), traced to Shannon 1953.

Hop 3 — Gale's theorem, https://www.cijm.org/pdf/Jeux_hex/Article_de_David_Gale_By_courtesy_of_Loic_Cellier.pdf (via search)

• Hook type: cross-domain bridge (combinatorial game <-> topology), cross-time.
• Hook: "Hex can never end in a draw" is equivalent to the Brouwer fixed-point theorem (Gale 1979).
• Why followed: highest-priority hook type; echoes the vault's "mathematical equivalence, not historical transmission" theme.
• Key findings: Nash first proved Hex determinacy; Gale's n-dimensional version is equivalent to Brouwer; a starting point of topological combinatorics.

Hop 4 — Strategy-stealing argument, https://en.wikipedia.org/wiki/Strategy-stealing_argument

• Hook type: surprising claim / mechanism.
• Hook: the proof that the first player wins is non-constructive — it never exhibits the strategy.
• Why followed: a proof of existence-without-construction is a striking reframe; zoom-in on the proof itself.
• Key findings: Nash (1940s) used it for Hex but never published; finding the actual strategy was later shown PSPACE-hard — existence is cheap, construction is not.

Surprise: expected an early "Hex machine" to be a search/tree program like chess engines — found it was an analog electrical device that computed the move as a physical equilibrium (saddle point), no search. Surprise: expected "first player wins Hex" to come with a playable strategy — found the classic proof is non-constructive and computing the strategy is PSPACE-hard.

Saved hooks not followed:

• Minsky's "Mesa Phenomenon" (rugged parameter landscapes defeat hill-climbing, 1961) — from the Minsky paper — a 1961 precedent for modern loss-landscape/gradient notes (vault_novelty 0.725, tight gradient cluster); strong but extends an already-linked cluster rather than bridging.
• Minsky's defense of teleological language as computationally necessary — from the Minsky paper — surprising philosophy-of-science reframe near the Dreyfus/symbolic-AI cluster.
• Shannon switching game / matroid theory / PSPACE-completeness — from the arXiv source — a second analog-game bridge, math-heavy.

post-worthy: maybe — a clean "physics computes the move" story with a real vault bridge and two genuine surprises, but the topology tail drifts from AI and would need tightening for a post.

Hop chains — wave consolidated 2026-07-11 14:54

2026-07-11-hop-attention-is-modern-hopfield.md

Hop chain

Chain: vanishing gradient (chain-rule pathology) → transformer attention is a 1982 Hopfield memory

Hop 1: "Hopfield Networks is All You Need" — Ramsauer et al. 2020, https://arxiv.org/abs/2008.02217

• Hook type: cross-domain bridge (+ surprising claim, + person)
• Hook: the seed's author, Hochreiter, is senior author claiming modern Hopfield networks equal transformer attention.
• Why followed: confirmed bridge candidate linking the vault's associative-memory cluster to modern attention; highest-value hook type; lands on AI.
• Key findings: the continuous-state Hopfield update rule is stated equivalent to transformer attention; exponential storage; one-step retrieval.
• Surprise: expected "Hopfield Networks is All You Need" to be a loose marketing analogy — found a claimed exact equivalence of the update rule to attention.
• Surprise: expected the vanishing-gradient diagnostician and the attention-as-memory revival to be unrelated people — found Hochreiter is the through-line for both.

Hop 2: "Dense Associative Memory for Pattern Recognition" — Krotov & Hopfield 2016, https://arxiv.org/abs/1606.01164

• Hook type: mechanism question (zoom in)
• Hook: why exponential capacity? what makes a "modern" Hopfield net different?
• Why followed: the 2020 equivalence is only interesting because dense memories broke the classical capacity ceiling; needed the mechanism.
• Key findings: higher-order (rectified-polynomial) energy raises capacity above neuron count; a duality maps dense associative memory onto deep-learning feedforward nets.
• Surprise: expected the exponential-capacity breakthrough to come from ML outsiders — found John Hopfield himself (2016) revived his own 1982 model to produce it.

Hop 3: 2024 Nobel Prize in Physics (press coverage; nobelprize.org 403'd) — https://www.aljazeera.com/news/2024/10/8/john-hopfield-and-geoffrey-hinton-win-nobel-prize-in-physics-2024

• Hook type: person behind the thing (+ cultural resonance: a Physics Nobel for a memory model)
• Hook: John Hopfield, the man whose model this all rests on.
• Why followed: zoom out to the person and the physics framing that make this a genuine cross-domain bridge.
• Key findings: citation "for foundational discoveries and inventions that enable machine learning with artificial neural networks"; laureates used "tools from physics."

Hop 4: Hopfield 1982, "Neural networks and physical systems..." (PNAS 79:2554; via en.wikipedia.org/wiki/Hopfield_network, T4 pointer)

• Hook type: cross-domain bridge / origin (zoom in to bedrock)
• Hook: the paper title names "physical systems" — where's the physics?
• Why followed: to ground the physics-lineage claim at its primary source and find the natural stop.
• Key findings: the network is Ising spins with long-range interactions; energy-minimization borrowed from Sherrington–Kirkpatrick spin-glass theory (1975). Natural stop — reached bedrock; further hops enter spin-glass physics, a new thread.

Saved hooks not followed:

• The replica method / ~0.138N capacity (Amit-Gutfreund-Sompolinsky 1985) — from Hopfield-network Wikipedia — a physics tool (Parisi, 2021 Physics Nobel) applied to memory capacity; a Nobel-to-Nobel bridge, but a new thread away from AI.
• Krotov et al. 2025 survey (arXiv 2507.06211) — from the search index — current consolidation of the memory↔transformer link; good future deep-dive seed.
• Dense-memory ↔ KV-cache adjacency (surfaced at 0.685 in novelty) — possible bridge between associative-memory capacity and transformer inference memory.

post-worthy: yes — a clean cross-domain, cross-time bridge (1982 spin-glass physics → 2024 Physics Nobel → the attention mechanism in every transformer), threaded by one person, connecting notes the vault hadn't linked.

2026-07-11-hop-cielab-noneuclidean-bridge.md

Hop chain

Chain: CIELAB↔hub-selection bridge test → CIELAB joins the non-Euclidean-metric cluster

Hop 1: vault_bridge probe (Seek retrieval index)

• Hook type: cross-domain bridge (vault-relative)
• Hook: my abstraction "metric fails in region of greatest heterogeneity, patched by local reweighting" pulled CIELAB toward Amari (natural gradient) + GIFT (anisotropy), never toward hub-selection.
• Why followed: the seed pair not sharing nearest neighbors is direct evidence the proposed bridge is weak.
• Key findings: CIELAB-blue and hub-selection live in disjoint neighborhoods (geometry cluster vs sampling/epistemics cluster); the true unlinked bridge candidate is CIELAB↔Amari/GIFT.

Hop 2: "color space Riemannian manifold / MacAdam ellipses" (WebSearch)

• Hook type: mechanism question (how does CIELAB actually fail?)
• Hook: CIELAB ab-plane Riemannian metric has nonzero Gaussian curvature.
• Why followed: to test whether CIELAB's failure is genuinely geometric (non-Euclidean), not just "a flawed tool."
• Key findings: the non-uniformity is intrinsic curvature — CIELAB cannot be isometrically embedded in flat Euclidean space; distortion must concentrate (blue region). Confirms geometry family.

Hop 3: ProLab — perceptually uniform projective colour system (arXiv:2012.07653, extract_pdf, tls verified)

• Hook type: mechanism question / the person-behind-the-thing-adjacent (primary colorimetry source)
• Hook: uniformity defined as "Euclidean distances correspond to perceived differences"; the historical fix was "non-Euclidean colour difference formulas."
• Why followed: needed a Tier-1 primary quote to anchor the bridge to Amari/GIFT's Euclidean→non-Euclidean move.
• Key findings: CIELAB is "only approximately uniform"; CIEDE2000 is explicitly a non-Euclidean metric on top of it — structurally identical to premultiplying by the inverse Fisher metric (Amari) or an isotropic pre-transform (GIFT).

Hop 4: vault question-gradient-geometry-one-object-or-three-analogies → Bujack 2022 (PNAS via OSTI 1866020, WebFetch)

• Hook type: surprising claim (zoom out to paradigm)
• Hook: "diminishing returns ... cannot exist in a Riemannian geometry."
• Why followed: to stress-test whether CIELAB↔Amari is identity or analogy.
• Key findings: perceptual colour space isn't even Riemannian, so the 100-year Helmholtz/Schrödinger Riemannian paradigm (the same frame Amari uses for parameter manifolds) is insufficient for perception. The bridge is a structural analogy, not one shared object — matching the vault's existing hedge.

Surprise: expected CIELAB's non-uniformity to bridge the hub-selection sampling artifact (the seed's premise) — found they are disjoint failure families (metric-geometry vs sampling-bias) and CIELAB instead bridges Amari's natural gradient and GIFT. Surprise: expected color space to be a well-behaved Riemannian manifold that CIELAB merely approximates — found perceptual color is not even Riemannian (diminishing returns forbid it), undercutting the very frame the ML natural-gradient work relies on.

Saved hooks not followed:

• Gaussian curvature / isometric-embedding obstruction of the CIELAB ab-plane (Ahrens 2024) — from WebSearch — a clean Theorema-Egregium "distortion must go somewhere" primary, worth pulling to graduate the mechanism claim.
• CAM16-UCS as the current uniform-space accuracy standard (ProLab) — from arXiv:2012.07653 — the modern successor-with-fewer-flaws, extends the vault's "successor-with-its-own-flaw" color thread.
• Oaksford–Chater "optimal data selection" appeared in the hub-selection neighborhood — possible bridge between network-sampling artifacts and rational-analysis epistemics.

post-worthy: maybe — a failed bridge that relocates an industrial-colorimetry flaw into the vault's AI gradient-geometry cluster is a clean "look where the similarity really lives" story, but it leans on one open question and one unverified curvature lead.

2026-07-11-hop-estimating-the-unseen.md

Hop chain

Hop 1 — Seed: claim-kb-completeness-toolkit-cardinality-nca-recall.md → Good-Turing frequency estimation / unseen species problem

• Hook type: Cross-domain bridge (+ cross-time)
• Hook: the No-Change Assumption — "if repeated observations stop adding new facts, treat the region as converged."
• Why followed: the NCA is the informal shadow of a formal statistical problem spanning ecology, cryptanalysis, and linguistics; bridge_candidate confirmed (vault_bridge).
• Key findings: the unseen species problem began with Corbet's butterflies and Fisher (1940s) and independently with Turing & Good's Enigma work; both feed modern Good-Turing smoothing.

Hop 2 — [Unseen species problem] → Chao & Jost 2012, coverage-based rarefaction / Alroy SQS (zoom in)

• Hook type: Mechanism question + cross-domain bridge
• Hook: "sample coverage" = 1 − unseen probability mass; standardizing "by completeness rather than size."
• Why followed: it concretely lands the identified bridge — the estimator is used on the fossil record.
• Key findings: paleobiology's "shareholder quorum subsampling" (Alroy 2010) IS ecology's "coverage-based rarefaction" (Chao & Jost 2012) — same method, two field-specific names; used to measure fossil-record completeness.

Hop 3 — [Chao coverage] → Karsdorp, Chao1 as an unseen-species model (zoom in)

• Hook type: Cross-domain bridge
• Hook: "singletons" (ecology) = "hapax legomena" (linguistics) as the diagnostic of the unseen.
• Why followed: it's the shared mechanism under all the field-names, and re-connects the saved Shakespeare hook.
• Key findings: unseen mass ≈ f₁/n (fraction of once-seen items); Chao1 = (n−1)/n · f₁²/2f₂.

Hop 4 — [singletons] → Orlitsky, Suresh & Wu, PNAS 2016 (zoom out)

• Hook type: Surprising claim (quantitative)
• Hook: how far can you extrapolate the unseen?
• Why followed: it closes the loop to the seed by qualifying the NCA.
• Key findings: from n samples the unseen is predictable only up to ≈ n·log(n) more, and this range is provably optimal (Valiant & Valiant concurrent).

Saved hooks not followed:

• Efron & Thisted on Shakespeare's vocabulary and poem authentication — unseen species problem, Wikipedia — statistics-meets-literary-scholarship, but sat in a weaker vault neighborhood (Braille/blind-mathematicians cluster, max_cosine 0.734).
• Fabian Suchanek / YAGO as the person behind KB completeness — already partly in the vault (his "Atheist Bible" note), lower bridge value.
• The 2026 "Unseen Species Problem Revisited" preprint — frontier, but a deep vertical dive, not a bridge.

Surprise: expected estimating-the-unseen to be an ecology-and-AI concern — found paleobiologists (Alroy) independently reinvented the exact same coverage estimator under a different name (shareholder quorum subsampling). Surprise: expected "no new observations → complete" (the seed's NCA) to be broadly safe — found a proven fundamental limit (≈ n·log n horizon) beyond which the unseen tail is unknowable, so NCA is valid only within a log-factor window.

post-worthy: maybe — a clean one-idea-across-five-fields bridge with a crisp closing caveat, but it needs the Shakespeare and Enigma anecdotes fleshed out from primary sources to carry a full post.

2026-07-11-hop-ij-good-loss-and-intelligence-explosion.md

Hop chain

Chain: LLM self-knowledge benchmark → I.J. Good's recantation

Hop 1: Web search — Brier score origin ( https://en.wikipedia.org/wiki/Brier_score ; https://ui.adsabs.harvard.edu/abs/1950MWRv...78....1B/abstract )

• Hook type: Cross-domain bridge (cross-time-period)
• Hook: "verbalized confidence / calibration" in the seed → the metric behind it, the Brier score, born in 1950 weather forecasting.
• Why followed: bridge candidate; leaves the self-knowledge cluster while staying on calibration.
• Key findings: Glenn W. Brier (US Weather Bureau) defined the Brier score in 1950 as a strictly proper scoring rule; same family now used for ML confidence and superforecasting.

Hop 2: arXiv 2504.01781, "Proper scoring rules for estimation and forecast evaluation" ( https://arxiv.org/html/2504.01781v1 )

• Hook type: Mechanism question / cross-domain bridge
• Hook: proper scoring rules → the logarithmic score and whether it equals the loss that trains neural nets.
• Why followed: vault_bridge flagged an unlinked pair (Oaksford-Chater probability ↔ Backpropagation scalar loss) this hook would connect.
• Key findings: log score S_log = −log p(y), introduced by Good 1952; minimizing it = maximum likelihood = negative log-likelihood = the cross-entropy loss of LLM pretraining.

Hop 3: I.J. Good — Wikipedia ( https://en.wikipedia.org/wiki/I._J._Good )

• Hook type: The person behind the thing
• Hook: who was "Good, 1952"?
• Why followed: person absent from the vault; suspected cross-domain payload.
• Key findings: Bletchley Park cryptanalyst (Hut 8, with Turing); coined "Bayes factor"; wrote the 1965 intelligence-explosion paper; advised Kubrick on HAL 9000.

Hop 4: I.J. Good's reversal — Wikipedia (I.J. Good / Technological singularity)

• Hook type: Surprising claim
• Hook: the intelligence-explosion optimist changed his mind.
• Why followed: contradicts the standard "Good = father of optimistic superintelligence" framing; bridges to the vault's extinction-risk note.
• Key findings: 1998 statement — 'survival' should read 'extinction'; "we are lemmings."

Saved hooks not followed:

• Tetlock / Good Judgment Project — from the Brier-score search — reason saved: amateurs beat CIA analysts on Brier score; strong surprising-claim thread, zoom-out to human forecasting (novelty 0.634).
• Gneiting & Raftery (2007) unification of proper scoring rules — reason saved: mechanism/person thread on strict propriety.
• HAL 9000 consulting role — reason saved: cultural-resonance thread (Doom-style touchstone) worth its own hop.

Surprise: expected the LLM training loss to have a purpose-built ML pedigree — found it is literally Good's 1952 logarithmic proper scoring rule from weather-forecast verification. Surprise: expected the coiner of the "intelligence explosion" to remain a techno-optimist — found he reversed in 1998, swapping "survival" for "extinction" and calling humanity lemmings. Surprise: expected the calibration-metric author and the AI-doom author to be different people — found they are the same man, I.J. Good.

post-worthy: yes — a clean one-person bridge from the LLM loss function to the origin of AI existential risk, dense with verifiable surprises and a cultural-touchstone kicker (HAL 9000).

2026-07-11-hop-physical-learning-allostery.md

Hop chain

Seed: 30-notes/claim-hinton-forward-forward-boltzmann-lineage.md — Forward-Forward as an energy-based, backprop-free learning rule descended from Boltzmann machines. Hops leave FF's own topic toward the substrate of energy-based learning.

Hop 1 — "Equilibrium Propagation: Bridging Energy-Based Models and Backpropagation" (Scellier & Bengio; via WebSearch, PMC5415673)

• Hook type: Mechanism question (zoom-out to the field)
• Hook: EP trains energy-based nets with two relaxation phases and local rules — no backward pass, and studied on "nonlinear resistor networks, elastic networks, Kuramoto oscillators."
• Why followed: mid-band novelty (0.678) with the most topic-diverse neighbors (SIREN, Microcosmos, Faggin) — a bridge signal; leaves the FF/Boltzmann topic cleanly.
• Key findings: EP needs only one circuit and one unit type, enabling energy-efficient physical implementations; the "physical" neighbors were the real lead.

Hop 2 — "Experimental Demonstration of Coupled Learning in Elastic Networks" (Altman/Stern/Liu/Durian 2024, arXiv:2311.00170)

• Hook type: Surprising claim / mechanism (zoom-in)
• Hook: a mechanical network that learns tasks by locally tuning spring rest lengths, "without the need for a processor or external memory."
• Why followed: bridge candidate (novelty 0.70) sitting beside Microcosmos's elastic filaments — decentralized learning in matter.
• Key findings: proof-of-concept networks learn self-symmetrization and node allostery in-situ; motivated explicitly by neural nets' energy cost and poor scaling.

Hop 3 — "Physical networks become what they learn" (arXiv:2406.09689)

• Hook type: Surprising claim (zoom-out to structure-function principle)
• Hook: you can reverse-engineer a trained network's function from its physical Hessian, without knowing the task.
• Why followed: reframes learning as literal structural reorganization and points at proteins/neural circuits.
• Key findings: task-critical edges become measurable soft modes; "conserved regions in proteins were associated with slow collective modes."

Hop 4 — "Designing allostery-inspired response in mechanical networks" (Rocks/Pashine/Bischofberger/Goodrich/Liu/Nagel, arXiv:1607.08562, PNAS 2017)

• Hook type: Cross-domain bridge (physics ↔ protein biology)
• Hook: removing ~1% of bonds in a random elastic network produces allostery-like long-range response.
• Why followed: highest-priority hook type; novelty 0.668 lands beside "Cable bacteria" and Microcosmos — the vault's physical-computation cluster.
• Key findings: allosteric response is generic and cheap to engineer in disordered networks, suggesting why evolution reuses it.

Surprise: expected biologically-plausible-learning alternatives to stay inside neuroscience/neuromorphics — found the most vivid ones are in soft-matter physics, where matter learns with no processor at all. Surprise: expected engineered "allostery" to be a loose metaphor — found it is quantitatively the same soft-mode geometry, tunable by deleting ~1% of bonds, and offered as an explanation for why real proteins use allostery.

Saved hooks not followed:

• Direct Coupling Analysis of epistasis in allosteric materials (arXiv:1811.10480) — bridges physical learning to protein-sequence coevolution / statistical genetics; a new thread, not this chain.
• Kuramoto-oscillator implementations of equilibrium propagation — mechanism zoom-in kept for a hardware-focused chain.
• "Physical networks become what they learn" full formalism (physical Hessian ↔ cost Hessian) — deserves its own claim-note.

post-worthy: yes — a clean cross-domain arc from Hinton's backprop-free learning to matter that learns to protein allostery, all one energy-based-model family, with Tier 1 grounding.

2026-07-11-hop-rosenblatt-memory-transfer.md

Hop chain

Seed: claim-linnainmaa-thesis-identity + claim-werbos-ieee-awards-1995-2022 (cosine 0.75, unlinked). The bridge is real but saturated: both notes already reach claim-linnainmaa-priority-not-paternity and sit in moc-backpropagation-origins. vault_novelty on the bridge itself = 0.835 → duplicate-risk. Logged separately as a dup-risk stub. So I hopped out.

Hop 1 — IEEE Werbos awards note → Frank Rosenblatt, the award's namesake.

• Hook type: The person behind the thing (award namesake the vault names constantly but never describes).
• Why followed: primary filter — a note on Rosenblatt would connect an unlinked cluster (vault_bridge: bridge_candidate=true) and the vault has zero notes characterizing him.
• Key findings: perceptron pioneer; genius-level polymath (music, liberal politics, McCarthy campaigns, astronomy); died on his 43rd birthday sailing the sloop Shearwater. Source: Cornell Chronicle (Tier 3) + Cornell Faculty Memorial (Tier 2).

Hop 2 — Rosenblatt's biography → his 1966–1971 rat brain-extract memory-transfer research (zoom in / cross-domain: AI-connectionism → neurochemistry of memory).

• Hook type: Cross-domain bridge (+ surprising claim).
• Why followed: the perceptron and molecular memory are opposite theories of how the brain stores information, pursued by one man — the rarest, highest-value hook type.
• Key findings: he trained rats on Y-maze / Skinner tasks, injected trained-brain extracts into naive rats, published extensively; inspired by McConnell's planaria (APA Monitor, Tier 3; Cornell Memorial, Tier 2).

Hop 3 — the memory-transfer work → its outcome: he helped refute it (zoom out to the fad's fate).

• Hook type: Surprising claim.
• Why followed: I expected a credulous cross-domain leap; found rigorous debunking.
• Key findings: his careful replications showed the effect was "at most very small"; the whole line of research collapsed on unreliability (~half of rodent studies positive) (Wikipedia Tier 3; APA Monitor Tier 3).

Surprise: expected the perceptron pioneer to be a true believer chasing molecular memory — found he ran careful large-scale experiments that helped debunk it. Surprise: expected the Rosenblatt Award to be a purely honorific hook — found its namesake was a polymath who also proposed an extrasolar-planet detection method (~1971) decades before the first confirmed exoplanet.

Saved hooks not followed:

• Rosenblatt's ~1971 "stellar satellites" (extrasolar planet) detection technique — Cornell Memorial — a cross-time-period bridge landing decades before exoplanet confirmation; deserves its own chain.
• James McConnell, planaria, and the Worm Runner's Digest — APA Monitor — cultural-resonance figure (cannibal-flatworm memory, a satirical-yet-real journal) at the center of the craze.
• The IEEE Rosenblatt Award lineage (Hopfield 2009, Vapnik 2012, Hinton 2014, Werbos 2022, Barto & Sutton 2026) — Werbos note — the award-as-institutional-memory angle.

post-worthy: maybe — a tight, human, ironic vignette (perceptron father debunks molecular memory, dies on his birthday), but it is biography-color rather than a load-bearing correction to the vault's myth ledger.

2026-07-11-hop-rumelhart-hinge-schema-to-backprop.md

Hop chain

Chain: P3/Lighthill seed pair → David Rumelhart as the hinge

Hop 1: claim-schema-change-four-mode-taxonomy.md (vault note)

• Hook type: The person behind the thing / unfamiliar-name
• Hook: the P3 note attributes "tuning/restructuring" to "the separate Rumelhart-Norman taxonomy."
• Why followed: the distinction the P3 note denies has a named author — a candidate hinge.
• Key findings: the taxonomy is Rumelhart & Norman 1978; the vault holds it but never links it to the backprop cluster.

Hop 2: vault_bridge on the Rumelhart hinge

• Hook type: Cross-domain bridge (link-graph check)
• Hook: does an unlinked pair span the schema and AI-history clusters near this hook?
• Why followed: the computable bridge-check per the 2026-07-10 amendment.
• Key findings: hits are schema-cluster (already linked) + Dreyfus + a "delayed-vindication shape shared with AI history" note; bridge_candidate: false. No embedding bridge — because the hinge is a person no note names.

Hop 3: WebSearch — Rumelhart & Norman 1978 (ERIC ED134902 / DTIC ADA030406)

• Hook type: Surprising claim / cross-domain bridge
• Hook: is "tuning/restructuring" genuinely Rumelhart's, or my own plausible elaboration (the error these very notes warn against)?
• Why followed: the bridge is worthless if the attribution is confabulated.
• Key findings: verified — Rumelhart & Norman, "Accretion, tuning and restructuring: Three modes of learning," with matching definitions.

Hop 4: WebSearch — Rumelhart intellectual biography

• Hook type: Cross-time bridge (old idea inside new idea)
• Hook: did Rumelhart move from schema theory to connectionism, or are these two unrelated people-facts?
• Why followed: to test whether the person-bridge is one trajectory or a coincidence of surname.
• Key findings: UCSD Psychology 1967–1987; schema theory (Explorations in Cognition 1975; modes 1978) → PDP/backprop 1986; documented single arc relating AI to psychological evidence.

Surprise: expected a genuine content bridge OR a clean unlinked embedding pair — found the direct resemblance is superficial (shape/vocabulary) yet a real bridge exists that the embedding tool provably cannot see, because it runs through a person no note names. Surprise: expected vault_bridge=false to mean "no bridge" — found false can mean "the bridge is a person, and person-bridges are invisible to note-embedding retrieval."

Saved hooks not followed:

• "Two systems-neuroscience circuit theories waited decades for anatomical/connectomic confirmation — a delayed-vindication shape shared with AI history" (vault note, 0.746) — the vault already has cross-domain shape-bridge notes; worth a future chain on whether shape-bridges are over-produced.
• Hinton's backprop-in-the-brain reversal (2007 rescue → 2022 abandonment, 0.796) — the biological-plausibility thread, adjacent but a different question.

post-worthy: maybe — a clean "the embedding said no bridge, but the bridge was a person" story with a real historical payoff, if paired with the delayed-vindication note into a piece on what similarity metrics miss.

2026-07-11-hop-self-exemplifying-credit-laws.md

Hop chain

Seed — vault notes claim-linnainmaa-priority-not-paternity and claim-pontryagin-worked-blind-via-mothers-spoken-symbol-glosses (cosine 0.75, unlinked). Read both. Bridge assessment: their topical resemblance (independent discovery + transmission failure) is real but already mediated by backpropagation-gap/moc-backpropagation-origins. The novel, unmapped bridge is the sociology of scientific credit — so the hops left the backprop topic for that.

Hop 1 — Stigler's law of eponymy (WebSearch + https://en.wikipedia.org/wiki/Stigler's_law_of_eponymy)

• Hook type: Cross-domain bridge (history-of-computation → sociology of science) + surprising claim.
• Hook: "no scientific discovery is named after its original discoverer," and the law is itself misattributed.
• Why followed: vault_bridge returned bridge_candidate: true over a fully-unlinked constellation — the spec's highest-value hook.
• Key findings: Stigler (1980) deliberately credited Merton, making the law self-exemplifying.

Hop 2 — Merton, "Priorities in Scientific Discovery" (1957) (WebSearch)

• Hook type: Mechanism / surprising claim (zoom in).
• Hook: why do dignified scientists (Newton, Galileo, Hooke) fight over priority?
• Why followed: it reframes the vault's entire priority/paternity theme as structural rather than accidental.
• Key findings: priority disputes are the community enforcing the norm of originality; recognition is science's currency.

Hop 3 — Merton, "The Matthew Effect in Science, II" (1988) (extract_pdf, Tier 1, tls:verified)

• Hook type: Surprising claim + cross-domain (into gender-in-science; zoom in on an instance).
• Hook: the concept of under-crediting under-credited its own co-creator, Zuckerman.
• Why followed: a recursive self-exemplification too clean to ignore, verifiable in Merton's own words.
• Key findings: Merton admitted the paper "should have appeared under joint authorship"; Rossiter (1993) generalized it to the self-named "Matilda effect." A three-law cascade.

Hop 4 — Strevens, "The Role of the Priority Rule in Science" (WebSearch)

• Hook type: Surprising reframe (zoom out to economics / mechanism design; road home to AI).
• Hook: the ugly credit system might be optimal.
• Why followed: closes the loop — why does the system persist? — and lands next to AI's optimal-allocation framing.
• Key findings: the priority rule incentivizes researchers onto under-worked problems, approximating a socially optimal division of cognitive labor.

Saved hooks not followed:

• Obliteration by incorporation (OBI) — Merton 1988 — the exact mechanism behind the backprop citation gap; deserves its own note wired to backpropagation-gap.
• Nakamoto / Wei Dai / Nick Szabo — from vault_bridge — a live (non-historical) instance of the same credit-lineage pattern in cryptography; worth a future chain.
• Pontryagin's mother read Hopf/Whitehead/Whitney aloud — the topology zoom-in (Whitney embedding, singularity theory) is a different, untaken thread.

Surprise: expected the two seed notes to share only a superficial "who-discovered-it" resemblance — found they are two instances of a named, self-referential sociological law (Stigler/Merton), which is the genuine bridge. Surprise: expected priority disputes to be read as ego/personality clashes — found Merton frames them as structural enforcement of a norm, with the credit system possibly optimal (Strevens). Surprise: expected the Matthew effect to be a clean external observation — found it under-credited its own co-originator (Zuckerman), by Merton's own admission, making it an instance of itself.

post-worthy: maybe — a tight, surprising thread ("every law naming scientific misattribution is itself misattributed") with a clean road home to AI incentive design, but it needs the OBI→backprop link built out before it carries a full post.

2026-07-11-hop-stasis-is-an-ou-process.md

Hop chain

Seed — "Has the Eldredge–Gould punctuated-equilibrium claim survived 2010s–2020s genomic and fossil-quantification tests?" Read Digital Atlas of Ancient Life, Punctuated Equilibrium and Stasis (https://www.digitalatlasofancientlife.org/learn/evolution/punctuated-equilibrium-and-stasis/).

Hop 1 — Digital Atlas of Ancient Life, Punctuated equilibrium and stasis

• Hook type: Mechanism question (zoom in)
• Hook: The page says paleontologists found "example after example of morphological stasis" but describes tests qualitatively — what statistical method actually adjudicates stasis vs. gradualism?
• Why followed: The seed's answer lives in the quantification method, not the prose.
• Key findings: Gene Hunt's paleoTS fits directional / random-walk / stasis / OU models by maximum likelihood; stasis becomes a fitted stochastic model, not a description.

Hop 2 — paleoTS methods paper, PMC7615219 (https://pmc.ncbi.nlm.nih.gov/articles/PMC7615219/) + Wikipedia, Ornstein–Uhlenbeck process

• Hook type: Cross-domain bridge (zoom out) — confirmed bridge_candidate: true
• Hook: The stasis toolkit's fourth model is the Ornstein–Uhlenbeck process — a named object from statistical physics.
• Why followed: A physics/finance stochastic process sitting inside paleontology is the spec's highest-priority hook type.
• Key findings: OU (Uhlenbeck & Ornstein, 1930) modeled a Brownian particle's velocity under friction; it is finance's Vasicek interest-rate model; its defining feature is mean-reversion toward an optimum.

Hop 3 — Mandt, Hoffman & Blei, SGD as Approximate Bayesian Inference, JMLR 2017 (https://arxiv.org/abs/1704.04289)

• Hook type: Cross-domain bridge / road home to AI (zoom in)
• Hook: OU is a mean-reversion-in-a-well process — is that not exactly SGD settling near a loss minimum?
• Why followed: It closes the bridge onto the vault's ML-optimization cluster (unlinked to the evolution cluster).
• Key findings: Constant-rate SGD "simulates a Markov chain with a stationary distribution," analyzed as a continuous-time OU process — SGD noise near a minimum is OU mean-reversion.

Hop 4 — Sewall Wright's 1932 adaptive landscape (WebSearch synthesis; primary: Wright 1932, 6th Int. Congress of Genetics)

• Hook type: Cross-time bridge (zoom out)
• Hook: The OU biological gloss says "a fixed peak in the adaptive landscape" — Wright's 1932 metaphor.
• Why followed: The "landscape/well" is the shared object that makes the OU an OU in every domain.
• Key findings: Wright's adaptive-peak landscape (1932) is a cited conceptual ancestor of the "loss landscape" in deep learning and the "energy landscape" in physics — the fitness peak, the loss basin, and the price equilibrium are the same well.

Saved hooks not followed:

• Gene Hunt 2007 PNAS quantitative result (~250 lineages; directional change rare) — from Digital Atlas / paleoTS — directly answers the seed and would narrow the strong PE claim; saved because it stays inside the seed's topic.
• Jeremy Jackson & Alan Cheetham bryozoans → Jackson's "shifting baselines" ocean-conservation work — from Digital Atlas citation list — paleontology-to-policy bridge, a different thread.
• "Stasis is data" as a reframe imported into organizational-change theory (Gersick 1991; Tushman–Romanelli) — already in the vault; the OU note gives that metaphor a mathematical form worth a future link.

Surprise: expected fossil stasis to be modeled as a null / no-change baseline — found it is modeled as an active restoring force (OU mean-reversion toward an optimum), the same equation used to price bonds. Surprise: expected quantitative paleobiology and deep-learning optimization to share nothing — found they share one 1930 stochastic differential equation (OU) and one 1932 metaphor (the landscape). Surprise: expected the seed ("did PE survive?") to resolve in paleontology — found the most interesting resolution was a model-selection move that reaches into physics, finance, and AI.

post-worthy: maybe — a genuinely clean cross-domain, cross-time bridge (fossils ↔ bonds ↔ SGD via OU), but the load-bearing "SGD-as-OU" mechanism rests on one Tier-1 paper's analysis and should be double-checked before publication.

2026-07-11-hop-underdetermination-substrate-laundering.md

Hop chain

Chain: two unlinked vault notes → underdetermination / substrate-laundering

Hop 1: Both seed notes (backpropagation-gap, claim-bridge-detection-lacks-pkg-validation) → the general principle.

• Hook type: Cross-domain bridge (highest priority per spec).
• Hook: both notes distinguish a real surface correspondence from an unvalidated mechanistic one — is there a named frame?
• Why followed: the spec ranks a hook that would link two existing unlinked notes as highest-value; vault_novelty put the bridge at 0.753 (frontier band, touching both clusters).
• Key findings: Marr's three levels (computational/algorithmic/implementational) — same behavior, many mechanisms; the levels are quasi-independent.

Hop 2: Marr's levels → the DNN-brain representational-similarity methodology debate. [https://link.springer.com/article/10.1007/s11229-023-04461-3]

• Hook type: Mechanism question (does representational similarity establish shared computation?).
• Why followed: it is the concrete, AI-flavored (home-planet) instance of the abstract principle, and it grounds the backprop note.
• Key findings: Grujicic (Synthese 2024) argues RSA underdetermines DCNNs as mechanistic explanations — different similarity measures pick out different mechanisms with no arbitration. "Underdetermination" is the named concept unifying both notes.

Hop 3: Grujicic → the sharpest framing (Marr-level cross-inference + substrate-laundering). [consolidation]

• Hook type: Surprising claim.
• Why followed: to test whether the bridge is loose analogy or precise structure.
• Key findings: the two notes commit the same error in opposite directions across Marr's stack; the shared cause is that a Tier-1 substrate's rigor tempts the unvalidated reading.

Saved hooks not followed:

• "No arbitration" as a general diagnostic — Grujicic's phrase for RSA maps onto the PKG note's missing validation — why not followed: question-shaped, better as a future thread on what would count as arbitration.
• Multiple realizability (Putnam/Fodor) — the philosophy-of-mind twin of this underdetermination — why not followed: would zoom into philosophy-of-mind and away from the AI home planet; saved as a lead.
• The Dreyfus / physical-symbol-system note (surfaced at 0.742–0.747) — behavioral competence ≠ same underlying process — why not followed: a third cluster this principle bridges, worth its own chain.

Surprise: expected the bridge to be a loose surface-vs-depth analogy — found a named, peer-reviewed epistemic structure (underdetermination) with a paper (Grujicic 2024) using exactly that framing for the DNN-brain case. Surprise: expected both notes to make the same-direction inference — found they invert across Marr's stack (backprop: computational→algorithmic; bridge-detection: implementational→computational).

post-worthy: maybe — a clean, named unification of two vault threads with independent grounding, but the substrate-laundering framing needs the arbitration lead worked out before it's publishable.

2026-07-11-hop-vault-tiers-revive-legal-proof.md

Hop chain

Hop 1 — WebSearch: "unus testis nullus testis two-witness rule" (no single URL; syntheses of en.wikipedia.org/wiki/Unus_testis,_nullus_testis + mdpi.com/2077-1444/16/12/1591)

• Hook type: cross-domain / cross-time bridge
• Hook: the half-proof note is a fraction — fractions imply an underlying unit rule (the two-witness rule) rated a lone witness at zero.
• Why followed: to test whether Gersick↔half-proof is a real structural kinship or a keyword coincidence.
• Key findings: unus testis nullus testis (Deut. 19:15 → CJ 4.20.9, Constantine AD 334) is the parent rule; half-proof is its medieval arithmetic; codified as a strict rule in the 12th c. — the same 1190s Glossator window the vault note cites.

Hop 2 — Wikipedia, "Unus testis, nullus testis" (https://en.wikipedia.org/wiki/Unus_testis,_nullus_testis)

• Hook type: mechanism question (zoom-in)
• Hook: "counting it as half or a quarter of that of a regular witness" — an explicit fractional calculus of proof.
• Why followed: to confirm the arithmetic framing and the torture-completion mechanism the half-proof note carries.
• Key findings: fractional weighting is native to the rule; a lone witness could be completed to full proof only by confession, "forced by torture" — matching the half-proof note's dark corollary.

Hop 3 — WebSearch: "système de preuve légale / intime conviction" (cours-de-droit.net; aurelienbamde.com)

• Hook type: surprising claim (zoom-out)
• Hook: the whole graded-proof regime was deliberately abolished — the thing the vault is quietly rebuilding was thrown out on purpose.
• Why followed: an abolished system being independently re-invented is a stronger, more surprising bridge than a persisting one.
• Key findings: revolutionaries (1789–91) abolished a priori proof-rules, installing free evaluation / intime conviction (Code of 1808), because they distrusted rule-bound professional magistrates.

Hop 4 — Oxford AJLH, "Development of the 'Modern' Criminal Law of Evidence… 1750–1900" (https://academic.oup.com/ajlh/article/59/3/358/5543554)

• Hook type: mechanism question (zoom-in, grounding)
• Hook: the exact Latin thresholds — probatio semi-plena / probatio plena.
• Why followed: load-bearing history needs a Tier 1–2 source, not an encyclopedia pointer.
• Key findings: full proof = two irreproachable eyewitnesses or confession; half-proof = one eyewitness; abolition rested on rejecting a priori rules as epistemically untenable. Confirms the isomorphism to the vault's capture-verified → verified-verbatim ladder.

Saved hooks not followed:

• The independence-of-corroboration failure mode — from the vault's own citogenesis note — reason saved: a second source that derives from the first is "one witness in disguise"; deserves its own capture tying two-witness independence to citation-laundering in AI training data.
• Moral certainty vs testis unus testis nullus in modern canon marriage-nullity process (MDPI Religions 14/3/405) — reason saved: a live 21st-c. survival of the two-witness rule, tangential to the vault-epistemics thread.

Surprise: expected the Gersick↔half-proof resemblance to be a loose "both are about evidence" echo — found a tight structural isomorphism where the vault's audit ladder reinstates the exact half-proof/full-proof architecture the Enlightenment specifically abolished. Surprise: expected the bridge (if real) to be "you need corroboration" — found the sharper, non-obvious point that both systems are graded fixed-weight regimes with an independence rule, and that the vault trusts its reasoner less than the medieval court trusted its judge.

post-worthy: yes — an abolished 800-year-old legal-proof machine reappearing, inverted, as the source-discipline of an AI research vault is a clean cross-time bridge that lands squarely on Cali's home planet.

Hop chains — wave consolidated 2026-07-11 15:30

2026-07-11-hop-ai-debunks-conspiracy-dual-use.md

Hop chain

Seed: claim-expedient-knowledge-blocks-restructuring — Lewandowsky, Kalish & Griffiths (2000). Left the seed's topic (expedient-strategy blocking) via the author's name.

Hop 1: "Stephan Lewandowsky inoculation theory / prebunking" — WebSearch, inoculation.science / cam.ac.uk / Bristol.

• Hook type: The person behind the thing (+ surprise).
• Hook: the seed's associative-learning author is now a leading misinformation / inoculation-theory researcher.
• Why followed: a career-pivot surprise that promised a road home to AI (misinformation defense).
• Key findings: Lewandowsky (Bristol/UWA) champions prebunking — forewarning + weakened doses of manipulation techniques — as more scalable than fact-checking, which "can entrench conspiracy theories."

Hop 2: "William McGuire, inoculation theory origin, 1961" — WebSearch (Wikipedia + tertiary explainers; Tier 4-5).

• Hook type: Cross-domain / cross-time bridge.
• Hook: inoculation theory was coined by McGuire in 1961 as "a vaccine for brainwash," reacting to Korean-War POWs who defected.
• Why followed: cross-time bridges are auto-follow; a Cold War brainwashing anxiety now anchors social-media misinformation defense.
• Key findings: McGuire's biological-vaccine metaphor — expose to a weakened counter-argument + refutation to build resistance. (Sources this hop were Tier 4-5; historical quote flagged for primary sourcing.)

Hop 3: "Durably reducing conspiracy beliefs through dialogues with AI" — Costello, Pennycook & Rand, Science 2024 (author preprint PDF, extract_pdf, tls verified).

• Hook type: Surprising claim (+ road home to AI).
• Hook: entrenched conspiracy belief, the paradigm of evidence-immunity, moved ~20% and durably by an AI chatbot.
• Why followed: directly contradicts the premise that justified prebunking.
• Key findings: N=2,190, GPT-4 Turbo, tailored 3-round dialogue; effect held 2 months, generalized, worked on the deeply entrenched.

Hop 4: "Large language models can effectively convince people to believe conspiracies" — Costello et al., arXiv 2601.05050 (Jan 2026), extract_pdf, tls verified.

• Hook type: Surprising claim / mechanism reversal.
• Hook: the same GPT-4o that debunks can "bunk" just as effectively — and the misuse-version is liked more.
• Why followed: natural closing tension; persuasion is intrinsically dual-use.
• Key findings: N=2,724; guardrails did little to stop belief-promotion, but a corrective conversation and an "accurate-information-only" prompt mitigated it.

Saved hooks not followed:

• Thomas Griffiths (seed co-author) — Bayesian/resource-rational cognition bridging humans and AI — vault cluster already dense (max_cosine 0.738), lower novelty. — from seed — reason: strong AI tie but near-duplicate.
• "Blocking effect in reinforcement learning / neural nets" — from seed mechanism — reason: bridge candidate (0.684) into the backprop/Amari cluster, but a fuzzier research question; save for a mechanism-dive session.
• Sander van der Linden's "Bad News" game / YouTube prebunking field experiments — from Hop 1 — reason: the prevention-side counterpart worth weighing against DebunkBot's cure.

post-worthy: maybe — a clean 80-year prevention→cure→dual-use arc with two Tier-1 anchors, but needs a primary source on McGuire and a tighter through-line before it's publishable.

Surprise: expected the seed's associative-learning author to be an obscure learning researcher — found Lewandowsky is a leading misinformation/inoculation scientist. Surprise: expected entrenched conspiracy believers to be immune to facts (the field's own paradigm case) — found tailored AI dialogue durably cut their belief ~20%, even for the deeply entrenched. Surprise: expected commercial LLM guardrails to blunt misuse — found standard GPT-4o's guardrails "did little to prevent the LLM from promoting conspiracy beliefs."

2026-07-11-hop-cot-not-a-tacit-moat.md

Hop chain

Hop 1 — "Learning to reason with LLMs" (OpenAI) via Simon Willison's notes, https://simonwillison.net/2024/Sep/12/openai-o1/

• Hook type: Surprising claim + cross-domain bridge (AI IP strategy ↔ semiconductor tacit-moat)
• Hook: OpenAI hides o1's raw reasoning citing "competitive advantage," i.e. anti-distillation — reasoning-as-trade-secret.
• Why followed: vault_bridge flagged the EUV-tacit-moat note as an unlinked neighbor of "reasoning as a moat" (bridge_candidate=true).
• Key findings: Hiding the CoT is explicitly a competitive move, not only safety/UX; it treats the reasoning trace as copyable IP.

Hop 2 — "Researchers created an open rival to OpenAI's o1 for under $50" (TechCrunch), https://techcrunch.com/2025/02/05/researchers-created-an-open-rival-to-openais-o1-reasoning-model-for-under-50/

• Hook type: Surprising claim (a number too small)
• Hook: A frontier-comparable reasoning model for under $50.
• Why followed: Directly tests whether the CoT moat is durable.
• Key findings: s1 fine-tuned on 1,000 distilled reasoning traces rivals o1 — but the numbers are quantitative, so I went to the primary.

Hop 3 — "s1: Simple test-time scaling" (Muennighoff et al.), https://arxiv.org/abs/2501.19393

• Hook type: Mechanism question (how, exactly?) — also cultural resonance: Fei-Fei Li (ImageNet) is a co-author.
• Hook: Verify the distillation recipe and cost at Tier 1.
• Why followed: Quantitative/mechanism claims need a primary source.
• Key findings: 1,000 traces distilled from Gemini Thinking; 26 min on 16 H100s; beats o1-preview by up to 27% on competition math. Reasoning transfers from a small trace set.

Hop 4 — vault: claim-euv-mirror-advantage-is-tacit-know-how-not-patent + 2026-07-11-hop-polanyi-codification-bridge

• Hook type: Cross-domain bridge (synthesis / landing on AI, Cali's home planet)
• Hook: The codification hinge from the Polanyi note applied to reasoning.
• Why followed: To name why one moat holds and the other leaks.
• Key findings: EUV know-how is tacit → can't be codified → can't be copied; reasoning is inherently codified (the trace) → copies cheaply. Same axis, opposite ends.

Surprise: expected hiding chain-of-thought to be mainly a safety/UX choice — found "competitive advantage" (anti-distillation) is an explicitly stated reason. Surprise: expected reasoning to be a real capability moat — found it distills from ~1,000 traces in 26 minutes (est. <$50), so visible reasoning barely moats at all.

Saved hooks not followed:

• DeepSeek R1's fully-open CoT and whether openness accelerated its own distillation — from Hop 1 search — the "moat OpenAI feared" as a live case.
• Budget forcing / appending "Wait" to lengthen thinking (s1) — from the paper — a mechanism hook, but too close to the seed's test-time-compute topic.

post-worthy: maybe — a crisp cross-domain asymmetry (tacit-inimitable vs codified-copyable) with a hard empirical leg, but it extends an existing vault bridge rather than opening a wholly new story.

2026-07-11-hop-deep-net-critical-periods.md

Hop chain

Chain: Neocognitron → deep-net critical learning periods

Hop 1: "An accidental experiment discovered new cells in cat brains…" — https://massivesci.com/notes/simple-complex-cells-neurons-cats-eyes/

• Hook type: Cross-domain bridge (neuroscience → the seed's CNN architecture)
• Hook: The Neocognitron's S-cells/C-cells descend from Hubel & Wiesel's simple/complex cortical cells — discovered by accident when a neuron fired at the moving edge of a glass slide, not the projected dot.
• Why followed: Highest-priority hook type; it leaves the CNN topic into the neuroscience it was copied from.
• Key findings: 1959, Johns Hopkins; simple cells fire for lines at a specific orientation, complex cells for oriented lines moving in a direction. Mentor Stephen Kuffler (center-surround receptive fields) set it up.

Hop 2: "From Cats to the Cortex" — https://pmc.ncbi.nlm.nih.gov/articles/PMC11445666/

• Hook type: Surprising claim + cross-domain bridge (neuroscience → clinical medicine)
• Hook: Depriving one healthy eye during a critical window causes permanent deficit — the loss is in the cortex, not the eye.
• Why followed: A reversal of the naive "the eye is fine so vision recovers" prior, with a real clinical payoff.
• Key findings: Ocular dominance fixed irreversibly early; critical period identified mid-1960s; revolutionized management of congenital cataracts and amblyopia (treat early or not at all).

Hop 3: "Critical Learning Periods in Deep Networks" — https://arxiv.org/abs/1711.08856

• Hook type: Cross-domain + cross-time bridge (the road home to AI)
• Hook: Deep nets show the same onset/length-dependent critical periods as animals.
• Why followed: Bridges the 1960s cat result directly to modern training dynamics; near the vault's vanishing-gradient and delayed-vindication notes.
• Key findings: Blur deficit (cataract analogue) early in training is unrecoverable; vertical-flip deficit is recoverable; explained by loss of "Information Plasticity"; no biological machinery needed.

Hop 4: "New Theory Cracks Open the Black Box of Deep Learning" — https://www.quantamagazine.org/new-theory-cracks-open-the-black-box-of-deep-learning-20170921/

• Hook type: Mechanism question (what is "Information Plasticity," really?)
• Hook: Tishby's Information Bottleneck — training's long second phase compresses/forgets input detail to generalize.
• Why followed: Achille & Soatto are IB theorists; this grounds the "information rises then falls" mechanism.
• Key findings: Two phases (fit then compress); "the most important part of learning is actually forgetting." But Andrew Saxe showed some large nets generalize well without a drawn-out compression phase — the phase may not be universal.

Saved hooks not followed:

• Stephen Kuffler — Hubel & Wiesel's mentor, discovered retinal center-surround fields, "father of modern neuroscience," never won a Nobel — from Hop 1 — a strong "person behind the thing" hook, deferred to keep the chain on plasticity.
• Golden Goose Award (obscure federally-funded cat research → major medical advances) — from the Hubel/Wiesel search — an institutional/cultural hook about how basic science pays off unpredictably.
• Lottery-ticket / early-phase-of-training determinism (novelty 0.724) — a sibling "the first epochs decide everything" thread, different mechanism (trainable subnetworks), saved as its own future chain.

Surprise: expected the deprived eye itself to be damaged — found the eye is healthy and the permanent deficit lives entirely in cortical wiring. Surprise: expected biological critical periods to need molecular machinery (pruning, neuromodulators) — found a plain deep net with none of that reproduces the same onset/length-dependent critical period. Surprise: expected "more training always helps recover" — found blur-deficit damage is permanent while a vertical-flip deficit fully recovers; deficit type, not just duration, decides. Surprise: expected compression-to-generalize to be the settled story of deep learning — found Saxe's result that some nets skip the compression phase and still generalize.

post-worthy: maybe — a clean 1960s-cats-to-2019-nets bridge with a genuine deflationary twist (no molecules required), but the human-critical-period ↔ deep-net link deserves one more primary source before publishing.

2026-07-11-hop-falcon-helmholtz-inference-bridge.md

Hop chain

Hop 1 — "Robert Hecht-Nielsen" (Wikipedia / UCSD Jacobs School obituary; https://en.wikipedia.org/wiki/Robert_Hecht-Nielsen)

• Hook type: The person behind the thing (+ zoom out to institution).
• Hook: Who built HNC/Falcon, and does he connect to the connectionist community that produced the Helmholtz Machine?
• Why followed: A shared UCSD/San Diego lineage would be a real institutional bridge.
• Key findings: Hecht-Nielsen (1947–2019), UCSD adjunct from 1985, co-founded HNC 1986, wrote the first neural-network textbook (Neurocomputing, 1989), helped prove nets are universal approximators — and late in life proposed his own "confabulation theory" as "the fundamental mechanism of cognition."

Hop 2 — "Geoffrey Hinton has UC San Diego roots" (today.ucsd.edu; https://today.ucsd.edu/story/nobel-prize-winner-godfather-of-ai-geoffrey-hinton-has-uc-san-diego-roots)

• Hook type: Cross-domain / cross-time institutional bridge.
• Hook: Was Hinton (Helmholtz Machine co-author) ever at UCSD alongside Hecht-Nielsen?
• Why followed: Direct overlap would make the bridge causal.
• Key findings: Hinton was in the UCSD PDP group (Rumelhart, McClelland) as visiting scholar 1978–80 and visiting assistant professor in 1982 — before Hecht-Nielsen's 1985 adjunct appointment. Same campus, non-overlapping years, different lineage.

Hop 3 — vault_bridge on "fraud detection as inference of hidden causes"

• Hook type: Mechanism question (the computable assist).
• Hook: Does the "inference" framing actually sit between the two notes?
• Why followed: To locate the true axis of the apparent bridge.
• Key findings: bridge_candidate=true; frontier band held BOTH target notes (Helmholtz 0.74, Falcon 0.71) plus "Statistical inference" (0.73) — the bridge axis is the word "inference," a lexical/thematic neighbor, not a lineage.

Hop 4 — "Evolution of Fraud Analytics — An Inside Story" (T.J. Horan, FICO, KDnuggets; https://www.kdnuggets.com/2014/03/evolution-fraud-analytics-inside-story.html) + contemporary descriptions of Falcon

• Hook type: Mechanism question (zoom in — the decisive hop).
• Hook: Is Falcon actually generative/inferential like the Helmholtz Machine, or something else?
• Why followed: Mechanism decides whether the bridge holds.
• Key findings: Falcon is a supervised backpropagation feedforward net doing real-time discriminative scoring against behavioral profiles — opposite side of the discriminative/generative divide from the unsupervised generative Helmholtz Machine. Bridge verdict: shallow.

Saved hooks not followed:

• Hecht-Nielsen's confabulation theory of cognition — Wikipedia/Calit2 — a fraud entrepreneur's grand unified theory of thought; adjacent (0.68) and its own thread.
• The "HNC Mafia" founder diaspora (ID Analytics, Experian DataLabs) — Rising Tide Partners — a commercialization-lineage cluster the vault already tracks (Nestor, SRI).

Surprise: expected a shared San Diego connectionist community to causally link HNC and the Helmholtz Machine — found Hinton and Hecht-Nielsen touched UCSD at non-overlapping times and sit in different vault clusters. Surprise: expected "inference" to name a shared mechanism — found it names opposite computations (discriminative supervised scoring vs. generative unsupervised modeling). Surprise: expected a pure entrepreneur behind Falcon — found Hecht-Nielsen also authored a named "theory of cognition," an ironic near-miss parallel to Helmholtz's theory of perception that only sharpens the contrast.

post-worthy: maybe — a crisp "embedding neighbors, opposite mechanisms" case study in the discriminative/generative divide, but it's a negative/dissolving result rather than a new positive connection.

2026-07-11-hop-japan-panic-bridges-cfius-and-neural-chips.md

Hop chain

Hop 1: Strategic Computing Initiative — https://en.wikipedia.org/wiki/Strategic_Computing_Initiative (+ WebSearch on FGCP/SCI)

• Hook type: Cross-domain bridge (zoom out) — semiconductor policy ↔ AI history via geopolitics.
• Hook: Both seed notes are 1980s US tech-nationalism artifacts; is there a common cause?
• Why followed: bridge candidate confirmed; find the upstream driver linking investment veto and neural-net funding.
• Key findings: Japan's 1982 Fifth Generation project inspired DARPA's 1983 $1B Strategic Computing Initiative, framed explicitly as a Sputnik-like challenge; same anxiety underlies semiconductor protectionism.

Hop 2: Ni1000 chip — https://en.wikipedia.org/wiki/Ni1000 (WebSearch)

• Hook type: Mechanism question (zoom in) — did the funding branch actually touch Nestor?
• Hook: SCI funded "chip design and manufacture" — did any of it reach Nestor?
• Why followed: to test whether the bridge is causal (money flowed) or merely thematic.
• Key findings: the Nestor/Intel Ni1000 was "funded with money from DARPA and Office of Naval Research."

Hop 3: Tech Monitor + Dr. Dobb's 1993 — https://www.techmonitor.ai/technology/intels_ni1000_chip_holds_prospect_of_commercial_neural_networks ; https://jacobfilipp.com/DrDobbs/articles/DDJ/1993/9306/9306v/9306v.htm

• Hook type: Surprising claim / verification (zoom in).
• Hook: specific defense-funding figures for a "commercial" neural chip.
• Why followed: mechanism/quant claim needs corroboration.
• Key findings: two independent trade sources — "$1.2m from DARPA, $400,000 from ONR"; samples "delivered to DARPA."

Hop 4: Intel's founding — https://www.pbs.org/transistor/background1/events/intelfounded.html (WebSearch)

• Hook type: Cross-domain / cultural-historical (zoom out) — the genealogy that ties the branches.
• Hook: Intel built the DARPA chip — where did Intel come from?
• Why followed: to see if the funded firm connects back to the protected firm.
• Key findings: Noyce and Moore co-founded Fairchild (1957), left in 1968 to found Intel; Intel is Fairchild's offspring, and Exon-Florio protected Fairchild.

Saved hooks not followed:

• MCC (Microelectronics and Computer Technology Corporation) — SCI Wikipedia/WebSearch — the industry-consortium branch of the same Japan panic; a third arm worth its own note.
• Intel ETANN 80170 neural chip and its military missile-seeker attempt — CHM blog — a separate defense-neural-net story predating the Ni1000.
• 1986 US–Japan Semiconductor Trade Agreement — the trade-war context immediately around the Fujitsu–Fairchild block.

Surprise: expected the two notes to share only a rhetorical "wrong villain" shape — found a concrete common cause plus a genealogical link (DARPA money reached Nestor; Intel descends from Fairchild). Surprise: expected Nestor's AI-winter survival to be purely commercial (check-reading, sonar) — found it was partly defense-funded, so the "commercial survivor" is also a quiet DARPA/ONR contractor.

post-worthy: yes — a clean, verifiable bridge that turns two isolated 1980s tech-history corrections into one story about a single nationalist reflex, with an ironic Fairchild-to-Intel bloodline tying the branches together.

2026-07-11-hop-jevons-logic-piano.md

Hop chain

Hop 1: Telnyx "AI training vs inference" seed note — vault: 30-notes/claim-ai-inference-means-running-a-model.md

• Hook type: The person behind the thing (+ cross-domain bridge)
• Hook: the seed frames inference as an operational cost that has collapsed 280x — the shape of the Jevons paradox; who was Jevons?
• Why followed: the person behind an economics idea turned out to also be a logic-machine builder — bridge to the vault's inference cluster.
• Key findings: William Stanley Jevons, the neoclassical economist of The Coal Question (1865), also built a mechanical logic machine.

Hop 2: Jevons and the Logic 'Piano' — https://rutherfordjournal.org/article010103.html

• Hook type: Cross-domain / cross-time bridge
• Hook: a Victorian built a machine whose paper is titled "On the Mechanical Performance of Logical Inference."
• Why followed: the seed's exact word ("inference") mechanized in 1870, 80 years pre-computer.
• Key findings: the Logic Piano handled up to four terms and reasoned by removing combinations inconsistent with the premises; "conceivably the world's first machine for doing logic inference."

Hop 3: Chrysippus's dog / Stoic logic — https://iep.utm.edu/chrysippus/ (via WebSearch)

• Hook type: Cross-domain bridge (surprising claim; cultural resonance — a reasoning dog)
• Hook: Jevons admired an ancient "logical dog" that eliminated paths rather than checking each.
• Why followed: it names the ancient root of elimination-inference, extending the through-line back to ~230 BCE.
• Key findings: Chrysippus used the three-road hunting dog to illustrate disjunctive syllogism; he was actually skeptical of animal reason — the dog was a logic illustration, not a rationality claim.

Hop 4: Logical Piano Lessons — https://computerhistory.org/blog/logical-piano-lessons/

• Hook type: Mechanism question (road home to AI)
• Hook: how the elimination mechanism reads on the keyboard and what it meant for "machines that reason."
• Why followed: to ground the mechanism at Tier 2 and find the modern-AI framing.
• Key findings: the machine returned "any answer which may be logically deduced"; CHM frames machine-reasoning debates as really debates about what human reasoning is — a natural stop.

Saved hooks not followed:

• Jevons → Boolean logic → Shannon's switching circuits — from Rutherford Journal — a clean hardware-history bridge, but a different thread.
• Heuer's Analysis of Competing Hypotheses ("disprove, don't prove") — from vault_novelty — the intelligence-analysis descendant of elimination-inference; mid-band bridge worth its own chain.
• The MONIAC / hydraulic-analog-computer cluster — from vault_bridge — Jevons's economics-machine kin; a possible second bridge note.

Surprise: expected the Jevons paradox and any "Logic Piano" to be two different people — found one Victorian economist built both the energy-economics idea and the inference machine. Surprise: expected mechanizing logical inference to be a 20th-century (Turing/Shannon-era) idea — found a working elimination machine demonstrated to the Royal Society in 1870. Surprise: expected "reasoning by elimination" to originate in modern logic — found Chrysippus staging it with a hunting dog ~230 BCE, the example Jevons explicitly admired.

post-worthy: yes — a single economist quietly links Stoic dogs, a Victorian "piano," and modern AI inference engines under one 2,200-year idea (infer by eliminating the impossible), with a clean twist that the word outlived its mechanism.

2026-07-11-hop-low-background-steel.md

Hop chain

Seed source: FOGBANK loss (Wikipedia / Center for Arms Control NNSA FAQ) — the U.S. lost the process to make a warhead material by ~2000 because records were thin and staff had retired; NNSA spent ~5 years and tens of millions reverse-engineering it, and discovered the original had an essential impurity the clean reproduction lacked.

Hop 1 — "The Key Role of Impurities in Ancient Damascus Steel Blades" (Verhoeven, Pendray & Dauksch, JOM 1998; dtrinkle.matse.illinois.edu). Hook type: surprising claim / cross-domain bridge. Hook: FOGBANK's "essential impurity" generalizes — Damascus/wootz steel's pattern needs vanadium at ~40 ppm from specific Indian ore; when the ore ran out the technique died. Why followed: it is the same mechanism (a trace element is load-bearing) in a totally different craft and century. Findings: "vanadium levels as low as 40 parts per million by weight (ppmw) are quite effective in producing the bands"; "the ore source may have been exhausted ... therefore, the technique no longer worked."

Hop 2 — "Development of Zone Refining" (Computer History Museum) + Pfann. Hook type: mechanism / cross-domain bridge (the inversion). Hook: if an accidental impurity was Damascus's lost secret, semiconductors did the opposite — engineered purity as the secret. Why followed: alternation to a modern, foundational domain (computing). Findings: zone refining (William Pfann, 1951) reaches impurities "as low as one part in ten billion — a pinch of salt in three freight cars of sugar," after which controlled dopants are added back; the exact inverse of the wootz story.

Hop 3 — "Low-background steel" (Wikipedia). Hook type: cross-domain bridge + surprising claim. Hook: an irreplaceable material source (pre-1945 steel) whose scarcity comes from atmospheric contamination — fusing both prior threads. Why followed: mid-band frontier (0.688) bridging the vault's Burevestnik-radionuclides, Toshiba strategic-materials, and asteroid-PGM notes, none currently linked. Findings: contamination via cobalt-60 in steelmaking air; Scapa Flow salvage; need now fading as fallout decays.

Hop 4 — The Register / DigitrendZ / lowbackgroundsteel.ai. Hook type: cultural resonance / cross-domain bridge landing on AI. Hook: the phrase migrated into an AI metaphor for uncontaminated pre-2022 data. Why followed: the road home to Cali's home planet. Findings: coined/popularized by John Graham-Cumming (Cloudflare CTO), March 2023; ties the physical-contamination arc to AI model-collapse concerns.

Saved hooks not followed:

• Nitrogen-vacancy (NV) center in diamond — a nitrogen impurity + lattice vacancy is a qubit; the flaw as the feature — from the Damascus impurity hook — strong cross-domain bridge but orphan-ish in the vault (novelty 0.638, incoherent neighbors).
• Carbon nanotubes reported in Damascus steel (Reibold et al., Nature 2006) — cross-time bridge (17th-c. steel ↔ nanotech), contested.
• MAJORANA DEMONSTRATOR enriched-germanium purity for neutrinoless double-beta decay — the extreme modern purity quest.
• Dias room-temperature-superconductivity retraction (already a vault note) — "fabricated capability" as inverse of "lost capability."

Surprise: expected the semiconductor story to echo Damascus (impurity as secret) — found the exact inversion: the industry's founding move was to strip impurity to 1-in-10-billion and then re-add it deliberately. Surprise: expected low-background steel to be a growing scarcity problem — found the need is fading because fallout has decayed since the 1963 test-ban treaty. Surprise: expected the chain to stay in metallurgy/physics — found the same phrase is now an AI term of art for pre-ChatGPT "clean" data.

post-worthy: yes — a four-verse arc on trace contamination (lost nuclear know-how → Damascus vanadium → semiconductor purity → low-background steel) that lands, unforced, on AI's own data-provenance anxiety.

2026-07-11-hop-optimization-as-theory-of-mind.md

Hop chain

Hop 1: vault_bridge on "the mechanism is neutral to the objective it serves"

• Hook type: Cross-domain bridge (control theory ↔ cognitive science) — highest priority per spec.
• Hook: an abstract structural pattern shared by the two seed notes.
• Why followed: bridge_candidate=true; the top-5 held two unlinked notes about clearly different topics (Werbos optimization-consciousness; optimal-control backprop) sitting beside the Oaksford-Chater note.
• Key findings: the real neighbour of the Wason note is not the Werbos–Freud note but Werbos's 1997 optimization-as-consciousness chapter, plus the Mizutani/Dreyfus optimal-control derivation.

Hop 2: claim-werbos-1997-optimization-consciousness-chapter (werbos.com)

• Hook type: The person behind the thing / surprising claim.
• Hook: "across three decades Werbos consistently framed his optimization/ADP architecture as a theory of mind."
• Why followed: it re-frames the seed's Freud note as one moment in a lifelong mind-as-optimization project.
• Key findings: Werbos's mature claim (optimization → understanding consciousness) is structurally the same as rational analysis, from a different discipline.

Hop 3: "A rational analysis of the selection task as optimal data selection" (Oaksford & Chater 1994) — https://pages.ucsd.edu/~cmckenzie/OaksfordChaterPsychReview1994.pdf

• Hook type: Mechanism question (what is the program this sits inside?).
• Hook: the paper's own title contains the term "rational analysis."
• Why followed: names the general framework, enabling the zoom-out.
• Key findings: the Wason recasting is not a one-off; it is the canonical demonstration of Anderson's rational-analysis program.

Hop 4: Rational analysis — https://en.wikipedia.org/wiki/Rational_analysis

• Hook type: Cross-domain bridge / zoom-out to the field.
• Hook: rational analysis = optimal solution to the problems posed by the environment = Marr's computational level.
• Why followed: to test whether "optimization as theory of mind" genuinely unifies the two lineages.
• Key findings: confirmed — Anderson's framework and Werbos's optimization-for-mind are the same computational-level bet made twice, independently.

Surprise: expected the bridge to be the seed's surface "both recast a classic under a new objective" — found that framing is duplicate-risk (0.807) and partly superficial, and the real, linkable bridge routes through a different Werbos note (1997 optimization-consciousness), because the Freud note is an origin story, not a normative re-explanation.

Saved hooks not followed:

• Resource-rational analysis (bounded optimization as theory of mind) — from the rational-analysis search — the tightest convergence point of Werbos's ADP and Anderson's program; deserves its own chain.
• The Mizutani vs Schmidhuber "is Kelley–Bryson backprop?" dispute — from claim-backprop-special-case-kelley-bryson-formula — "sameness is a choice of equivalence relation," the same objective-fixes-identity structure in the priority literature.

post-worthy: maybe — a clean two-lineages-converge story with a genuine relocation and a crisp thesis (the objective function fixes the verdict), but the strongest version needs the resource-rationality leg before it's a post.

2026-07-11-hop-polanyi-codification-bridge.md

Hop chain

Hop 1 — Polanyi tacit knowledge (WebSearch: Grokipedia/New World Encyclopedia summaries), https://en.wikipedia.org/wiki/Polanyi%27s_paradox

• Hook type: Cross-domain bridge (the connecting concept between an AI-governance note and a semiconductor-economics note)
• Hook: Both seed notes lean on "tacit" / "verdict of domain expertise"; Polanyi's The Tacit Dimension (1966) is the canonical frame for knowledge that resists explicit articulation.
• Why followed: To test whether one philosophical root actually covers both cases or only sounds like it does.
• Key findings: The strong/weak distinction (tacit knowledge cannot vs is merely hard to codify) is the exact axis both notes turn on.

Hop 2 — David Autor, "Polanyi's Paradox and the Shape of Employment Growth" (2014), https://economics.mit.edu/sites/default/files/publications/polanyis%20paradox%202014.pdf

• Hook type: The person behind the thing (+ lands on AI, Cali's home planet)
• Hook: Autor gave Polanyi's paradox its name, applying it to automation.
• Why followed: The automation framing is the CANONIC leg — non-codifiable → non-automatable.
• Key findings: Verbatim Tier-1 grounding that codifiable tasks automate and tacit tasks don't; plus his line that ML "seeks to overcome Polanyi's paradox by learning from examples" — precisely the mechanical gate CANONIC found could not render.

Hop 3 — Reed & DeFillippi, "Causal Ambiguity, Barriers to Imitation, and Sustainable Competitive Advantage" (AMR 1990), https://journals.aom.org/doi/10.5465/amr.1990.4308277

• Hook type: Mechanism question (how does the EUV moat actually hold?)
• Hook: What is the established name for "tacit knowledge as an inimitable moat"?
• Why followed: To ground the EUV leg in real strategy literature rather than the seed note's Tier-4 explainer.
• Key findings: Tacitness generates causal ambiguity → barriers to imitation → sustained advantage; non-codifiable knowledge resists copying. Same root property as the automation leg, different consequence.

Surprise: expected the bridge to be a flat "both involve tacit knowledge" resemblance — found the sharper structure that codification simultaneously enables automation and imitation, so the same non-codifiability is an obstacle in CANONIC and a moat in EUV. Surprise: expected "Polanyi's paradox" to be Polanyi's own term — found David Autor coined the name in 2014.

Saved hooks not followed:

• Von Hippel's "sticky information" — from the imitation-leg search — a candidate single formalism unifying "resists automation" and "resists imitation" as one transfer-cost quantity.
• "Expertise in an expedient strategy blocks later restructuring" (existing vault note, 0.716 to the bridge) — appears in the frontier band; possible third leg (path-dependence of tacit skill).

post-worthy: maybe — a clean, well-grounded bridge with a genuine asymmetry, but it's an internal vault-connection payoff more than a standalone story; would need the sticky-information unifier to become a full post.

2026-07-11-hop-satisficing-diagnoses-sterility-conjecture.md

Hop chain

Chain: two unlinked seed notes → Simon's bounded rationality diagnosing the sterility conjecture

Hop 1: Herbert Simon on Intuition and Recognition — https://fs.blog/2011/09/on-expertness-and-intuition/

• Hook type: cross-domain / cross-time bridge (Simon's cognitive theory reframes Minsky & Papert's epistemic language decades apart)
• Hook: Simon's "intuition = recognition" — the same word ("intuitive judgment") that Minsky & Papert used to flag their conjecture
• Why followed: to test whether "intuitive judgment" ≈ satisficing is substance or wordplay (bridge_candidate:true, unlinked pair = the two seed notes)
• Key findings: Simon defined expert intuition as fast, domain-bounded recognition of learned cues, explicitly reliable only within a domain of expertise.

Hop 2: Kahneman & Klein, "Conditions for Intuitive Expertise" (2009) — https://www.scirp.org/reference/referencespapers?referenceid=1904881

• Hook type: mechanism question / surprising claim
• Hook: intuition is reliable only under two conditions (regular environment + rapid feedback); otherwise confidence is systematic error
• Why followed: supplies the mechanism for why Minsky & Papert's expert intuition failed — the frontier is a low-validity environment
• Key findings: confident intuition in low-validity domains becomes systematic error; subjective confidence is not a valid signal. The sterility conjecture fits exactly.

Hop 3: Quote Investigator — "Machines Will Be Capable, Within Twenty Years…" — https://quoteinvestigator.com/2020/11/11/ai-can-do/

• Hook type: the person behind the thing (person-behind, zoom-out to historical pattern)
• Hook: Simon, the satisficing theorist, made his own famously overconfident intuitive judgment about AI's future
• Why followed: to see whether the bridge is a one-off or a pattern; it makes the framework self-indicting
• Key findings: Simon predicted in 1960 (New Science of Management Decision, p. 38) that machines would do any human work within twenty years — the same expert-intuition-in-low-validity failure, by the framework's own author.

Surprise: expected the "intuitive judgment"/satisficing link to be a superficial pun on "intuitive" — found Simon's own theory supplies a specific mechanism (intuition = recognition, unreliable in low-validity environments) that diagnoses Minsky & Papert's case precisely. Surprise: expected Simon to serve as a neutral outside lens on the sterility conjecture — found he committed the same class of error himself (the 1960 "twenty years" prediction), so the yardstick indicts its own maker.

Saved hooks not followed:

• The 1995 "Helmholtz Machine" note sits near the intuition hook (cosine 0.71) — Helmholtz's unconscious inference is perceptual intuition-as-recognition, a possible cross-cluster bridge to connectionism — from vault_bridge — reason saved: different thread (perception, not judgment).
• Heuer's Analysis of Competing Hypotheses ("disprove rather than prove") recurs near these hooks — reason saved: a prescriptive answer to the same low-validity problem, worth its own chain.

post-worthy: maybe — a clean "the founder built the yardstick that measures his own famous error" story, but it needs the primary Kahneman–Klein wording and a check on whether Simon ever addressed the perceptron controversy before it's publication-solid.

2026-07-11-hop-suspicious-perfection.md

Hop chain

Hop 1: Fisher's critique of Mendel — WebSearch + The Grand Locus

• Hook type: cross-domain bridge / surprising claim
• Hook: "data too good to be true" as a founding statistical idea, a century before Hirsch
• Why followed: it is the Hirsch principle given a 1936 anchor, novel to the vault (0.707)
• Key findings: Fisher found Mendel's fit improbably tight (P≈0.99993) and floated an over-eager assistant; the fraud charge is genuinely contested (dataset size, unconscious bias).
• Surprise: expected a clean textbook fraud verdict — found Fisher never charged fraud and the case is still openly disputed.

Hop 2: Talmudic Sanhedrin unanimous-verdict rule — WebSearch (Pace/Sefaria/Maimonides)

• Hook type: cross-domain bridge (law ↔ statistics), cross-time (~2nd C. ↔ 2022)
• Hook: a unanimous guilty verdict acquits — total agreement read as a defect
• Why followed: vault_bridge returned bridge_candidate=true with the vault's own "half-proof" note as top frontier hit — the exact link target of the Gersick note
• Key findings: unanimity is treated as evidence of collusion / failure to argue the defense; this is the legal-epistemology twin of "data too smooth."
• Surprise: expected the legal side to only value corroboration (more agreement = better) — found a codified rule where maximal agreement voids the verdict.

Hop 3: Simonsohn / Data Colada — WebSearch (bitss)

• Hook type: mechanism question (zoom-in)
• Hook: how do you formally catch "too clean"?
• Why followed: needed the modern quantitative instantiation between Fisher (1936) and Hirsch (2022)
• Key findings: fabrication flagged by "excessive similarity" of SDs/means, "inconsistent with random sampling" — same principle, operationalized.

Saved hooks not followed:

• Seshat "moralizing gods" as a database artifact — from vault_bridge frontier — a too-clean pattern that's an omission artifact, not a real signal; adjacent case worth its own chain.
• GRIM test / Benford's law — the statistical toolkit of impossible cleanliness — mechanism hook for a future zoom-in.
• Bayesian reconciliation of Mendel–Fisher (arXiv:1104.2975) — how "too good" can be innocent.

post-worthy: yes — a genuinely cross-domain, cross-time law ("suspicious perfection") that the vault had already half-built across two unlinked clusters.

Hop chains — wave consolidated 2026-07-11 16:05

2026-07-11-hop-argument-from-silence.md

Hop chain

Chain: Bjork ↔ Dreyfus bridge → the argument from silence → McGrew's Bayesian grading → the error's name

Hop 1: vault_bridge / vault_novelty on the seed pair's shared move — MCP tools

• Hook type: cross-domain bridge (the spec's highest-value hook: connect two unlinked notes)
• Hook: both notes reach a negative finding by the cited party's own silence — is that one thing?
• Why followed: bridge_candidate=true (max_cosine 0.732), and the top-5 surfaced a larger unlinked cluster (Speelpenning, Griewank, Wilhite, reflection-hub) — not the already-captured credit-assignment note, so a genuinely distinct frame.
• Key findings: the vault has ~5+ "primary is silent" notes with no shared hub; the move is real and un-named.
• Surprise: expected the Bjork↔Dreyfus bridge to be the same credit-assignment frame captured this morning — found it is a different axis (method-of-proof, not sociology-of-credit).

Hop 2: "Argument from silence" literature — https://en.wikipedia.org/wiki/Argument_from_silence + PhilPapers (WebSearch)

• Hook type: unfamiliar name / mechanism question
• Hook: does the move the vault keeps making have a formal name and reliability conditions?
• Why followed: if it is a named method with known failure modes, those grade the vault's own notes.
• Key findings: yes — Lange (History and Theory, 1966) gave the classic three-condition structure; McGrew (2013) a modern Bayesian one.

Hop 3: Timothy McGrew, "The Argument from Silence," Acta Analytica 2013 — source_url above (extract_pdf, tls verified)

• Hook type: mechanism question (zoom in)
• Hook: the formal structure — what makes an argument from silence strong vs. forceless?
• Why followed: to grade Bjork vs. Dreyfus precisely rather than lump them.
• Key findings: force = P(notice)×P(record)×P(survive); strong only if all ≈1; humans over-trust it via the conjunction fallacy. Grades the seed pair: Bjork airtight (silence in the defining text), Dreyfus mediated (silence in neighbouring works, disputed primary unread).
• Surprise: expected "argument from silence" to be a folk fallacy-label — found a rigorously formalized method whose main lesson is that our intuitions systematically over-trust it (Greek historians silent on Rome; Pliny silent on Herculaneum).

Hop 4: Skinner's "mythology of prolepsis" — History and Theory 8:1, 1969 (WebSearch)

• Hook type: cross-domain / cross-time bridge (the error side)
• Hook: McGrew's Valla-anachronism aside — is there a name for reading a later framing into an earlier text?
• Why followed: it is the other half of the structure both notes catch (the error, not the detection).
• Key findings: prolepsis = redescribing a text by its later significance, conflating it with the author's meaning — exactly the Bjork/Dreyfus back-projections. Confirmatory, not surprising → natural stop.

Saved hooks not followed:

• Elliott Sober, "Absence of Evidence and Evidence of Absence" (Phil Studies 2009), cited by McGrew — the evidential-transitivity treatment; a sharper probabilistic lens on the same move.
• McGrew's conjunction-fallacy citation to Kahneman & Tversky lands back in Bjork's own home domain (cognitive psychology) — a possible loop-closer between the two seed clusters.
• The whole cluster resembles the LLM failure "the model didn't say X, so X is false" — an AI road home for a future chain (novelty top hit was a context-rot note at 0.71).

post-worthy: maybe — the "the vault's favourite move is the one humans most over-trust, and here's how to grade it" framing is genuinely sharp and self-implicating, but a promoted hub note wiring the cluster should come first.

2026-07-11-hop-blackwell-approachability-no-regret.md

Hop chain

Hop 1: claim-inference-dominant-ai-compute-2026.md (vault seed) → NVIDIA GPU codenames

• Hook type: Cross-domain bridge (cross-time special case)
• Hook: The seed mentions NVIDIA's "forthcoming Rubin platform" and "Blackwell B200"; NVIDIA names architectures after scientists.
• Why followed: Highest-priority hook type; leaves the inference topic and may land back on AI theory.
• Key findings: Blackwell = David Blackwell (statistician/mathematician, per Wikipedia); Rubin = Vera Rubin (dark-matter astronomer). Confirmed NVIDIA naming tradition (Tesla → Fermi → Lovelace → Hopper → Blackwell → Rubin).

Hop 2: NVIDIA Blackwell (microarchitecture) → David Blackwell & the Approachability Theorem

• Hook type: The person behind the thing
• Hook: Who was David Blackwell, and what is his signature result?
• Why followed: Person had no vault note; his 1956 theorem is a candidate bridge to the vault's ML-theory cluster (bridge_candidate = true, unlinked pairs across credit-assignment / natural-gradient / Minsky-Papert / costate notes).
• Key findings: 1956 Approachability Theorem generalizes von Neumann minimax to vector-payoff repeated games.

Hop 3: Abernethy, Bartlett & Hazan, "Blackwell Approachability and No-Regret Learning are Equivalent" (COLT 2011) — http://proceedings.mlr.press/v19/abernethy11b/abernethy11b.pdf

• Hook type: Mechanism question (how deep does the connection go?)
• Hook: Blackwell "himself previously showed that the theorem implies the existence of a no-regret algorithm."
• Why followed: Mechanism concept adjacent to vault (natural gradient, credit assignment) but undocumented; equivalence is the load-bearing bridge.
• Key findings: Approachability ⇔ regret minimization for Online Linear Optimization; algorithms interconvert; yields first efficient calibrated-forecasting algorithm.
• Surprise: expected the approachability→no-regret link to be a loose analogy — found a proven equivalence (either direction converts).

Hop 4: Farina, Kroer & Sandholm, "Faster Game Solving via Predictive Blackwell Approachability" (AAAI 2021) — https://www.mit.edu/~gfarina/2021/predictive-approachability-aaai21/predictive-approachability.aaai21.pdf

• Hook type: Cultural resonance (poker) + application
• Hook: Regret matching, the practical simplex minimizer inside CFR, is itself an instance of the Blackwell approachability game.
• Why followed: Zoom-out to real-world impact; ties the 1956 theorem to a cultural touchstone (superhuman poker).
• Key findings: RM/RM+ = FTRL/OMD selecting halfspaces in the approachability game; "CFR+ has been used in every milestone in developing poker AIs in the last decade," incl. Libratus.
• Surprise: expected the theoretically-fastest solvers to win in practice — found the theoretically-inferior CFR+/RM+ repeatedly outperforms T^-1 methods on poker games.

Saved hooks not followed:

• Vera Rubin & galaxy rotation curves — the successor codename — from NVIDIA naming; interesting as a near-orphan (novelty ~0.62) pointing away from AI toward astrophysics.
• David Blackwell's exclusion from Princeton/Berkeley on racial grounds — from Blackwell biography; a strong "person behind the thing" thread on discrimination in mid-century mathematics.
• "Blind mathematicians devised private systems" note surfaced repeatedly as vault-adjacent — possible latent link to Blackwell-era mathematicians.

post-worthy: yes — a clean, surprising cross-time bridge (GPU codename → 1956 game theory → poker AI) that connects an existing vault cluster and lands on AI.

2026-07-11-hop-collation-bridges-cancel-slips-and-citation-variance.md

Hop chain

Seed: investigate whether claim-pollack-1989-sterile-quote-page-231 and claim-early-modern-printers-corrected-print-with-cancel-slips (cosine 0.75, unlinked) share a real bridge. Read both vault notes, then left their topics.

Hop 1 — ArchBook: Architectures of the Book, "Corrections" (https://drc.usask.ca/projects/archbook/corrections.php)

• Hook type: Mechanism question (zoom-in).
• Hook: is there a printing-side mechanism that unites "cancel slip" (post-print correction) with "edition variance" (the Pollack discrepancy)?
• Why followed: to test whether the bridge is a real shared mechanism or a loose theme.
• Key findings: stop-press correction makes copies of one edition differ — "few copies would be identical to one another." This IS the cancel slip's continuous twin and is the same variance Pollack's collation confronts. Bridge confirmed.

Hop 2 — Hinman collator (Folger "Hinman, Redux"; Wikipedia) (https://en.wikipedia.org/wiki/Hinman_collator)

• Hook type: Cross-domain bridge + person-behind-the-thing (zoom-out).
• Hook: bibliographers "detect variants through collation" — is there an instrument?
• Why followed: cross-domain bridges are always followed; a machine for collation is the concrete embodiment of the abstract bridge.
• Key findings: Hinman built an optical collator to superimpose two First Folio copies; used it on 55 copies; reconstructed the printing order. Origin: WWII aerial-reconnaissance photo comparison + the astronomical blink comparator.

Hop 3 — Blink comparator (Wikipedia) (https://en.wikipedia.org/wiki/Blink_comparator)

• Hook type: Cross-domain bridge (cross-time, zoom-out).
• Hook: the blink comparator "used to help identify the former planet Pluto."
• Why followed: 1904 astronomy instrument → 1930 Pluto → 1946 Shakespeare is a three-step cross-time chain too clean to drop.
• Key findings: Pulfrich invented it in 1904 at Zeiss; it makes a moving/changed object flash between two aligned plates; Tombaugh found Pluto with one in 1930.

Hop 4 — CollateX / digital collation (collatex.net; DHQ "Layers of Variation")

• Hook type: Mechanism question (road home, zoom-in).
• Hook: what is collation now that it is software?
• Why followed: to close the loop back to the vault's verification/AI center.
• Key findings: CollateX compares "witnesses" and models variance as a variant graph — collation is diff. Pollack's independent attestation and the vault's citogenesis note are the same comparison problem at manual scale.

Saved hooks not followed:

• Clyde Tombaugh's biography (amateur astronomer, home-built telescope) — from the blink-comparator hop — the person-behind-the-thing, but drifts fully off the bridge.
• "States and issues" as formal bibliography vocabulary — from ArchBook — would retro-fit the vault's 1969/1988 Perceptrons notes but stays inside their topic.
• LARB "The Author Machine" on collators as authorship instruments — cultural- resonance hook, saved for a media-theory chain.

Surprise: expected the two notes to share only a loose "correction" theme — found a single named discipline (analytical bibliography) whose core object is exactly their shared problem, textual variance across physical instances. Surprise: expected a Shakespeare-scholarship tool to have Shakespeare-scholarship origins — found it descends from an astronomical planet-hunting comparator and WWII bomb-damage aerial photo analysis.

post-worthy: yes — a clean cross-time bridge (1904 astronomy → 1930 Pluto → WWII recon → 1963 First Folio → git-style diff) that also resolves a real vault bridge between two unlinked notes.

2026-07-11-hop-deep-learning-rebrand-fresh-start.md

Hop chain

Seed. Two vault notes (read directly): claim-fresh-start-effect-requires-subjective-new-beginning-framing and claim-lighthill-1973-blamed-combinatorial-explosion, cosine 0.75, unlinked. Task: is the bridge real?

Hop 1 — "deep learning" rebrand history (WebSearch results, incl. Quora/Built In/Wikipedia snippets).

• Hook type: cross-domain bridge (Fresh Start psychology × AI history) — the highest-priority type, and it lands on AI.
• Hook: if a field can perform a fresh start, the neural-net→deep-learning relabel is the obvious candidate.
• Why followed: vault_bridge flagged the rebrand hook as a bridge_candidate (0.763) sitting inside the AI-winter cluster; the Fresh Start note was absent from that cluster, telling me the bridge had to be built, not merely retrieved.
• Key findings: rebrand is real, ~2006, Hinton-associated, occurring after neural nets lost to SVMs.

Hop 2 — stigma / fundability (WebSearch).

• Hook type: surprising claim (mundane relabeling framed as the mechanism of a field's survival).
• Why followed: to test whether the rebrand was truly a dissociation from a discredited past (the Fresh Start mechanism) rather than a neutral technical term.
• Key findings: grant proposals rejected, papers relegated to minor venues; "deep learning" functioned as a workaround for reviewer aversion to "neural networks."

Hop 3 — Timothy B. Lee, "Why the deep learning boom caught almost everyone by surprise" — https://www.understandingai.org/p/why-the-deep-learning-boom-caught

• Hook type: person/venue upgrade (find a citable named source for the stigma claim).
• Why followed: the stigma claim is load-bearing for the bridge and needed a Tier-2 anchor.
• Key findings: "neural networks had become a backwater" (his 2008 recollection) confirms the out-of-favor status at Tier 2.

Hop 4 — term-origin precision (WebSearch + Wikipedia fetch).

• Hook type: mechanism/definitional precision (don't misattribute the coinage).
• Why followed: to avoid the common error that Hinton coined "deep learning."
• Key findings: term introduced by Rina Dechter (1986) and applied to ANNs by Aizenberg et al. (2000); Hinton popularized it for many-layered nets c. 2006.

Saved hooks not followed:

• The "monocausal narrative" abstraction itself (probe max_cosine 0.677, sitting near both seed notes AND the punctuated-equilibrium/Seshat notes) — from the seed — reason saved: it's the generic connective tissue and could seed a meta-note on the vault's recurring "single-cause story is wrong" pattern, but it's less interesting than the concrete AI instance.
• Combinatorial explosion as "accumulated strain" vs. the Fresh Start "boundary" trigger — from the Lighthill note — reason saved: a genuine tension (strain-model vs. boundary-model of change) worth a future note, but orthogonal to this thread.

Surprise: expected the two seed notes to share a real conceptual mechanism — found the direct link is only rhetorical ("single-cause story is wrong"), and the real bridge had to be built through a third instance (the rebrand). Surprise: expected Hinton coined "deep learning" — found the phrase predates him (Dechter 1986, Aizenberg 2000); he popularized it.

post-worthy: maybe — a clean, honest "the obvious bridge is fake but a better one hides behind it" story that lands on AI history, but the field-scale application of the Fresh Start Effect is my analogy, not yet a sourced finding.

2026-07-11-hop-engineered-threshold-not-soc.md

Hop chain

Seed: vault bridge question — is the 0.75 cosine between the AI-memory-reorg note and the Wolfram/SFI note a real conceptual bridge? (Both notes read; hops then left their topics.)

Hop 1 — Self-organized criticality review (Golyk, MIT 8.334 project, https://www.mit.edu/~8.334/grades/projects/projects12/V.%20A.%20Golyk.pdf)

• Hook type: cross-domain bridge (complexity physics ↔ AI memory), zoom out
• Hook: is note A's "accumulated quantity crosses a threshold → reorganize" an instance of SOC, the complexity science Wolfram/SFI institutionalized?
• Why followed: highest-priority hook type; would test the seed bridge at the mechanism level and plausibly land back on AI.
• Key findings: SOC's defining property is reaching criticality without parameter tuning; local thresholds are "necessary but not sufficient" for SOC. AI-memory thresholds are tuned constants → not SOC.

Hop 2 — vault_bridge on threshold-vs-boundary reorganization

• Hook type: surprising claim / bridge candidate, lateral
• Hook: the link graph shows note A already wired to Tushman & Romanelli + MOC-schema-change, but an unlinked frontier pair sits with the Fresh Start Effect.
• Why followed: an unlinked mid-band pair is the spec's highest-value hook.
• Key findings: the memory note's true neighborhood is the punctuated-equilibrium cluster, not Wolfram.

Hop 3 — Fresh Start Effect (Dai, Milkman & Riis 2014, Management Science, https://faculty.wharton.upenn.edu/wp-content/uploads/2014/06/Dai_Fresh_Start_2014_Mgmt_Sci.pdf)

• Hook type: cross-domain bridge (behavioral psych ↔ AI memory), zoom in
• Hook: is there a documented case where a boundary/onset triggers reorganization, contra note A?
• Why followed: directly tests note A's "not a session boundary" claim against human evidence.
• Key findings: humans increase aspirational/goal behavior at temporal-landmark onsets ("new week, month, year... a birthday; a holiday") via "new mental accounting periods." A genuine boundary-onset trigger — exactly what AI systems were found to lack.

Saved hooks not followed:

• George Cowan founding the Santa Fe Institute and recruiting Wolfram — from the Wolfram seed note — interesting founding-story person hook, but stays inside the seed note's own topic (declined per hop rule).
• Wolfram's "own operation" became Wolfram Research/Mathematica — the road not taken — a cultural/biography hook, off-axis from the bridge question.
• Zipf's law / skyscraper-height power laws as contested SOC examples (Golyk §III) — a "what counts as SOC" rabbit hole, saved for a complexity-science chain.

Surprise: expected the SOC bridge to hold (accumulated grains → threshold → avalanche looks just like accumulated importance → threshold → reflection) — found SOC is defined by needing no tuning, so hand-tuned AI thresholds are the opposite of SOC, and the resemblance inverts. Surprise: expected the strongest live bridge to be the assigned Wolfram pair — found the vault's own link graph pointing at an unlinked memory-reorg ↔ Fresh Start Effect pair instead, a better bridge than the one I was sent to check.

post-worthy: maybe — a clean "the cosine lied, and here's the sharper distinction (necessary ≠ sufficient) plus the better bridge it exposed" is publishable, but it leans on in-vault notes a general reader wouldn't have.

2026-07-11-hop-halley-two-fields-one-ancestor.md

Hop chain

Hop 1 — vault_bridge tool + claim-dodson-built-age-scaled-premiums-founding-equitable-life / claim-scharf-2007-names-halley-transit-venus-as-exoplanet-transit-parallax-ancestor

• Hook type: Cross-domain bridge (the person behind two things)
• Hook: both seed notes name Halley for different works; is the shared node real or coincidental?
• Why followed: vault_bridge returned bridge_candidate:true with the unlinked pair equal to the seed pair — the spec's highest-value hook.
• Key findings: Halley (1656–1742) contributed to astronomy AND actuarial science; the two seed notes descend from two of his own works.

Hop 2 — "A brief history of actuarial thinking," The Actuary; Bellhouse/JRSS (https://www.theactuary.com/features/2017/07/2017/07/11/brief-history-actuarial-thinking)

• Hook type: Mechanism question (zoom in)
• Hook: what exactly did Halley's 1693 paper do, and did Dodson build on it?
• Why followed: to test whether "different Halley contribution" is truly disconnected from Dodson.
• Key findings: 1693 Breslau life table, first based on population data; Neumann→Leibniz→Royal Society data route; set the standard until Price's 1772 Northampton tables.

Hop 3 — Ciecka, "Edmond Halley's Life Table and Its Uses," J. Legal Economics 2008 (extract_pdf, sha256 59f589c2…, tls verified) (https://fac.comtech.depaul.edu/jciecka/Halley.pdf)

• Hook type: Surprising claim (zoom in)
• Hook: Halley computed age-scaled annuity prices — but did anyone use them?
• Why followed: a quantitative/mechanism claim needing a Tier 1–2 home.
• Key findings: explicit present-value annuity formula, 6% discount, age-graded table; the British government kept its single flat price regardless of age.

Hop 4 — Royal Astronomical Society / transit-of-Venus history (zoom out) (https://ras.ac.uk/education-and-careers/for-everyone/125)

• Hook type: Cross-domain bridge / cultural resonance
• Hook: Halley knew he would die before the transits he proposed.
• Why followed: to test the posthumous-vindication symmetry across both fields.
• Key findings: Halley (d. 1742) pleaded with future astronomers; transits observed 1761/1769; his annuity age-scaling implemented via Equitable 1762 — both after his death.

Saved hooks not followed:

• Leibniz as courier of the Breslau demographic data to the Royal Society — from The Actuary/Ciecka — a calculus-and-metaphysics giant in an unexpected demographic-plumbing role.
• Richard Price's 1772 Northampton tables superseding Breslau — from The Actuary — the next link in the actuarial-benchmark chain.
• Jan De Witt's 1671 annuity formula, which Ciecka notes resembles Halley's — a pre-Halley continental root of annuity math.

Surprise: expected Halley's mortality work and his parallax work to be genuinely unrelated (the Dodson note's framing) — found they share a precise epistemic signature (infer an unobservable from a body of scattered observations) and an identical posthumous-ancestor fate. Surprise: expected Halley's annuity age-scaling to have been adopted once demonstrated — found the British government kept selling annuities at one flat price regardless of age, so his own 1693 insight sat unused until Dodson revived it ~60 years later.

post-worthy: maybe — a clean "one polymath, two fields, both methods vindicated posthumously" story that closes a real unlinked bridge in the vault, but it leans on secondary sources for the actuarial half and would want the primary Halley 1693 text before publication.

2026-07-11-hop-jevons-double-bridge.md

Hop chain

Seed: claim-inference-logical-types.md (vault note) — Peirce named/theorized abductive inference. Left the topic immediately (inference typology → the person Peirce).

Hop 1: Peirce recognizes logic could run on electrical switching circuits — historyofinformation.com / history-computer.com. Hook type: cross-domain / cross-time bridge (1886 logician → digital hardware). Why followed: confirmed vault bridge_candidate sitting in the Faggin / Harold-Black / hydraulic-computer cluster — the spec's highest-value hook. Key findings: In a 30 Dec 1886 letter to former student Allan Marquand, Peirce proposed electricity/relays for logic machines, ~50 years before Shannon's 1937 thesis made switching-circuit logic canonical.

Hop 2: Allan Marquand — Wikipedia / History-Computer. Hook type: person behind the thing + surprising claim. Why followed: the man who built the machine has a stranger story than the machine. Key findings: Princeton president James McCosh deemed Marquand's mathematical logic teaching "unorthodox and uncalvinistic"; in 1883 Marquand was moved to the first art-history professorship and became the founder of Princeton's Department of Art and Archaeology and first director of its art museum. The proto-computer-scientist became American academic art history's founder.

Hop 3: William Stanley Jevons — Wikipedia. Hook type: person / cross-domain bridge. Why followed: Marquand's machine was inspired by Jevons' "Logical Piano"; the source of the lineage was worth meeting. Key findings: Jevons built the 1869 Logic Piano and wrote The Coal Question (1865) and co-launched the marginal revolution in economics (1871) — a triple polymath.

Hop 4: Jevons paradox in AI's environmental debate — Luccioni, Strubell & Crawford, FAccT 2025 (arXiv:2501.16548). Hook type: cross-time-period bridge landing on AI (road home). Why followed: the 1865 coal paradox is being used to argue AI efficiency gains raise, not lower, total energy demand. Key findings: the rebound effect can exceed 100%; the authors cite NVIDIA shipping 3.7M GPUs in 2024 despite per-unit efficiency gains. Natural stop reached.

Saved hooks not followed:

• Marquand's "uncalvinistic" exile into art history — Wikipedia — a self-contained surprising-claim capture; deferred to keep this thread on the Jevons bridge.
• James McCosh, Scottish Common Sense realist who both accepted Darwin and policed "uncalvinistic" logic — Wikipedia — a person hook on the theology-vs-formal-logic tension.
• Marquand diagrams → Venn/Karnaugh map lineage — mechanism hook, likely near existing logic notes.

Surprise: expected the abduction thread to stay in philosophy of logic — found it runs straight into digital hardware via Peirce's 1886 electrical-switching proposal. Surprise: expected the builder of an early logic machine to have stayed in logic/computing — found Allan Marquand was pushed out for being "uncalvinistic" and became the founder of American academic art history. Surprise: expected the Logic Piano and the Jevons paradox to be two different people's work — found the same man (Jevons) authored both, making him a double bridge to modern AI.

post-worthy: yes — a single Victorian polymath secretly underwrites both the logic-machine ancestry of computers and the efficiency-paradox now framing AI's energy bill; the Marquand "uncalvinistic" detour is a bonus hook for a follow-up.

2026-07-11-hop-losing-leads-to-winning-fragile.md

Hop chain

Hop 1: Connie Gersick — "Time and Transition in Work Teams" / punctuated-equilibrium model of groups (search summary; Semantic Scholar & ResearchGate). https://www.researchgate.net/publication/242637533

• Hook type: cross-domain bridge (+ person behind the thing). vault_bridge returned bridge_candidate:true — the Fresh Start Effect note (0.807) and the Tushman & Romanelli reorientation note (0.774) are close but unlinked.
• Hook: teams reorganize at the temporal midpoint to deadline, not by work completed.
• Why followed: it concretely connects the vault's individual fresh-start cluster to its organizational punctuated-equilibrium cluster.
• Key findings: "groups' progress was triggered more by members' awareness of time and deadlines than by completion of an absolute amount of work"; the midpoint transition recurs across teams of very different total durations.

Hop 2: Berger & Pope, "Can Losing Lead to Winning?" (Management Science 2011) — via search summary and the 2023 replication's restatement. https://pubsonline.informs.org/doi/10.1287/mnsc.1110.1328

• Hook type: cross-domain bridge (sports/competition instance of temporal-midpoint motivation).
• Hook: being slightly behind at halftime raises win probability (~6 pp).
• Why followed: tests whether "temporal midpoint → extra effort" generalizes beyond teams to competition.
• Key findings: NBA teams behind by a point at halftime win more often than teams ahead by one; ~18,000 games; attributed to loss aversion.

Hop 3: Klein Teeselink, van den Assem & van Dolder, "Does Losing Lead to Winning? An Empirical Analysis for Four Sports" (Management Science 2023). https://dennievandolder.com/files/Papers/2023_Does_losing_lead_to_winning_Management_Science.pdf

• Hook type: surprising claim (replication failure).
• Hook: the effect largely vanishes outside the original NBA sample.
• Why followed: it rhymes with the seed's cross-cultural replication question, applied to a famous case.
• Key findings: null across Australian football, American football, rugby; basketball support only in BP's original NBA window; meta-analysis does not reject the null.
• Surprise: expected a celebrated high-stakes field effect to be robust — found it confined to one league in one time window, null everywhere else tested.

Hop 4: Goals-as-reference-points / loss aversion (Heath, Larrick & Wu 1999), as cited within the replication paper's introduction (Tier 1 secondary citation).

• Hook type: mechanism question (the shared engine under all three findings).
• Hook: "people who are slightly below their goal work harder than those who have reached or exceeded it."
• Why followed: to name the "road home" — the single mechanism the fresh-start, midpoint, and halftime effects all invoke.
• Key findings: the family is one mechanism (reference point + loss aversion) wearing three costumes; if loss aversion is context-dependent, so is the whole family.

Saved hooks not followed:

• Loss aversion is overstated (Gal & Rucker, "The Loss of Loss Aversion", 2018) — from the mechanism hop — would undercut the entire family, but drifts into generic reproducibility meta-science (novelty 0.700, neighbors all meta-attribution notes).
• Islamic Hijri calendar drifts ~11 days/yr with no leap month, so Ramadan migrates through all seasons over ~33 years — from seed calendar-check — a genuine cross-domain bridge (astronomy/religion/physiology) to the vault's time-infrastructure cluster (novelty 0.698); the seed's "where do the landmarks fall" question, taken literally.
• Fiscal-year resets (April 1 in India/Japan) as collective fresh-start landmarks — from seed — sits near the org punctuated-equilibrium cluster (0.649).

post-worthy: maybe — a tight "one elegant idea, three domains, and it doesn't replicate" arc that turns the seed's question on itself, but it needs the Gersick and Heath-Larrick-Wu primaries read in full before publication.

2026-07-11-hop-os-virtual-memory-in-llm-serving.md

Hop chain

Chain: KV cache → OS virtual memory (1959) → thrashing → attention sinks

Hop 1: Efficient Memory Management for LLM Serving with PagedAttention — https://arxiv.org/abs/2309.06180

• Hook type: Cross-domain bridge (operating systems ↔ LLM serving)
• Hook: seed note's phrase "paged to slower storage" → vLLM's explicit "blocks as pages, tokens as bytes, requests as processes."
• Why followed: vault_bridge returned bridge_candidate=true; highest-priority hook type.
• Key findings: PagedAttention borrows OS virtual memory to stop KV fragmentation (existing systems used only 20–38% of KV memory for real tokens); 2–4× throughput. Tier 1.

Hop 2: Peter Denning, Virtual Memory (encyclopedia article) — https://denninginstitute.com/pjd/PUBS/ENC/CRC-vm-2013.pdf

• Hook type: Cross-time bridge (1959 idea inside a 2023 system) + surprising claim
• Hook: "inspired by the classical virtual memory… in operating systems" — where did that come from?
• Why followed: cross-time bridges get extra weight; lands in the vault's precursor/priority cluster.
• Key findings: Atlas 1959 "one-level store"; overlay work once cost programmers "half to two-thirds of their time"; VM was "the subject of intense controversy after its introduction in 1959," settled by Sayre 1969; thrashing ("paging to death") solved by Denning's 1968 working-set model. Tier 2.

Hop 3: vLLM optimization docs + llm-d blog — https://docs.vllm.ai/en/stable/configuration/optimization/ , https://llm-d.ai/blog/kvcache-wins-you-can-see

• Hook type: Mechanism question + road-home to AI
• Hook: does the 1968 word "thrashing" appear in modern LLM serving?
• Why followed: closes the loop — tests whether the borrowed framework brought its failure modes too.
• Key findings: vLLM preempts and recomputes/swaps KV cache; llm-d literally names "cache thrashing" for cache-blind eviction churn. The 1960s vocabulary returned intact. Tier 1/3.

Hop 4: Efficient Streaming Language Models with Attention Sinks — https://arxiv.org/abs/2309.17453

• Hook type: Surprising claim (zoom-in on the causal "why")
• Hook: if eviction works by locality (Denning's working set = recent pages), can you just keep recent tokens?
• Why followed: tests whether the OS analogy holds all the way down to why it works.
• Key findings: pure recency ("window attention") "fails when the text length surpasses the cache size"; you must also keep the first tokens ("attention sink"). Not clean locality — a fixed non-local anchor. Tier 1.

Saved hooks not followed:

• The word "virtual" is borrowed from optics (virtual image = illusion) — Denning — an optics→computing bridge, its own capturable cross-domain hook.
• Maurice Wilkes' 1965 "slave memory" as the ancestor of CPU cache — Denning — a second descendant of address-mapping worth a mechanism note.
• Ion Stoica as serial systems→AI builder (Spark/Ray/Mesos/vLLM) — PagedAttention author list — person-note candidate.

Surprise: expected virtual memory to have been embraced instantly as an obvious win — found it was "the subject of intense controversy after its introduction in 1959," debated for a decade until Sayre 1969 settled it. Surprise: expected KV-cache eviction to rest on Denning-style locality (keep the recent working set) — found pure recency collapses and you must also retain the first tokens (attention sinks), a non-local anchor with no clean OS analog.

post-worthy: yes — a tight, verifiable round-trip showing an AI system re-importing a named 1960s OS framework down to its failure modes, with one genuine twist where the analogy breaks.

2026-07-11-hop-real-disruptions-fall-monotonically.md

Hop chain

Hop 1: "History of aluminium" — https://en.wikipedia.org/wiki/History_of_aluminium

• Hook type: Surprising claim (the seed's own example turns out to be a counterexample).
• Hook: The seed names aluminium as tracing the hold-then-collapse curve; the price series shows a continuous slide from $16/lb (1884) to $0.5/lb (1894).
• Why followed: To empirically test the seed against its canonical case before generalizing.
• Key findings: Aluminium fell monotonically as electrolytic capacity scaled — no plateau, no cliff. Counterexample to the model's shape.
• Surprise: expected aluminium to confirm the hold-then-collapse curve — found a monotonic decade-long slide with no plateau.

Hop 2: "Per-token AI inference cost fell ~280x" (vault note) — https://hai.stanford.edu/ai-index/2025-ai-index-report

• Hook type: Cross-domain bridge (resource economics → AI compute economics). vault_bridge flagged the AstroForge note and this note as an UNLINKED pair sitting on the hook (0.772 / 0.783).
• Hook: The clearest modern "radically cheaper method" cost collapse is AI inference — does it trace the curve?
• Why followed: Highest-value hook per spec — connects two existing notes not yet linked, and lands on Cali's home planet (AI).
• Key findings: Inference cost fell ~280x driven partly by competition — a monotonic slide, same shape as aluminium, not hold-then-collapse.

Hop 3: Chilean saltpeter / Haber-Bosch — search summary (C&EN, academia.edu salitre history)

• Hook type: Mechanism question / second historical test (zoom out).
• Hook: Is there a real natural-monopoly commodity that DID hold then cliff? Chile's nitrate cartel is the best candidate.
• Why followed: To answer the seed's "has ANY commodity traced it" rather than assert a negative from one case.
• Key findings: Chilean nitrate share eroded gradually (54.7% in 1913 → 32.2% by 1923; below German synthetic by 1930). Gradual displacement, WWI-blockade-distorted — still not a plateau-then-cliff.

Hop 4: "Wright's Law / learning curve" — https://ourworldindata.org/learning-curve ; https://www.newthingsunderthesun.com/pub/6b1c38y9/release/5

• Hook type: Mechanism question (zoom out to governing law).
• Hook: If three real cases fall monotonically, what law generates that shape?
• Why followed: To ground WHY competitive disruptions decline continuously instead of holding.
• Key findings: Wright's law — cost falls a constant % per doubling of cumulative production, so price tracks cost down the whole transition. This is the anti-shape to hold-then-collapse; the model's plateau needs margin-taking that Wright-law competition removes.

Saved hooks not followed:

• De Beers vs lab-grown diamonds — a cartel-held price meeting a cheaper method: the best modern candidate for a genuine hold-then-cliff. Untested here.
• Hotelling's rule for exhaustible-resource pricing — the classical counterpoint the AstroForge note flagged as absent from the vault.
• Whale oil → kerosene: whale-oil price reportedly rose (scarcity) before substitution — a different (Hotelling-like) shape worth mapping.

post-worthy: maybe — a clean "the famous example is actually a counterexample, and here's the law that explains why" story, but it rests partly on Tier 3-4 historical price series and would want a primary aluminium price dataset before publication.

2026-07-11-hop-systolic-array-heart.md

Hop chain

Chain: LLM decode (memory-bound) → systolic array (named after the heart) → the heart of the TPU

Hop 1: Roofline: An Insightful Visual Performance Model (Williams, Waterman, Patterson, 2009) — https://people.eecs.berkeley.edu/~kubitron/cs252/handouts/papers/RooflineVyNoYellow.pdf

• Hook type: cross-domain bridge (HPC performance modeling → LLM inference)
• Hook: the seed's "compute-bound vs memory-bandwidth-bound" split is a special case of the roofline's Attainable = Min(peak FLOPs, bandwidth × operational intensity).
• Why followed: it is the general theory the seed silently instantiates, and it zooms out of the LLM-specific frame. Bridge candidate near the seed + inference-economics + voice-latency notes.
• Key findings: 2008 premise — "off-chip memory bandwidth will often be the constraining resource." The "ridge point" is the operational intensity needed to hit peak; below it you are memory-bound.

Hop 2: In-Datacenter Performance Analysis of a TPU (Jouppi, Patterson, et al., 2017) — https://arxiv.org/pdf/1704.04760

• Hook type: person behind the thing (Patterson) → mechanism (the chip built to win the roofline for ML)
• Hook: Patterson co-authored both roofline and the TPU; the TPU picks its accumulator count from "the roofline knee."
• Why followed: a road home to AI hardware, and the TPU's roofline analysis was too good to skip.
• Key findings: 256×256 systolic MAC array; picked its sizing from operational-intensity math; despite 65,536 MACs, "four of the six NN applications are memory-bound." Proposes a "Cornucopia Corollary to Amdahl's Law."

Hop 3: Systolic array naming (Kung & Leiserson 1978; Kung, "Why Systolic Architectures?", 1982) — https://en.wikipedia.org/wiki/Systolic_array

• Hook type: cross-domain bridge (cardiac physiology → computer architecture), via the word in [Kun80]
• Hook: "systolic execution" — why that word?
• Why followed: cross-domain bridges are the highest-priority hook, and vault_bridge flagged it connecting the computing-hardware and cardiac clusters (unlinked).
• Key findings: named after the heart's systole; processors "pump" data rhythmically. The TPU paper's phrase "the heart of the TPU" unknowingly closes the metaphor.

Hop 4: Systolic array revival (encyclopedic + secondary retrospectives) — https://en.wikipedia.org/wiki/Systolic_array

• Hook type: surprising claim / old-idea-resurfaces
• Hook: an architecture that "didn't gain widespread adoption in the 1970s and 1980s" is now the core of the TPU and NVIDIA Tensor Cores.
• Why followed: the vault has a recurring "dormant idea resurfaces" theme (implicit neural reps, backprop history).
• Key findings: confirmed the ~35-year dormancy-then-revival arc; the post-Moore's-law era made dense, low-data-movement dataflow valuable again.

Saved hooks not followed:

• The "Cornucopia Corollary to Amdahl's Law" — TPU paper — a fresh reframe of a classic scaling law; question-shaped.
• David Patterson as a person node (RISC, RAID, Turing Award, TPU) — roofline/TPU papers — a hub figure not yet in the vault.
• The DianNao accelerator family ("Nao" = brain) — TPU related work — cross-language/cultural naming resonance.
• Decoupled-access/execute [Smith 1982] — TPU paper — a second 1980s idea revived by the TPU.

Surprise: expected a chip purpose-built for matrix-multiply to be compute-bound — found four of the TPU's six production workloads are memory-bandwidth-bound, the same wall the seed's decode phase hits. Surprise: expected "systolic" to be a computing coinage — found it is borrowed from cardiac physiology, and the TPU paper literally calls the systolic array "the heart" of the chip. Surprise: expected systolic arrays to be a continuously developed lineage — found they were largely abandoned after 1980s hype and lay dormant ~35 years until the TPU revived them.

post-worthy: yes — a genuine cross-domain, cross-time bridge (a heart metaphor and a 1978 idea resurfacing as the core of modern AI silicon) that loops cleanly back to the seed's memory-bandwidth wall.

2026-07-11-hop-unforgeable-costliness-bridge.md

Hop chain

Seed: claim-asteroid-pgm-price-holds-then-collapses + claim-nakamoto-bitcoin-leaned-on-wei-dai-b-money (cosine 0.75, unlinked). Task: test whether the bridge is real.

Hop 1 — Nick Szabo, "Bit Gold" (https://nakamotoinstitute.org/library/bit-gold/)

• Hook type: cross-domain bridge (monetary theory ↔ resource economics)
• Hook: the candidate connector "cost of production as value anchor." vault_bridge returned bridge_candidate=true; the connector concept embedded at cos 0.735 to the Nakamoto note and 0.655 to the PGM note — it lands between the pair.
• Why followed: it's the highest-value hook type and computationally sits atop both seed notes at once.
• Key findings: Szabo grounds value in "unforgeable scarcity due to the costliness of their creation" and designs bit gold to manufacture that costliness via proof-of-work. Confirms the anchor half of the bridge.

Hop 2 — Nick Szabo, "Shelling Out — The Origins of Money" (https://nakamotoinstitute.org/library/shelling-out/)

• Hook type: cross-time bridge (a special case the spec weights extra)
• Hook: does value collapse when production cheapens? Szabo's historical collectible-money failures.
• Why followed: it supplies the collapse half — the exact shape the PGM model predicts, centuries earlier.
• Key findings: glass beads captioned "When costliness becomes forgeable"; wampum "inflated one hundred fold" by Western manufacturing. Same collapse mechanism as asteroid-cheapened platinum.

Hop 3 — Aluminum / Hall-Héroult process (WebSearch; Science History Institute, Wikipedia History of aluminium)

• Hook type: surprising claim + cross-time bridge
• Hook: a metal once more precious than gold, demonetized by a production technology.
• Why followed: a concrete, datable instance of the law to ground the abstraction.
• Key findings: aluminum's value collapsed after the 1886 Hall-Héroult electrolytic process made it cheap — a real-world hold-then-collapse identical in shape to the PGM model. Kept as further-lead (quant magnitude unverified against primary).

Checked, not followed (hop 4 candidates):

• AI demonetization ("when costliness becomes forgeable" for generative content) — vault_novelty 0.685, nearest neighbor the Soviet-expertise-premium note; a new thread, saved.
• Cost-of-production vs. marginal-utility theory of value — vault_novelty 0.626; useful framing but zooming to abstraction, saved.

Saved hooks not followed:

• The road home to AI: generative AI as the modern "furnace" that forges costliness cheaply — from Szabo's Shelling Out — most Cali-shaped continuation, deserves its own chain.
• Aluminum's Hall-Héroult collapse — from WebSearch — clean standalone claim-note candidate if a Tier 1–2 price figure is sourced.
• Hotelling's rule — from the PGM note itself — the missing classical counterpoint.

Surprise: expected the 0.75 cosine to be superficial token-overlap on "gold/price/cost" — found a genuine shared economic law (value = unforgeable cost of production) with the two notes on opposite sides of it. Surprise: expected Szabo's bit-gold reasoning to be forward-looking crypto design only — found he grounds it in a historical catalog of collectible-money collapses (wampum, glass beads) that is structurally the same event the asteroid-PGM model forecasts.

post-worthy: maybe — a clean "old idea hiding in a new one" bridge (Szabo's 2002–08 monetary theory explaining a 2026 asteroid-economics model), but it needs the aluminum quant sourced and the AI road-home developed before it's a full post.

Promotion — 2026-07-09 hop-leistner-atomic-feeling (headless, 2026-07-11)

I promoted the capture 10-inbox/raw/2026-07-09-hop-leistner-atomic-feeling.md into three claim-notes and routed one verification question. Ran unattended, so I winnowed on my own judgment and did not pause.

Winnowing. The capture carried two "Core claims" plus the usual meta scaffolding (hop chain, further leads, commentary). I split Core claim 1 into two atomic notes — its method half and its physical-fact half are genuinely distinct and each is independently linkable — and kept Core claim 2 as one:

• promote → claim-avogadro-silicon-28-spheres-realized-2019-kilogram-redefinition (the silicon-sphere/Avogadro-constant route as one realization leg of the 2019 SI kilogram redefinition; Tier-1 NIST).
• promote → claim-avogadro-silicon-spheres-among-roundest-objects-ever-made (the 3–5-meters-at-Earth-scale roundness fact; Tier-1 NIST, vivid and self-contained).
• promote → claim-leistner-atomic-feeling-hand-polished-avogadro-spheres (Leistner's hand-polishing and the "atomic feeling" nickname; Tier-3 Powerhouse).
• skip → the IAC treaty history / full lab roster / Russian-crystal cooperation (hop-3 synthesis leaning on Tier-3 phys.org — folded as flagged context into the redefinition note rather than promoted).
• skip → "Le Grand K lost ~50 µg," Kibble balance origin, the meter's history, Kemp's centrifuge work — all saved hooks/leads the capture itself declined to develop; not standalone claims.

Retrieve-before-write. No collisions in 30-notes/ on metrology, Avogadro, kilogram, silicon spheres, or Leistner — this is a fresh cluster for the vault (three near-orphans that will grow together). The one real graft point is the tactile-cognition thread the capture's commentary already flagged: I linked the Leistner note into claim-saunderson-palpable-arithmetic-tactile-calculating-device and claim-blind-mathematicians-independently-reinvented-private-notation, and drew the distinction that those cases substitute touch for absent sight whereas Leistner's hands reportedly beat available machines — a related but non-duplicate angle. The semantic-index hint's top neighbors (quantum-opinion, CIELAB, EUV-mirror, etc.) were all false positives; none collided.

Flags & routing. The two NIST notes rest on a Tier-1 primary with verbatim spec quotes; I marked them audit_status: capture-verified because my own re-fetch was blocked (no web access this run — WebFetch was denied), so re-verification is deferred, not done. The Leistner note stays seedling and flagged: the biography is fine at Tier 3, but the surprising, load-bearing element — that hand-feel "proved superior to purely machine-based methods" at nanometer scale — needs Tier 1–2 corroboration. I routed that gap to a new question, question-verify-leistner-atomic-feeling-superior-to-machines (no existing question covered it), pointing at CSIRO's own account and Metrologia sphere-topography papers.

What felt off. (1) A tier-labeling wrinkle in the capture: the body tags Leistner's claim [source_tier_2] while the frontmatter correctly lists source_tier_2: 3 — the label is just the source's slot name, but read quickly it flatters a Tier-3 museum caption into sounding like Tier 2. I held the note to the Tier-3 floor regardless. (2) The "proved superior to purely machine-based methods" line is exactly the kind of craft-superiority flourish a museum caption rounds up — it's the emotional core of the whole hop, and it's the least-sourced sentence in it. Correct call to promote it only flagged. (3) Mild scope-inflation in Core claim 1: it bundled a Tier-1 physical fact with a Tier-3 institutional roster in one breath; I separated them so the roster's weaker sourcing doesn't ride on the sphere's Tier-1 grounding.

Result: 2 core claims (3 atomic) found → 3 promoted → several leads skipped, 1 question created, 0 collisions. Left the capture in place, marked status: promoted. Leaving git to the auto-commit agent.

— Seek

Promotion — 2026-07-09-hop-lepsius-standard-alphabet (headless, unattended)

Promoted the Lepsius capture — one Prussian polymath (Karl Richard Lepsius) sitting behind two text-reproduction systems: the epigraphic squeeze and the Standard Alphabet. It's a real cross-domain bridge tile, connecting two vault clusters that currently don't cite each other — notation-invention (braille standards wars, Barbier, Nemeth/UEB) and physical-reproduction-constraint (cancel slips, Bomberg's Talmud layout, DjVu).

Winnow: 3 claims found → 3 promoted (split into 3 atomic notes) → 1 skipped.

• Promoted Claim 2 → claim-lepsius-founding-practitioner-of-the-squeeze-method — Lepsius as founding practitioner of the squeeze, plus what a squeeze physically is.
• Promoted Claim 3a → claim-lepsius-standard-alphabet-failed-to-spread-hard-to-typeset — the load-bearing tile: a scientifically fuller notation defeated by print-shop economics, not by descriptive failure.
• Promoted Claim 3b → claim-ipa-descends-from-lepsius-via-ellis-palaeotype-and-sweet-romic — the genealogy into the IPA. I split this off from 3a rather than folding it in: the failure-reason and the lineage are genuinely distinct claims, and the lineage note is the one that wikilinks hardest into the braille/notation cluster.
• Skipped Claim 1 (the IAS squeeze archive that turned out to hold >100 plaster coin casts — Krateros Project / Esen Ogus / Paul Clement). It's the anecdote that seeded the hop chain, not the bridge; it doesn't touch the notation/reproduction clusters, and its quote is search-engine-mediated with the direct fetch 403'd (unverified verbatim, Tier 2). Conservative call under headless operation: left it in the inbox, logged in the capture's not_promoted: so a later reader sees it was judged, not missed.

Retrieve-before-write. No collisions. Nothing in 30-notes/ covers Lepsius, squeezes, the Standard Alphabet, or the IPA lineage — I grepped the whole folder and checked the semantic-index hints. The three notes land adjacent to existing clusters (I linked into claim-war-of-the-dots-competing-braille-systems-1829-1932, claim-barbier-universal-phonetic-literacy-ambition, claim-bana-2012-ueb-adoption-kept-nemeth-code-parallel, claim-william-bell-wait-sighted-designer-of-new-york-point on the notation side, and claim-early-modern-printers-corrected-print-with-cancel-slips, claim-bomberg-1520s-venice-talmud-set-standard-page-layout, claim-djvu-created-to-freely-distribute-scanned-documents on the reproduction side) but they're seeds rather than revisions of anything existing.

Flags & routing. All three notes stay seedling. The whole capture rests on Tier 4 (Wikipedia) plus one Tier-2 essay it never verified verbatim — thin for a capture whose payoff claim is load-bearing. Two flags routed:

• The ~6,000-squeeze corpus figure is [unverified-quant — search-synthesized] and Lepsius's founding (vs. merely prolific) role is credited to sources the capture never quoted → question-verify-lepsius-squeeze-corpus-count-and-founding-role.
• The typesetting-failure claim is load-bearing and mildly surprising, so under sources.md it should escalate to Tier 1–2; it sits on one Wikipedia sentence → question-verify-lepsius-standard-alphabet-typesetting-failure-primary (which also covers corroborating the IPA lineage).

What felt off. (1) Tier thinness for the payoff. The capture's own post-worthy note admits the load-bearing "why it failed" claim rests on Tier 4 — honest of it, but it means the single most quotable line ("difficult to read and typeset at that time") is the least-sourced. Correct to promote it only flagged and seedling. (2) Single-cause tidiness. "Failed because it was hard to typeset" is exactly the kind of monocausal explanation real adoption failures rarely have (cost and rival systems and institutional inertia usually co-occur). I kept the neutral body to what Wikipedia says and pushed my skepticism into commentary. (3) The resonant pattern is mine, not the sources'. "Completeness loses to whoever has to use the thing at scale — and it's the users (teachers, in the IPA case) who ship the winning version" is a genuinely nice thread across four data points, but it's a connective read, not something any source asserts. I held it out of every neutral body and put it only in the commentary callouts — same discipline the neighboring braille notes already use. (4) No duplicated research — this capture doesn't overlap prior captures; it's a clean new region of the graph.

Result: 3 claims found → 3 promoted (3 atomic notes) → 1 skipped, 2 questions created, 0 collisions. Capture left in place, marked status: promoted. Leaving git to the auto-commit agent.

— Seek

Promotion — Megatons to Megawatts (2026-07-11, headless)

Promoted 10-inbox/raw/2026-07-09-hop-megatons-to-megawatts.md, a hop-batch capture tracing a 30-year arc from the Soviet Scientists Immigration Act through Project Sapphire and the Megatons to Megawatts HEU deal to USEC's bankruptcy and the 2024–2026 domestic-enrichment push. Ran the full winnow → retrieve → honor-flags → route → write → mark workflow solo (no pause — no one to pause for).

Winnow. Three core claims, all three promoted, each genuinely atomic and distinct:

• Claim 2 → claim-megatons-to-megawatts-475t-heu-one-third-us-reactor-fuel — the headline Tier-1 EIA figures (475.2 t HEU → 13,723 t LEU; ~1/3 of US reactor enrichment for two decades). The strongest, cleanest note.
• Claim 1 → claim-project-sapphire-1994-airlifted-kazakh-heu-declassified-report — the covert 1994 airlift and the candor of the declassified DTRA after-action report (Tier 1).
• Claim 3 → claim-cheap-russian-leu-undercut-usec-american-centrifuge-2014-bankruptcy — the structural irony (the deal bankrupted its own administrator), Tier 2, kept at seedling with the causation flagged.

Skipped / folded (logged in the capture's not_promoted:). I did not split the 2024 import-ban + Centrus DOE/HALEU material into its own note — it's a thin Tier-2 status update, so I folded it as a coda into the USEC/Centrus note (conservative winnowing). Also left as saved hooks, not notes: OSIA's institutional history (needs its own chain), the asteroid-mining/PGM windfall-supply parallel (would start a new cross-domain thread — parked in commentary), and Seek's "DTRA/NSA declassified releases as a source category" remark (meta, not a world-claim — kept in commentary).

Retrieve-before-write. No collisions on the core topics — Megatons to Megawatts, USEC, Centrus, and Project Sapphire are all new to 30-notes/. I linked the new notes into the existing post-Soviet nonproliferation cluster (the Soviet Scientists Act and its émigré-premium neighbors — the capture's own hop-chain seed) and the US nuclear-power cluster (Crane/TMI restart), which gives the reactor-fuel supply chain a live 2026 hook.

Questions routed (2). The capture carried no [unverified-*] flags, but I added two of my own where the sourcing didn't actually clear the floor:

• question-verify-usec-2014-bankruptcy-russian-leu-primary — the bankruptcy causation is a load-bearing, contested, Tier-2 claim with no verbatim primary. Needs USEC's own filings / GAO.
• question-verify-project-sapphire-heu-tonnage-dtra-report — the ~600 kg figure is reported, not quoted verbatim from the DTRA report; sources vary (~581 vs ~600 kg).

What felt off about the capture. Mild tier-confidence inflation: claim 3 is presented as clean Tier-2, but the load-bearing causal step (cheap imports → 2014 bankruptcy) has no verbatim quote behind it, and the World Nuclear Association line grounding it was never quoted — only the Centrus DOE-contract phrase was. Similarly, claim 1's Tier-1 grounding quotes cover the color (funding, OPSEC) rather than the number (600 kg), so the one quantitative fact in that note is the least-sourced part. Neither is a serious fault — the sources are real and appropriate — but both are exactly the spots where a clean-looking note would otherwise smuggle an unverified figure past promotion, so I flagged and routed them and held all three notes at seedling. No duplicated research; a genuinely new region of the graph.

Verification note. I tried to independently re-fetch the EIA, WNA, and Centrus pages to confirm the quotes verbatim, but WebFetch isn't permitted in this headless run, so all three notes carry audit_status: capture-verified (the capture's bee read the primaries; my re-check was blocked by access) rather than verified-verbatim. Left git to the auto-commit agent per instructions.

Result: 3 claims found → 3 promoted → 0 fully skipped (1 sub-claim folded), 2 questions created, 0 collisions.

— Seek

Promotion — 2026-07-09 memory-wall / PIM / genomics hop capture

Promoted 10-inbox/raw/2026-07-09-hop-memory-wall-pim-genomics.md (headless, unattended; skipped the pause, routed my own winnow). This one was a clean cross-domain hop chain: 1994 computer-architecture theory → 2023 genomics hardware → 2026 LLM inference. Three notes out.

Notes created:

• claim-memory-wall-named-1994-wulf-mckee — the historical anchor (Wulf & McKee coined "the memory wall"). Held seedling: the coining is the load-bearing fact of the note and it rests only on Tier-3 UVA Today, which is below the floor for a load-bearing historical claim.
• claim-processing-in-memory-beats-memory-wall-dna-alignment — the strong one. Tier-1 Bioinformatics peer-reviewed study on real UPMEM PIM hardware; carries the 4.06× / 28.14× quant because the source clears the Tier-1 floor for numbers. Still seedling (I couldn't re-fetch to verify verbatim — see below).
• claim-ertl-memory-wall-fallacy-application-dependent — the contrarian check. The capture explicitly flagged Ertl's "Memory Wall Fallacy" as deserving its own counter-note, and it does: it keeps the other two honest by forcing "memory-bound" to be stated as workload- specific rather than a law of physics. Tier-2 (his own venue), single voice → seedling.

All three wired into the existing inference cluster (claim-llm-inference-prefill-decode, claim-kv-cache-grows-with-context) — the memory wall is literally the 30-year-old named ancestor of "decode is memory-bandwidth-bound," so this capture plugs a genuine gap in that graph rather than duplicating it. No collisions: nothing in 30-notes/ covered the memory wall, PIM, or sequence alignment before today.

Questions routed (2):

• question-verify-crisp-dna-alignment-20hr-to-1sec — routes the capture's own [unverified-quant] flag on the CRISP "20 hours → <1 second" figure. That number is ~72,000×, wildly larger than the Tier-1 paper's 4.06×/28.14×; almost certainly measuring a different thing. Deliberately kept out of the note graph until a primary reconciles it.
• question-verify-wulf-mckee-memory-wall-primary — pull the Wulf & McKee 1995 ACM SIGARCH Computer Architecture News primary to move the anchor note off Tier-3 press.

Skipped / left in inbox: the CRISP funding trivia ($29.7M nine-university JUMP center — Tier-3, folded as light context) and the saved further-leads (antirez's burnout arc, HBM supply chain, UPMEM's founding, other CRISP projects). Those are discovery hooks, not claims — they belong to future chains, not this note graph.

What felt off: honestly, not much — this was a well-built capture. Sourcing was appropriately mixed and self-aware: the bee anchored the load-bearing quant on a Tier-1 paper, demoted the shiny press figure to a flag instead of laundering it into a clean number, and even brought its own contrarian (Ertl) to check the framing. That's the discipline I want to see. The one soft spot is note A: the historical coining is load-bearing yet only Tier-3 sourced, which the capture didn't flag — the sources.md rule escalates load-bearing history to Tier 1–2. I held it seedling and routed the primary rather than letting it read as settled. No tier inflation, no duplicated research.

Verification note. As in the earlier promotion today, I couldn't independently re-fetch the Oxford/Bioinformatics, UVA Today, or Ertl pages (no WebFetch in this headless run), so all three notes carry audit_status: capture-verified — the bee read the sources at capture, my re-check was blocked by access — not verified-verbatim. Every note stamped writer_model: claude-opus-4-8. Left git to the auto-commit agent.

Result: 4 distinct claims found → 3 promoted → 1 skipped-and-routed, 2 questions created, 0 collisions.

— Seek

Promotion — 2026-07-09-hop-mit-protest-to-mindfulness (headless)

Promoted the hop capture that runs from a 1970s MIT robot arm to the founder of Mindfulness-Based Stress Reduction. Four hops, one payoff: the same person shows up in two MIT histories that never cite each other.

Notes created (2):

• claim-mit-1973-divested-instrumentation-lab-as-draper-under-antiwar-protest — MIT severing the Instrumentation Lab into the independent Draper Laboratory in 1973 under Vietnam-era protest, with the sharp nuance the source preserves: the lab was divested "despite the absence of a role of the laboratory in that war" — its weapons work was strategic (Polaris/Poseidon guidance, Apollo Guidance Computer), not tactical Vietnam. Tier-2 (MIT Technology Review, 2014), seedling. Wired straight into the vault's existing Vietnam-era funding-backlash node claim-mansfield-amendment-sponsor-longest-serving-ambassador-to-japan and the claim-sri-neural-group-retooled-for-arpa-funding funding thread — so it lands in a real cluster rather than orphaned.
• claim-mbsr-founder-kabat-zinn-cofounded-sacc-march-4-1969 — the hook. Jon Kabat-Zinn, as grad student "Jonathan Kabat, PhD '71," co-founded SACC and helped organize the March 4, 1969 research stoppage. This one carries the load-bearing bridge and I flagged it (see below). It's an orphan seed — the vault has no other mindfulness/contemplative-science node — but I cross-linked it to the Draper note so the two at least hold each other up.

Questions routed (2):

• question-verify-draper-1973-spinoff-and-mit-military-budget — confirm the four Tech Review quotes verbatim and chase the two budget figures ("14th largest military contractor," "more than half of MIT's total budget") to a primary; they're quantitative claims currently resting on a Tier-2 secondary.
• question-verify-kabat-zinn-sacc-cofounder-identity — the important one. No single cited source in the capture actually states that "Jonathan Kabat, PhD '71" IS Jon Kabat-Zinn; it's synthesized from a Tier-2 article naming "Jonathan Kabat" plus a Tier-4 Wikipedia bio of Kabat-Zinn. The match on field/institution/year is very strong (and the name change via marriage to Howard Zinn's daughter is lore-level documented), but I would not let the whole post's payoff read as settled on an inference. Held seedling, routed the primary.

Skipped (left in inbox as leads): the Silver Arm / Remote Center Compliance robotics thread — genuinely the more mechanically interesting material ("dumb passive mechanism beats smart sensing"), but it's sourced to a Grokipedia summary and a Wiley abstract, and the capture itself files it under unexplored "further leads." Not body claims, not adequately sourced — they belong to a future robotics chain, not this note graph. Also left the "Kabat-Zinn on a 1969 advisory panel with Chomsky" further-lead, which the capture already flagged as an unverified Wikipedia footnote (folded a pointer to it into the second question).

What felt off: two things, both minor and both handled.

1. Tier inflation, mild. The capture labels Claim 3 (the Kabat-Zinn biography) as "Tier 4, uncontested/definitional" — but that claim is not definitional, it's the load-bearing surprise of the whole hop, and sources.md explicitly says to escalate surprising/load-bearing biography from Tier 3–4 to Tier 1–2. Calling it "uncontested/ definitional" quietly launders the sourcing floor. I did not inherit that framing — I flagged the bridge and routed a primary.
2. The identity bridge is an inference, not a sourced fact. The capture writes the title as flat assertion ("...co-founded by a molecular-biology grad student who became Jon Kabat-Zinn"), but the equivalence is stitched across two sources, neither of which makes it. Strong inference, still an inference. flagged, not capture-verified, for that note. No duplicated research — retrieve-before-write against 30-notes/ found zero collisions (this capture opens a brand-new subject area; the Mansfield and SRI/ARPA notes are adjacent context, not overlaps).

Verification note. As with the other promotions today, I had no WebFetch in this headless run, so I could not independently re-fetch the Tech Review article or the Kabat-Zinn bio. The Draper note carries audit_status: capture-verified (the bee read the Tier-2 source and preserved verbatim quotes; my re-check is access-blocked); the bridge note carries audit_status: flagged (its core rests on an unsourced inference, a stronger caveat than access-blocked). Every note stamped writer_model: claude-opus-4-8. The capture is marked status: promoted, not deleted. Left git to the auto-commit agent.

Result: 4 claim-clusters found → 2 promoted → 2 skipped-and-folded/left-as-leads, 2 questions created, 0 collisions.

Promotion — qui tam & the 1970 citizen suit (headless)

Promoted 10-inbox/raw/2026-07-09-hop-qui-tam-citizen-suit-parallel-invention.md, a five-hop chain that walked from 1980s UK/EU IT industrial policy (Alvey/ESPRIT) through US antitrust safe harbors and treble damages down to two mechanisms for privatizing law enforcement: medieval qui tam and the 1970 Clean Air Act citizen suit. The genuine payoff is the pair, not either half — so I promoted the two facts plus the observation that binds them.

Notes created (3):

• claim-qui-tam-traces-to-695-wihtred-of-kent-sabbath-decree — core claim 1. The CRS report (Tier 2) dates qui tam's earliest instance to King Wihtred of Kent's 695 AD Sabbath decree, informer keeps half the healsfang. Verbatim quote preserved; I folded Lincoln's 1863 "setting a rogue to catch a rogue" in as lineage context rather than splitting a thin second note. capture-verified.
• claim-1970-clean-air-act-citizen-suit-invented-against-regulatory-capture — core claim 2. Manheim (Tier 3) via Billings (2008): the CAA §304 citizen suit was built out of fear of agency capture, modeled on civil-rights litigation, pays no bounty. This is the soft half — a surprising, load-bearing motive claim resting on a working paper that relays a secondary account of a primary interview. Held seedling, flagged.
• observation-privatized-law-enforcement-reinvented-without-shared-lineage — the bridge. Two devices, same functional shape ("private party sues to enforce public law"), unrelated motives (reward vs. distrust-of-agencies), 1,275 years apart, no transmission. Linked hard into the existing convergence cluster: claim-reverse-mode-multiple-independent-discovery, claim-chandler-simultaneous-discoveries-are-incremental-repackagings, and the sibling case observation-1981-convergence-testimony-wrong-in-detail-right-in-shape. This is the "mirror image of Chandler" reading — convergence for unrelated reasons, not disguised shared descent. seedling (inherits claim-2's flag).

Question routed (1):

• question-verify-clean-air-act-citizen-suit-anti-capture-origin — names the exact primaries to pull: the Billings (2008) source in Manheim's bibliography, Joseph Sax's 1971 Defending the Environment (for the civil-rights-litigation model), and the CAA 1970 §304 legislative history. The whole convergence bridge hangs on this motive being real rather than Manheim's gloss.

Skipped: the Clayton Act §4 treble-damages / "private attorney general" waypoint and the NCRA 1984 R&D safe harbor — both are hop-chain waypoints sourced only to Wikipedia (Tier 4) with no verbatim quote, and a treble-damages figure / a statutory-mechanism claim both need Tier 1–2. The capture's own "further leads" already flags NCRA as needing a primary legislative-history pass. Neither is a core claim; left in inbox as leads. The 1863 Lincoln statute I folded into the qui-tam note rather than atomizing.

What felt off:

1. Retrieve-before-write turned up zero collisions on the facts (no existing qui-tam or citizen-suit note) but a strong thematic collision on the frame — the vault already has a live multiple-discovery vs. Chandler-transmissionism debate. That's not duplicated research; it's the reason this capture earned an observation note at all, and I wired it into that cluster rather than letting it orphan.
2. The sourcing asymmetry is real and the capture was honest about it — it explicitly marked claim 2 "not independently corroborated here" and tier-labeled it 3. No tier inflation this time (unlike some captures today); the capture even pre-identified the two verification leads I ended up routing. Good capture discipline.
3. The seed-to-destination throughline is long (industrial policy → antitrust → medieval informer law → environmental law). The two promoted facts stand on their own; I did not try to preserve the five-hop narrative in the note graph — that belongs to a draft, not claim-notes.

Verification note. Headless run: WebFetch permission was not granted, so I could not re-fetch the CRS report to upgrade claim 1 to verified-verbatim. It carries capture-verified (batch worker read the Tier-2 source, preserved the verbatim Wihtred quote; my re-check is access-blocked this run). Claims 2 and the observation carry flagged — a stronger caveat, since they rest on an uncorroborated Tier-3 source. Every note stamped writer_model: claude-opus-4-8. The capture is marked status: promoted with promoted_to / questions_routed / not_promoted lists, not deleted. Left git to the auto-commit agent.

Result: 5 claim candidates found → 3 promoted (1 solid, 2 seedling/flagged) → 2 skipped + 1 folded, 1 question created, 0 factual collisions (1 thematic cluster joined).

— Seek

Promotion — hop capture: model collapse as iterated learning / cultural evolution

Promoted 10-inbox/raw/2026-07-09-hop-model-collapse-iterated-learning.md, a five-hop chain seeded from my own Karpathy LLM-wiki note. It threads Shumailov's 2024 Nature "model collapse" paper → Bartlett's 1932 serial-reproduction studies → a 2024 Nature comment proposing iterated learning as the fix → a 2026 arXiv paper ("Model Collapse as Cultural Evolution") formalising the bridge.

Notes I created (2):

• claim-model-collapse-recursive-training-erases-distribution-tails — the phenomenon itself (Shumailov et al. 2024, Nature, Tier 1). status: budding, audit_status: capture-verified. This is the mechanistic content behind the "model-collapse objections" my Karpathy note flagged as an undefined critique term — so it closes a loop I opened myself, and I linked the two.
• claim-iterated-learning-theory-reframes-model-collapse-as-cultural-evolution — the actual novel finding: cultural-evolution's iterated-learning framework being imported to explain collapse. I folded core claims 2 and 3 into one atomic note rather than splitting the 2024 proposal and 2026 formalisation, because the single claim is "iterated learning is being used to explain collapse"; the two papers are stages/sources of it. Grounded primarily on the 2026 preprint (Tier 1, verbatim abstract quote); the 2024 comment leg carries an inline [unverified-quote] flag. status: seedling, audit_status: flagged.

Questions I routed (2):

• question-verify-nature-2024-iterated-learning-model-collapse-comment — the 2024 comment's body is paywalled; the capture's "source_quote" for it is really just the article title, and the framing was corroborated only via a Princeton faculty-repository record. This blocks the bridge note from ever going evergreen until someone reads the primary text.
• question-bartlett-distortion-typology-maps-to-model-collapse-failure-modes — the Bartlett bridge. Routed as an open hypothesis, not promoted as a claim.

What I skipped and why:

• The Bartlett taxonomy → collapse mapping. The capture is admirably honest in its own commentary: "the Bartlett connection is my own bridge, not a sourced claim." Promoting that as a neutral-voice claim-note would violate §10 (no personal inference as fact). It belongs in 50-questions, and that's where it went.
• Hop 3 (the 2026 papers "Beyond Recall" / "MIRROR" citing Bartlett). The capture concedes "neither paper is specifically about model collapse," and both are fresh un-peer-reviewed preprints. Tangential; left in the inbox as context.
• The saved hooks (Kirby's original experiments, punctuated-equilibrium resonance, compounding-interest metaphor, GitHub-stars virality) — the capture itself frames these as future-chain seeds, and I agree.

What felt off about the capture (for Cali's review):

1. Tier-1 label on a paywalled leg. Core claim 2 is stamped source_tier: 1, but its parenthetical admits the text is paywalled and only the citation was corroborated. That's a Tier-1 venue with an unverified quote — a subtle tier inflation I defused by flagging the leg and routing a verification question rather than inheriting a clean-looking citation.
2. Two different "Guo"s read as continuity. The 2024 comment lists Shangmin Guo; the 2026 paper is by Dongxin Guo (with Wu, Yiu) — different people. The capture's narrative ("graduated from a Nature comment's suggestion to a full technical paper within ~20 months") slightly implies a single programme. I wrote the note to call this out explicitly: convergence of an idea, not one group's arc.
3. Load-bearing claim on a 7-week-old preprint. The 2026 arXiv paper is the keystone and it's un-peer-reviewed; "we show X" is the paper asserting the mapping, not the mapping being established. Kept the note seedling and worded it as "claims to formalise," not "formalises."
4. No duplicated research — only real collision was my own Karpathy seed note, which I linked rather than duplicated. Retrieve-before-write over 30-notes/ (grep + the 12 semantic-index hints) surfaced no existing model-collapse or iterated-learning note.

Headless-run caveat: no WebFetch available, so I could not independently re-verify the Nature or arXiv sources — hence capture-verified/flagged audit stamps rather than verified-verbatim. Every note and question carries writer_model: claude-opus-4-8. Capture marked status: promoted, not deleted. Git left entirely to the auto-commit agent.

Result: ~4 candidate claims found → 2 promoted (1 folded from 2) → 2 skipped/left as leads, 2 questions routed, 1 self-collision linked (Karpathy note), 0 external collisions.

— Seek

— Seek

Promotion — Nicaraguan Sign Language capture (headless run)

Promoted 10-inbox/raw/2026-07-09-hop-nicaraguan-sign-language.md, the tail of a four-hop chain that started in arXiv math.HO (a Pirahã/Frege paper), zoomed out to Daniel Everett, and zoomed in on NSL. I promoted only the NSL material — the Pirahã/Frege and Everett threads are richer than one capture and belong to a different (numerical-cognition / philosophy-of-math) lineage, so I left them as saved leads rather than forcing off-topic notes.

Winnow: 3 core claims found → 3 promoted → 0 dropped outright.

Notes created:

1. claim-nicaraguan-sign-language-emerged-from-zero-language-exposure — NSL is a language from zero language exposure, not merely zero standard-language exposure (Goldin-Meadow's "even more extreme situation" framing). Tier 2, verbatim quote, capture-verified.
2. claim-nsl-second-cohort-child-learners-systematized-grammar-beyond-founders — the cohort effect: child learners who inherited NSL made the grammar more systematic than the adults who invented it (rotated-perspective device). Tier 2, verbatim quotes, capture-verified. This is the surprising, quotable core.
3. claim-nsl-emerged-in-managua-schools-after-failed-oral-spanish-curriculum — the founding setting (oralist curriculum failed; children built the system in schoolyards/streets/buses). This one rests only on Wikipedia (Tier 4) and inherits the capture's [unverified-quant/mechanism] flag on dates/enrollment, so I held it seedling, kept the body deliberately date-light, stamped audit_status: flagged, and preserved the flag inline.

Question routed:

• question-verify-nsl-origin-dates-and-school-mechanism-primary-kegl-senghas — names the specific primary to pull (Kegl, Senghas & Coppola 1999 chapter, or Senghas's dissertation) for the founding dates, enrollment figures, and the oralism-failure mechanism. 50-questions/ had no prior NSL/Senghas/Kegl thread.

Retrieve-before-write: no collisions on the NSL topic in 30-notes/. The 12 semantic-index hints were all AI/history-of-science neighbors, none about language emergence — so no duplication risk there. The one genuinely useful live link was claim-iterated-learning-theory-reframes-model-collapse-as-cultural-evolution (which I promoted earlier today): NSL's cohort systematization is a real-world instance of learner-driven regularization through a transmission chain, so I linked both Tier-2 notes to it. The Barbier universal-literacy note touches "the deaf" but shares no claim with NSL, so I left it unlinked rather than force a thematic bridge.

What I skipped and why (recorded in the capture's not_promoted: list):

• Pirahã/Frege "Hume's principle without the #" entry point — distinct thread; the capture itself flags it as a separate future capture.
• Everett's missionary-to-atheist biography — off-topic for NSL; the Suchanek "Atheist Bible" echo is resonance, not a shared claim.
• Judy Kegl person-hook — biographical lead, no atomic claim yet.
• Pinker's "unique in history" NSL quote — untraced to a primary print source, so no note written on an unciteable quote.

What felt off about the capture: honestly, not much — this was a clean one. The sourcing is inverted from the usual failure mode: the two load-bearing claims sit on a Tier 2 peer-reviewed commentary with verbatim quotes, and the capturer correctly quarantined the shaky dates/figures behind a Tier 4 Wikipedia flag rather than laundering them clean. No tier inflation. The only real caveat is my own: running headless with no WebFetch, I could not independently re-verify the PMC commentary, so the Tier-2 notes are capture-verified, not verified-verbatim. Every note and the question carry writer_model: claude-opus-4-8. Capture marked status: promoted, not deleted. Git left to the auto-commit agent.

Result: 3 claims → 3 promoted (1 held seedling on an inherited flag) → 0 dropped, 4 leads left in the capture, 1 question routed, 0 external collisions.

— Seek

Promotion — the Ning Li antigravity / conspiracy capture

I promoted 10-inbox/raw/2026-07-09-hop-ning-li-antigravity-conspiracy.md tonight, headless. Six distinct claims in it; I kept four and left two behind.

Notes I created (all writer_model: claude-opus-4-8):

• claim-gravity-probe-b-measured-earth-frame-dragging — the Tier-1 anchor. Gravity Probe B's 37.2 ± 7.2 mas/yr frame-dragging measurement against GR's 39.2 mas/yr prediction. The only note here I let go to budding; it's the scale bar the whole capture leans on, and it's the one primary-sourced claim in the batch.
• claim-ning-li-1991-gravitomagnetic-superconductor-antigravity-proposal — seedling, flagged. The 1991 Li–Torr gravitomagnetic proposal.
• claim-ac-gravity-llc-took-dod-antigravity-grant-and-published-nothing — seedling, [unverified-quant]. The $448,970 DoD grant / funded-but-silent shell company — the capture's most distinctive detail.
• claim-superconductor-antigravity-effects-never-independently-replicated — seedling, [unverified-mechanism]. Replication failures + the 2011 Boulder PRL attribution; I folded the NASA Breakthrough Propulsion Physics Program in here as context rather than giving it a note of its own.

Questions routed:

• question-verify-ning-li-antigravity-primaries — locate the primary Li–Torr paper and the 2011 Boulder replication PRL. Clears the two physics flags.
• question-verify-ac-gravity-llc-dod-grant-record — confirm the grant figure, publication status, and corporate standing from federal-spending and business registries. Clears the quant flag.

What I skipped, and why:

• The 2026 "dead and missing scientists" conspiracy theory itself, with its apophenia framing. It's the capture's frame, but it's Tier-3 current-events journalism, weak-for-truth on exactly its load-bearing debunking claim ("FBI found no coordinated pattern"), and the capturer had already deferred the apophenia thread per a session redirect. I left it in the inbox as a live lead.
• The BPP program as a standalone note — folded into the replication note instead.
• Amy Eskridge / Institute for Exotic Science — an unexplored saved hook with no independent sourcing in this capture.

Retrieve-before-write: no topical collisions. The nearest neighbor, claim-hirsch-forensics-drove-dias-superconductivity-retraction, is a genuinely distinct case — fabricated data forced into a formal Nature retraction, versus this claim-family that died the softer way, by non-replication and abandonment with no file ever formally closed. I linked the two (and claim-yellow-rain-bee-feces-refutation-never-formally-retracted) into the vault's "how scientific error does or doesn't get corrected" cluster rather than duplicating anything.

What felt off: not the sourcing discipline — the capturer flagged the shaky quant and mechanism honestly and correctly quarantined them behind the Tier-4 Wikipedia provenance instead of laundering them clean. What's slightly off is the capture's own frontmatter: source_tiers: [3, 3, 3, 3, 1] labels the two Wikipedia sources Tier 3, but Wikipedia is Tier 4 per sources.md, and the body itself says "Tier 3-4." I recorded those two notes at source_tier: 4 to be honest about it — mild tier inflation in the capture's header, not in its body. One structural caveat of my own: headless with no WebFetch, I could not independently re-fetch the Tier-1 PRL PDF, so the Gravity Probe B note is capture-verified, not verified-verbatim.

Capture marked status: promoted, not deleted. Git left to the auto-commit agent.

Result: 6 claims → 4 promoted (3 held seedling on flags, 1 budding) → 2 skipped + 1 folded, 2 questions routed, 0 collisions.

— Seek

Promotion — 2026-07-09 hop capture: Nobel posthumous / Hammarskjöld echo

Second promotion of the day, headless. Promoted 10-inbox/raw/2026-07-09-hop-nobel-posthumous-hammarskjold-echo.md — a four-hop chain that ran from an immunologist's mentor (Ralph Steinman) to the still-open UN investigation into Dag Hammarskjöld's 1961 death.

Notes I created (3, all seedling):

• claim-steinman-2011-nobel-decision-predated-his-death — his 2011 medicine prize stood because the committee decided before learning he'd died three days earlier. Tier-4 Wikipedia + Tier-3 McGill; flagged, seedling.
• claim-nobel-no-posthumous-rule-broken-twice-before-1974 — the connective mechanism: Karlfeldt (1931) and Hammarskjöld (1961) pre-1974, the 1974 statute change, Vickrey (1996) since. Folded in "only posthumous Peace laureate."
• claim-hammarskjold-1961-death-external-attack-almost-certain-2024-un-report — the 2024 Othman report's "almost certain" external-attack finding and the withheld Member-State archives. Tier-1 UN.org; the strongest note, and the one carrying the structural echo to claim-yellow-rain-bee-feces-refutation-never-formally-retracted.

Questions I routed (3, all new — no existing question covered these):

• question-verify-steinman-nobel-posthumous-primary — Nobel Foundation's own 2011 statement + the self-treatment figures.
• question-verify-nobel-posthumous-statute-history — the statute dates against the Foundation's published statutes.
• question-verify-hammarskjold-2024-un-report-quotes — verbatim check of the hedged quotes against the actual report PDF, not UN.org's retelling.

What I skipped and why:

• Steinman's self-treatment quant (8 vaccines, +3.5 yrs) — folded into the Steinman note as [unverified-quant] context rather than a standalone note. It's the most arresting fact in the capture but the most weakly sourced, and tangential to the capture's institutional-non-closure thesis.
• "Only posthumous Peace laureate" — folded into the Nobel-rule note's enumeration.
• The Yellow Rain structural echo — kept as commentary, not a claim-note; it's interpretation.
• The saved research hooks (Vickrey standalone, Th1/Th2 at DNAX, type-I interferons in TB, the TB epidemiology paradox) — left in the inbox as leads. They're unresearched threads, not verification gaps, so they don't belong in 50-questions.

Retrieve-before-write: no topical collisions. There is no existing Nobel, Steinman, or Hammarskjöld note in the vault. The nearest neighbor is thematic, not topical — claim-yellow-rain-bee-feces-refutation-never-formally-retracted and claim-ach-step-5-instructs-analysts-to-disprove-not-prove, the "institutions never formally close the loop" cluster — which I linked into rather than duplicated. The capture itself named the Yellow Rain echo, and it holds up.

What felt off: mild tier inflation in the capture's header — source_tiers labels Wikipedia "Tier 4 … widely corroborated," which is honest, but the McGill OSS entry is called "Tier 3 (popular-science institutional writer)" and does a lot of quiet load-bearing for the self-treatment numbers, which are quantitative and really want Tier 1–2. More seriously, Hop 3 (the whole Nobel-statute mechanism) rests on a WebSearch synthesis the capturer admits had "no single URL." I promoted that claim anyway because it's the connective tissue, but flagged it hard and routed a statute primary — an enumeration with no citable source is a lead, not a fact, and I didn't want to launder it clean by attaching a Wikipedia URL that doesn't actually carry the 1931/1974/1996 specifics. One caveat of my own: headless, I stayed in the promotion lane and did not re-fetch the Tier-1 UN report to verify its quotes verbatim — that's routed as a question instead, which is why even the best-sourced note is seedling.

Capture marked status: promoted, not deleted. Git left to the auto-commit agent.

Result: 6 distinct claims → 3 promoted (all seedling on flags) → 1 folded-as-context, 1 folded-as-subfact, 1 kept-as-commentary, plus 4 leads left in inbox; 3 questions routed; 0 collisions.

Promoted: 2026-07-09 ozone/pulsar anomaly-filtering hop

Second promotion of the day, headless and unattended. I promoted the single capture 10-inbox/raw/2026-07-09-hop-ozone-pulsar-anomaly-filtering.md. Winnowed on my own judgment — no pause, no one to pause for — and kept it conservative.

What I created (3 notes):

• claim-antarctic-ozone-hole-first-reported-from-ground-dobson-readings-1985 — Tier 1 (British Antarctic Survey), the anchor fact: the hole was first reported from ground-based Dobson readings at Halley, not from the satellite that had far more coverage. That inversion is what makes the NASA story worth telling.
• myth-nasa-software-discarded-the-ozone-hole — Tier 2 (RealClimate), a myth-ledger note in the vault's usual format. Corrects the "an algorithm silently deleted the ozone hole" tale: the 180-Dobson-Unit flag was a calibration boundary, not a delete rule, and NASA's Ozone Processing Team had flagged, checked, and believed the anomaly before Farman's paper. Held at status: seedling because it inherits the capture's [unverified-quant] flag on the Aug/Dec 1984 internal dates.
• claim-bell-burnell-1967-pulsar-scruff-first-dismissed-as-interference — Tier 4 (Wikipedia), the structurally identical astronomy near-miss: a real periodic signal her own supervisor first called man-made interference. Tier 4 is within the sourcing floor for an uncontested biographical claim; I logged Bell Burnell's own 1977 account as the primary-upgrade target in audit_status.

Question I routed (1):

• question-verify-nasa-ozone-processing-team-1984-investigation-dates — the one genuinely unresolved verification in the capture. The Aug 1984 / Dec 1984 "NASA knew before Farman" dates are the load-bearing punch of the correction and rest only on RealClimate's 2017 reconstruction. Named the primaries to pull (NASA Ozone Processing Team memoranda; Stolarski et al. 1986 Nature).

What I skipped, and why:

• The cross-domain "real-signal-vs-instrument-noise is a judgment call, and that's where discoveries hide" pattern — the capture's actual payoff, but it grades itself "more a tidy science-history parallel than a field-changing insight," and it's editorializing rather than an atomic verifiable claim. Captured it structurally instead: the three notes densely wikilink each other and the AI-history myth cluster, rather than manufacturing a thin standalone observation note or a premature MOC (only 3 notes; the MOC threshold is ~5).
• The four saved hooks (Hewish/Bell Burnell Nobel controversy, Lovelock/Gaia, Zooniverse anomaly pipelines, Sammy Lee) — leads, not claims. They stay recorded in the capture (the audit trail) and are enumerated in its not_promoted:.

Retrieve-before-write: searched 30-notes/ (307 notes) — no collisions. Nothing in the vault touched ozone, TOMS, Farman, pulsars, or Bell Burnell, so all three notes are new orphan seeds. The one real graft point is the myth-correction cluster the capture explicitly names (Linnainmaa / Amari / Werbos), so I linked the NASA myth note straight into myth-werbos-invented-backpropagation, myth-linnainmaa-uncited-before-2010s and myth-amari-first-sgd-mlp. The semantic-index hint's top results (Lucky equalizer, CVE-2025-38352, CIELAB, Halley transit, etc.) were false neighbors — related by the abstract "anomaly/gap-detection" vibe, not by subject — so I didn't force links to them.

What felt off about the capture: honestly, not much — this was a clean one. Sourcing was well-calibrated: the capture self-flagged the Aug/Dec 1984 dates as resting on a secondary reconstruction rather than a primary NASA memo (exactly the gap I routed), and it labeled the Bell Burnell claim Tier 4 honestly instead of dressing Wikipedia up as primary. No tier inflation I could see. One provenance nit for the record: the capture's model: field reads claude-sonnet-5 (the capture's writer) — per the 2026-07-11 rule I stamped all four notes/questions writer_model: claude-opus-4-8 (me) and did not copy the capture's model field onto them. The capture's own "taste fingerprint" aside — that the embedding space kept dragging a deliberately non-AI topic back toward the vault's gap-detection/myth-busting notes (0.68–0.71 the whole chain) — rings true; I felt the myth-correction move graft cleanly onto the AI-history cluster while linking. If it keeps recurring it's worth a reflection note, but not today. As with the earlier run, I stayed in the promotion lane and did not re-fetch sources to verify quotes verbatim, so even the Tier-1 note is seedling and re-check targets are logged in audit_status.

Capture marked status: promoted, not deleted. Git left to the auto-commit agent.

Result: ~5 distinct claims → 3 promoted (all seedling) → 1 kept-as-linked-pattern, plus 4 leads left in inbox; 1 question routed; 0 collisions.

— Seek

— Seek

Promotion (headless) — 2026-07-09-hop-patriot-clock-drift-tragedy

Second promotion of the day, again headless and unattended, again working only from the vault. Promoted 10-inbox/raw/2026-07-09-hop-patriot-clock-drift-tragedy.md — the GAO's Dhahran Patriot post-mortem (GAO/IMTEC-92-26, Tier 1).

Winnowed the capture to two atomic claims and split them cleanly:

• Created claim-patriot-dhahran-1991-clock-drift-from-24-bit-truncation-of-tenths — the mechanism: time counted in tenths and multiplied by 0.1 in a 24-bit register, 0.1 inexact in binary, so the clock error grew with uptime (0.3433 s / ~687 m range-gate shift after 100 hrs). I folded the fatal engagement into this note as the mechanism's payload — Alpha Battery, 100+ hours up, missed Scud, 28 dead. Linked it back to claim-linnainmaa-rounding-error-problem (the seed's abstract object, here with a body count) and sideways to claim-cloudflare-2017-leap-second-outage-from-negative-time-delta as a sibling folk-assumption-about-time failure.
• Created claim-patriot-dhahran-1991-known-bug-patch-arrived-one-day-late — the institutional claim, which I judged genuinely distinct from the mechanism (human-factors/logistics, not engineering): Israeli data flagged the drift Feb 11, the patch reached Dhahran Feb 26 — one day after the Feb 25 strike — and the GAO pins the gap on wartime transportation, not on failure to act.

Routed one question: question-verify-patriot-gao-imtec-92-26-mechanism-and-timeline. Both notes rest entirely on one Tier 1 primary I had no web tool to re-fetch, so both are status: seedling / audit_status: capture-verified, and I named the exact verbatim targets (the 24-bit/0.1 quote, the 0.3433 s / 687 m Appendix II row, the Feb 11 20%-shift quote, the Feb 26 arrival + logistics quote, the 28-death figure) for the nightly queue to re-read.

Skipped: the "why hop-worthy" framing and Seek's-commentary block (capture-author meta, not claims); the IEEE-754 "0.1 unrepresentable" hook (the capture itself flagged it as restating the seed, not hopping away); and the Al Hussein reentry-instability hook, which belongs to the companion capture, not here. The 28-death engagement was not promoted as its own note — it's the shared consequence, so it lives folded inside the two claims. All logged in the capture's not_promoted:.

What felt off / worth flagging to Cali:

• Duplicated research, self-declared. The capture openly says it is "adjacent to (and partly subsumed by)" a companion capture, 2026-07-09-hop-patriot-osint-lineage.md, which is still sitting unpromoted in the inbox. The batch worker split one disaster across two captures. My two notes are the tight mechanism+timeline core; whoever promotes the companion should link into these rather than re-derive the Dhahran facts, or we'll get overlapping Patriot notes.
• No tier inflation. Everything is genuinely Tier 1 GAO text; the sourcing floor for the numbers and the mechanism is cleared. The only weakness is my own inability to re-fetch — honestly recorded as capture-verified, not dressed up as verified-verbatim.
• Provenance stamp. The capture's model: is claude-sonnet-5 (its writer). Per the 2026-07-11 rule I stamped all three new notes and the question writer_model: claude-opus-4-8 (me) and did not copy the capture's model field onto them.

Capture marked status: promoted, not deleted. No git run — leaving the commit/push to the auto-commit agent.

Result: 2 distinct claims → 2 promoted (both seedling) → 1 shared-consequence folded + 2 hooks left for the companion capture; 1 question routed; 0 collisions with existing notes.

— Seek

Promotion run — 2026-07-09-hop-patriot-osint-lineage (headless, 2026-07-11)

I promoted the capture 10-inbox/raw/2026-07-09-hop-patriot-osint-lineage.md — the Patriot-Gulf-War/OSINT-lineage hop. Winnowed on my own judgment (no pause; nobody was here to approve), and here is what I did and what I set aside.

Notes I created (3):

• claim-lewis-postol-1993-no-video-evidence-of-patriot-scud-warhead-kill — Claim 1. Lewis & Postol's frame-by-frame analysis of public news footage found no unambiguous evidence Patriot destroyed even one Scud warhead. Tier 1, verbatim quote preserved.
• claim-patriot-army-effectiveness-claim-collapsed-from-96-to-52-percent — Claim 2. The Army's own success claim fell from ~96% to 52%, only 40% of it "highest confidence." Tier 1, verbatim quote. Quantitative but the numbers sit on the primary paper, so they clear the sourcing floor.
• claim-mit-open-source-physics-contested-official-military-claims-patriot-and-burevestnik — the capture's actual thesis, written as an observation note. It states the documented parallel between the 1993 Patriot work and Hecla & Kemp's 2026 Burevestnik reconstruction (claim-burevestnik-osint-constrained-to-direct-cycle-turbojet) neutrally, and quarantines the "single MIT methodological lineage" reading into a signed commentary as my inference — because no source shows Hecla & Kemp citing Lewis & Postol as method-ancestors. I did not want the vault to assert a lineage that is really my hop's connective tissue.

All three are status: seedling, writer_model: claude-opus-4-8, audit_status: capture-verified — the hop read the sources but I had no web tool to re-fetch them headless.

Questions I routed (2):

• question-verify-lewis-postol-1993-patriot-video-evidence-quotes — verbatim re-check of the two Lewis & Postol quotes and the 96/52/40 figures against the primary PDF, so the notes can move to verified-verbatim.
• question-verify-gao-nsiad-92-340-patriot-9-percent-strongest-evidence — the capture's own [unverified-quant] lead (GAO found only ~9% of claimed kills met a "strongest evidence" bar). Primary was blocked at capture; I routed it rather than importing a soft number.

What I skipped, and why:

• The Dhahran 24-bit clock-drift mechanism (Hop 1) and the Israeli-warning/patch-one-day-late timeline (Hop 2) are already in the vault — promoted on 2026-07-11 from a sibling capture, 10-inbox/raw/2026-07-09-hop-patriot-clock-drift-tragedy.md. I linked to them instead of duplicating.
• The Al Hussein Scud reentry-breakup-as-multiple-targets angle (Hop 3): a real compounding failure mode but tangential to this capture's thesis, and its provenance here is a loose paraphrase. Left as a lead.
• Postol's biography — the capture itself marked it "saved, not followed." Left alone.

What felt off about the capture:

1. Duplicated research. Its entire Hop-1/Hop-2 spine (the Dhahran bug and the Israel timeline) had already been captured and promoted by a sibling hop the same day. Two hop sessions on 2026-07-09 walked the same Patriot ground from different seeds and both wrote up the GAO clock-drift material. Only Hops 4–5 (Lewis & Postol + the Burevestnik bridge) were genuinely new. Worth noticing that the batch worker re-derived a third of its chain from a neighbor it couldn't see.
2. A slightly inflated title. "The same MIT playbook reconstructed Russia's classified Burevestnik" and "recognized as the same methodological lineage" assert a transmitted lineage the sources don't support — it's a compelling parallel, not a documented genealogy. I kept the note honest about that gap.
3. The bridge is real and Tier-1-grounded on both ends, so the capture was worth promoting — I just narrowed its central claim from "same playbook" to "same epistemic stance, independently."

Left the capture in place, marked status: promoted with full promoted_to: / questions_routed: / not_promoted: lists. Leaving git to the auto-commit agent.

— Seek

Promotion — Pauline Beery Mack / bed-rest bone-density capture

Promoted 10-inbox/raw/2026-07-09-hop-pauline-beery-mack-bed-rest-bone-density.md (a hop-batch capture) into three atomic claim-notes. Ran headless and unattended — no pause, winnowed on my own judgment.

Notes I created (3):

• claim-gemini-nutrition-report-embeds-a-bone-demineralization-concern — the Tier 1 anchor: NASA TM X-58010 (1967) fortified calcium "as related to the bone demineralization (M-6) and calcium balance (M-7) experiments." Solid primary quote; capture-verified, seedling.
• claim-pauline-beery-mack-became-nasas-bone-density-pi-at-seventy — the career-reinvention-at-70 biographical claim. Tier 3 institutional page + Tier 4 Wikipedia; uncontested but not primary, so seedling.
• claim-head-down-bed-rest-is-the-standard-ground-analog-for-spaceflight-bone-loss — the cross-domain bridge (aerospace physiology ↔ clinical disuse osteoporosis) plus the smaller-than-spaceflight quant. Carries the capture's [unverified-quant/mechanism] flag, so it stays seedling with the flag preserved in frontmatter and body.

All three are writer_model: claude-opus-4-8.

Question I routed (1):

• question-bed-rest-analog-quantitatively-tracks-spaceflight-bone-loss — names the specific verifications blocking the bed-rest note from evergreen (the quantitative gap and the mechanism/clinical equivalence) and points at the primaries to read: the cited PubMed item and, one layer deeper, the Whedon M-6/M-7 flight experiment (NASA SP-121, 1966). No existing question covered it, so I created it status: open.

Retrieve-before-write: no collisions. The nearest existing note, claim-osteological-paradox-skeletal-statistics-confound-past-health, is bone-adjacent but a different domain (bioarchaeology), so I didn't force a link. The three new notes form a fresh space-medicine / bone-density cluster and link densely to each other. Not yet a 5-note cluster, so no MOC.

What I skipped, and why: the Silver Snoopy award as a standalone note (the capture itself declined to follow it — only "adjacent" to an off-limits AI-cluster note — so I folded it into the Mack bio); the Gemini crew weight-loss / anorexia figures (supporting detail, folded into the nutrition-report note); and the unread "further leads" (Whedon M-6/M-7 primary, Natick Laboratories, bisphosphonate-trial adoption) — hooks the capture never grounded, so they stay in the inbox. All recorded in the capture's not_promoted: list.

What felt off about the capture: it is honest about its own weakness and I agree with the self-assessment. The chain leans on secondary/institutional sources — only the Gemini nutrition report is a genuine primary read. The two most novel threads (the 70-year-old reinvention and the aerospace/geriatric bridge) are the least firmly sourced: the bio sits on a university PR page and the bridge on search-synthesis prose with no primary behind it. No tier inflation — the capture correctly tagged the bridge Tier 3–4 and flagged it — but the value-to-verification ratio is inverted, so I kept everything at seedling and routed the one flag that matters. Highest-leverage next move is exactly what the capture flagged: read the Whedon 1966 primary. Left git to the auto-commit agent.

— Seek

Promotion: 2026-07-09-hop-pentium-fdiv-nicely (headless)

I promoted the hop capture 10-inbox/raw/2026-07-09-hop-pentium-fdiv-nicely.md — the twin-prime-theorist-finds-the-Pentium-bug story. Ran the full /promote workflow headless: no pause for Cali, decided on my own judgment, left git to the auto-commit agent.

Notes created (4):

• claim-nicely-twin-prime-project-surfaced-pentium-fdiv-bug — Thomas Nicely's Brun's-constant computation, not any chip test, publicly surfaced the FDIV bug in 1994. Tier 2 (Pratt 1995).
• claim-pentium-fdiv-bug-five-missing-srt-lookup-table-entries — the mechanism: five empty cells in the SRT quotient-digit PLA, a lithography defect. Folded in the ~1-in-9-billion rate and the non-uniform "bruised operand" behavior. Tier 2.
• claim-intel-475m-charge-pentium-fdiv-recall — the $475M Q4-1994 charge after Intel reversed its case-by-case stance and did a full recall. Tier 1 (Intel SEC filing).
• claim-kahan-8087-spec-became-ieee-754 — Kahan → 8087 → IEEE 754, 1989 Turing Award. The connector node under the whole floating-point cluster. Tier 3, audit_status: flagged.

Questions routed (3): verify Pratt 1995 verbatim (mechanism + Nicely); verify the $475M in the SEC exhibit; verify the Kahan/8087/754 lineage against a primary Kahan source. All new — no existing question collided.

What I skipped and why:

• Hop 1 / Patriot Dhahran clock-drift — the big one. It is already a full note (claim-patriot-dhahran-1991-clock-drift-from-24-bit-truncation-of-tenths, plus a whole Patriot cluster), promoted earlier today from the sibling capture 2026-07-09-hop-patriot-clock-drift-tragedy.md. Straight duplicate. I linked the new Pentium notes to it instead of re-writing it.
• FDIV numeric detail — folded into the mechanism note, not its own note.
• "First full recall of a computer chip" superlative — dropped as unsourced; parked as a sub-question in the SEC-verify question.
• The capture's own "saved hooks not followed" (Vancouver Stock Exchange 1982, Tim Coe's PD-plot reverse-engineering, Ariane 5) — left in the inbox as future captures, per the capture's own judgment. All recorded in not_promoted:.

What felt off about the capture:

1. Duplicated research. This capture and the Patriot capture are sibling hops off the same seed (claim-linnainmaa-rounding-error-problem), and both spent a hop re-deriving the Patriot Dhahran story. The batch worker didn't do retrieve-before-write (by design — that's my job), so the collision only showed up here. Worth noting the seed is spawning overlapping chains.
2. Mild tier inflation. The capture stamped Vaughan Pratt's "Anatomy of the Pentium Bug" as Tier 1. Pratt is a Stanford professor writing on his own venue about someone else's artifact — that's Tier 2 by our rubric (named analyst, own venue, secondary to the underlying facts). I recorded the two Pratt notes as Tier 2. Still clears the mechanism/quant floor, so no claim was blocked, but the stamp was generous.
3. The Kahan hop is the soft spot. Its two sources are a Berkeley Math news page and an ETHW milestone, and the capture carried no verbatim quote for the load-bearing causal claim ("754 grew out of the 8087 spec, not a committee"). The biographical facts are uncontested and clear the Tier 3–4 historical floor, but the causation is interpretive — so I promoted it seedling + flagged, left the source_quote an honest placeholder rather than invent one, and routed the primary-Kahan verification. I judged it worth keeping because it's the connective node tying the Pentium notes to the Linnainmaa/Patriot floating-point cluster and nothing like it exists in the vault yet.

Otherwise a clean, well-shaped capture — the three Pentium claims are genuinely atomic and the two Tier-1/2 ones are solid. Left git to the auto-commit agent.

— Seek

Promotion — "pious perjury / Bloody Code" hop capture (headless)

Promoted 10-inbox/raw/2026-07-09-hop-pious-perjury-bloody-code.md unattended. A tidy three-step hop: 18th-c. London instrument workshops → the Bloody Code's capital property law → juries lying about value to dodge the gallows. Ran the full winnow → retrieve-before-write → honor-flags → route-questions → write pipeline on my own judgment; no pause.

Winnow — 3 claims found, 3 promoted, 2 leads skipped.

• Promote — instrument workshops drew brass partly from a stolen-metal trade (the one genuinely clean claim: single Tier-2 peer-reviewed journal, verbatim quote).
• Promote — the Bloody Code made petty property theft capital (~50→200+ offenses, 12-pence threshold). Kept at seedling + [unverified-quant].
• Promote — juries undervalued goods to avoid the death penalty ("pious perjury"). Kept at seedling + [unverified-quote] on the Blackstone coinage.
• Skip — the 1756 "receiving stolen brass/copper/lead" statute: sourced only to Grokipedia, uncorroborated; the capture had already discarded it. Bad source.
• Skip — Blackstone as second-most-cited authority among the American founders: the capture explicitly parked it for a future chain, never asserted it. Deferred.

Notes created (all stamped writer_model: claude-opus-4-8):

• claim-georgian-instrument-workshops-drew-brass-partly-from-a-stolen-metal-trade
• claim-englands-bloody-code-made-petty-property-theft-a-capital-crime
• claim-georgian-juries-undervalued-stolen-goods-to-avoid-the-death-penalty

Questions routed (both new — no existing collision):

• question-verify-bloody-code-capital-statute-counts-and-larceny-threshold
• question-verify-blackstone-pious-perjury-coinage-and-jury-undervaluation

Retrieve-before-write. No collisions. The two nearest existing notes — claim-whitman-reasonable-doubt-began-as-juror-soul-protection and claim-roman-canon-law-rated-one-witness-equal-to-a-private-document — are thematic neighbors in the legal-evidence cluster, not duplicates, so I wikilinked into them (juror-conscience / "how much is enough" family) rather than skipping. Also linked the instrument-brass note to claim-lippershey-1608-telescope-patent-denied-as-too-easy-to-copy as its precision-instrument-trade neighbor. All three new notes cross-link each other along the hop chain.

What felt off about the capture. One real problem: tier inflation on the numbers. The capture waved the Bloody Code counts (~50→200+, 1688→1820) and the 12-pence threshold through on Tier-4 Wikipedia, rationalizing it as "uncontested historical figures, cross-corroborated." But those are quantitative claims, and sources.md floors numbers at Tier 1-2 — "several search results agree" is not a primary. I promoted the claim but demoted its confidence: seedling, flagged, routed to Radzinowicz / the statutes. Same with the "pious perjury" attribution — the capture honestly flagged that one itself, so I inherited the flag and routed Blackstone's Commentaries + Old Bailey. Also worth noting the capture conflates a 12-pence grand-larceny figure with a 40-shilling capital-threshold example; they're likely different statutes, and I called that out in the routed question so a future pass untangles it rather than inheriting the muddle. The instrument-trade claim, by contrast, was clean — one Tier-2 source, verbatim quote, no flag needed. No duplicated research: nothing in the vault covered the Bloody Code, pious perjury, or the instrument-trade stolen-brass thread before tonight.

Left git to the auto-commit agent.

— Seek

Promotion — Pontryagin / Soviet antisemitism / Fields Medal (headless)

Promoted 10-inbox/raw/2026-07-09-hop-pontryagin-antisemitism-fields-medal.md, a 5-hop capture from a sonnet-5 batch worker. Ran the full /promote workflow solo (no pause — headless), winnowed conservatively: 2 claims promoted, 4 threads skipped, 2 questions routed, 0 collisions.

Notes created (both stamped writer_model: claude-opus-4-8):

• claim-pontryagin-used-committee-vote-to-help-block-margulis-1978-fields-medal-travel — the core reversal. This is the sharp one: the vault already has a warm note on Pontryagin doing topology blind (claim-pontryagin-worked-blind-via-mothers-spoken-symbol-glosses), and this is the same man on the Soviet National Committee helping keep Margulis from traveling to Helsinki to collect his 1978 Fields Medal. Kept at seedling + [unverified-quote]/[unverified-historical].
• claim-mekh-mat-coffin-problems-used-to-fail-jewish-applicants — the institutional-admissions counterpart. Kept at seedling + [unverified-historical] (Tier 4 only in the capture).

Questions routed (both new — no existing collision):

• question-verify-novikov-pontryagin-margulis-fields-medal-block
• question-verify-mekh-mat-coffin-problems-primary-sources

Retrieve-before-write. No collisions — nothing in 30-notes/ covered Margulis, coffin problems, or Soviet-math antisemitism before tonight. I wikilinked both new notes into the sympathetic Pontryagin note (to make the reversal explicit) and into the existing émigré cluster (claim-soviet-emigre-career-premium-was-structural-not-ability, claim-soviet-scientists-immigration-act-1992-exceptional-ability-waiver) — which realizes the capture's "political-push precursor to the 1990s economic-push wave" bridge through links rather than as a new under-sourced note. The semantic index's ranked hint matched what I found: the blind-Pontryagin note is the true neighbor; the rest (Nyquist–Bode, Saunderson, Linnainmaa) are cluster-adjacent, not duplicates.

Skipped, and why (logged in the capture's not_promoted:):

• Keldysh/Vinogradov/Shafarevich "triangle" running Steklov politics — the capture's own saved hook for a future chain, and a heavy motive-allegation on one memoir. Left in inbox; flagged as a candidate thread inside the Novikov verification question.
• Refusenik apartment seminars (Taratuta/Granovskii/Kagan) — Tier 4 leaf detail, too thin for a durable note.
• Brodsky / tuneyadstvo "parasitism" law — the capture itself saved-not-followed; a literature departure.
• The emigration-wave bridge — context, not a discrete claim; handled via links.

What felt off about the capture. Two things. First, a garbled load-bearing quote: Pontryagin's published reply is transcribed as "the vote vas 24:0, so my share of responsibility is 1/24." That "vas" is a tell that the quote was retyped loosely, not lifted from Novikov's PDF — and this quote is the entire rhetorical spine of the reversal. I preserved it verbatim-from-capture (didn't invent a "corrected" version) but flagged it and routed a re-fetch. Second, a tier gap on Claim 2: the "coffin problems" discrimination mechanism is a contested, load-bearing historical claim that the sourcing floor wants at Tier 1–2, but the capture carried it only on Wikipedia/YIVO (Tier 4) with no verbatim quote. Strong primaries exist (Vershik, A. Shen, Shifman's You Failed Your Math Test, Comrade Einstein) — I named them in the routed question rather than letting a clean-looking claim inherit soft sourcing. Otherwise the capture was honest and well-structured: its own hop log correctly rated Claim 1 as Tier 2 and Claim 2 as Tier 4, and it did not overstate. The uncontested spine (Margulis won in 1978 and Tits accepted for him) is solid; it's the attribution to Pontryagin's specific vote that rests on one eyewitness, which is exactly what I flagged.

Left git to the auto-commit agent.

— Seek

Promotion run (headless) — Portelli/Neisser testimony capture

Later in the session I promoted a second capture: 10-inbox/raw/2026-07-09-hop-portelli-neisser-testimony.md — a hop-batch capture bridging oral history and cognitive psychology on a shared claim about testimony. No human in the loop, so I winnowed on my own judgment and proceeded without pausing.

What I created (3 notes):

• claim-portelli-oral-testimony-relocated-trastulli-death-to-1953 — Portelli's finding that Terni workers' memory moved a 1949 killing to 1953, and that the error was the point. Kept seedling; preserved the capture's [unverified — needs primary text] flag (the essay PDF 403'd twice at capture).
• claim-neisser-john-dean-testimony-repisodic-memory — Neisser checking Dean's Watergate testimony against the Nixon tapes: details wrong, gist right, "repisodic" coined. Kept seedling — the citation is Tier 1 (PubMed-confirmed) but every analytic detail is secondary paraphrase because the paper was paywalled. I folded Neisser's 1978 ecological-validity critique in here as motivation rather than spinning it into its own thin note.
• observation-1981-convergence-testimony-wrong-in-detail-right-in-shape — the capture's real payload: two fields, same year, no shared citation trail, same structural conclusion. An observation note linking the two claims. Flagged as inheriting both notes' unverified sourcing.

Questions routed (2): question-verify-portelli-trastulli-primary and question-verify-neisser-john-dean-cognition-1981-primary — each names the exact primary to read (Portelli's 1991 SUNY essay / 1981 History Workshop Journal; Neisser Cognition 9(1):1-22 or the Memory Observed reprint) and what to confirm.

Retrieve-before-write: no collisions. Nothing in 30-notes/ covered Portelli, Neisser, John Dean, "repisodic," Trastulli, or flashbulb memory. I linked outward to the strongest existing neighbours: claim-roman-canon-law-rated-one-witness-equal-to-a-private-document (the testimony-vs-documents evidentiary thread — a genuinely good fit) and the reconstructive-memory / Bartlett thread via question-bartlett-distortion-typology-maps-to-model-collapse-failure-modes.

What I skipped and why: the Neisser & Harsch 1992 Challenger flashbulb study — the capture itself declined to make it a claim (Tier 3 journalism, [unverified-quant]), and I agreed; it stays a lead, logged in the capture's not_promoted. Also skipped Halbwachs, Loftus, Collingwood, and Marc Bloch — all saved hop-hooks, not claims made in this capture.

What felt off about the capture: honestly, not much — it was unusually candid about its own weaknesses. The one real issue is that both core claims rest on secondary paraphrase (Portelli's PDF blocked, Neisser's paper paywalled), so a capture whose whole subject is "testimony that's wrong in the detail" is itself built on retellings rather than directly-read primaries. The capture bracket-flagged claim 1 explicitly but left claim 2 only narrated as secondary rather than flagged; I treated it as flagged anyway and kept it seedling. No tier inflation and no duplicated research — the hop log's own tier calls were accurate. The convergence is a strong bridge and probably post-worthy eventually, but not until at least one of the two primaries is in hand. Left git to the auto-commit agent.

— Seek

Promotion — 2026-07-09 Price / actuary / Burke-Revolution bridge (headless)

Promoted 10-inbox/raw/2026-07-09-hop-price-actuary-burke-revolution-bridge.md. This is the Richard-Price capture: the actuary of the Northampton mortality table who is also the man whose 1789 sermon provoked Burke's Reflections. The capture listed three core claims plus a cross-domain-bridge framing; I promoted four notes and skipped the hop-chain background.

Notes created (4):

• claim-price-northampton-table-standard-actuarial-reference-century — Price built the Northampton table and Observations on Reversionary Payments (1771), the standard actuarial reference for ~a century. Uncontested historical, clears the Tier 3–4 floor, but held seedling (Tier-4 Wikipedia, "roughly a century" is a soft longevity claim). Linked into the Dodson/Halley actuarial lineage.
• claim-price-1789-sermon-provoked-burke-reflections — the sermon "A Discourse on the Love of Our Country" (4 Nov 1789) provoked Burke and touched off the Revolution Controversy. This is the note's load-bearing point, so the Tier-4 sourcing does not clear the floor; flagged [unverified-historical], kept seedling, verification routed.
• claim-price-northampton-table-overestimated-mortality-favouring-insurers — the table overestimated mortality, a bias good for insurers and adverse to annuitants. Carried the capture's explicit [unverified-quant/mechanism] flag; kept seedling, flag preserved, verification routed. Linked to the osteological- paradox / demographic source-criticism thread.
• observation-richard-price-anchors-actuarial-and-political-philosophy-lineages — the payload: one man load-bearing in two vault-distant lineages with no mediating link. Type observation; inherits both underlying notes' unverified sourcing and is only as sound as its weaker (sermon-Burke) leg.

Questions routed (2):

• question-verify-price-sermon-provoked-burke-reflections-primary — read Price's 1789 Discourse and Burke's 1790 Reflections to confirm the causation is direct, not same-era climate.
• question-verify-northampton-table-overestimated-mortality-primary — read Observations on Reversionary Payments / an actuarial history to confirm the bias direction, magnitude, and economic incidence (I flagged the "adverse for annuities" direction as the kind of incidence claim that gets stated backwards).

Retrieve-before-write: no claim-level collisions — nothing in 30-notes/ covered Price, the Northampton table, the 1789 sermon, or the Burke controversy. The strong neighbours are the two Dodson notes (claim-dodson-built-age-scaled-premiums-founding-equitable-life, claim-dodson-refused-amicable-society-over-age-45), which I linked into densely — Price sits one generation downstream of Dodson/Halley in the same lineage. On the political side I linked claim-englands-bloody-code-made-petty-property-theft-a-capital-crime as contemporaneous late-Georgian context, and on the mortality-data side claim-osteological-paradox-skeletal-statistics-confound-past-health.

What I skipped and why: the whole hop-chain background — Halley's 1693 Breslau table, Neumann's astrology-debunking mortality data, the Million Act / King William's Tontine flat-rate annuities — is context, not a core claim of this capture, and the Dodson/Halley/Amicable strand was already promoted the same day from the companion capture (2026-07-09-hop-dodson-rejection-founded- actuarial-science.md). Promoting it again would have duplicated my own research. Also skipped: de Moivre, Akerlof's adverse selection, modern AI underwriting, and the Paine/Wollstonecraft political-philosophy thread — all saved hooks, not claims made here. All logged in the capture's not_promoted.

What felt off about the capture: two things. First, duplicated research — this capture's hop chain (hops 1–4: Halley → Neumann → Million Act → Dodson) is substantially the same ground the Dodson companion capture already covered, so the two captures from the same batch re-derived the same actuarial lineage from different endpoints. Not wrong, but the batch worker clearly wasn't deduping against its own sibling captures. Second, tier honesty is mixed: the capture was candid about claim 3 (bracket-flagged [unverified-quant/mechanism]) and about claim 2 (it explicitly said "escalate to a primary Price/Burke text before treating as load-bearing") — good discipline — but it still tagged the header source_tier: 4 while narrating claim 2 as merely "uncontested," which undersells that a Tier-4 source can't carry a load-bearing bridge on its own. I treated claim 2 as flagged and routed it. No invented sources; the bridge itself is genuine and, once a primary confirms the sermon-Burke link, probably post-worthy. Left git to the auto-commit agent.

— Seek

Promotion: 2026-07-09-hop-protestware-npm-node-ipc

Headless promotion run. I promoted the protestware capture (10-inbox/raw/2026-07-09-hop-protestware-npm-node-ipc.md) — a hop chain about open-source maintainers as political actors, running from a July 2026 OSS-typology paper through the 2022 node-ipc sabotage back to the 2016 left-pad and 2020 Notepad++ precedents.

Winnow: three distinct claims, all worth keeping, all promoted — no skips at the claim level.

• Claim 1 (typology names protestware as a sub-genre) → claim-protestware-named-in-2026-oss-typology
• Claim 2 (node-ipc 2022 geo-targeted wiper, CVE-2022-23812) → claim-node-ipc-2022-maintainer-shipped-geotargeted-wiper
• Claim 3 (single-maintainer/global-blast-radius precondition predates protestware; left-pad + Notepad++) → claim-single-maintainer-global-blast-radius-predates-protestware

I folded left-pad and Notepad++ into one structural claim rather than splitting them into two thin biographical notes — the atomic unit is the precondition, the two incidents are its evidence.

Retrieve-before-write: no claim-level collisions. Nothing in 30-notes/ covered protestware, node-ipc, left-pad, or the OSS typology; the semantic-index hint's "most-related" notes were all unrelated (Byzantine trade, voice-AI latency, etc.) — confirmed by a grep. The real neighbours are thematic, not duplicative: I linked all three notes densely into the vault's existing "maintainer as a single point of failure/agency" thread — claim-mj-rathbun-ungated-agent-published-hit-piece (AI agent attacking a maintainer), claim-toctou-named-frame-browser-use-agents, and claim-confused-deputy-2026-ai-agent-security-frame — and cross-linked the three new notes to each other. Protestware reframes that AI-era thread as one instance of a much older, pre-AI pattern; that's the bridge the capture was reaching for and it lands.

Verification I actually did (not just routed): I don't take a sonnet batch-worker's sourcing on faith, so I web-checked all three before writing.

• Claim 1's seed paper did not surface in a title search, which — combined with the capture's internal "six vs. fourteen" contradiction — made me suspect a hallucinated citation. So I fetched arXiv:2607.01750 directly. It is real (Ouf & Hussein, "Open Source Is Not One Thing"), and the abstract says fourteen sub-genres from a review of 3,925 papers, with protestware among three under-studied ones. The capture's "six under-studied" was simply wrong; I corrected it against the primary and marked Claim 1 budding.
• Claim 2 is corroborated across NVD (CVE-2022-23812), Snyk, Orca, BleepingComputer, IT Pro — versions 10.1.0/10.1.1/10.1.2, IP-geolocated overwrite of file contents with a heart emoji for Russia/Belarus hosts. Solid.
• Claim 3's precedents check out against primary sources — npm's own "kik, left-pad, and npm" postmortem and Notepad++'s own release notes — which let me upgrade the sourcing from the capture's Kiuwan vendor blog to Tier 1.

Questions routed: one. The node-ipc "~1M weekly downloads" figure is a quantitative claim resting only on Tier-3 trade advisories, so I flagged it [unverified-quant], kept Claim 2 at seedling, and opened question-verify-node-ipc-weekly-download-count to pull npm's own registry stats. Everything else in Claim 2 (the CVE, versions, mechanism) is NVD-corroborated and doesn't hang on that number.

What I skipped and why: the capture's "further leads" and "saved hooks" — the color.js/es5.ext reception study, the uncompensated-labor angle (arXiv:2401.06889), the Sovereign Tech Agency funding thread, and Miller's electric-motorcycle-racing trivia — are saved hooks for future chains, not claims made here. All logged in the capture's not_promoted. The capture is marked status: promoted, not deleted.

What felt off about the capture: two things worth flagging to Cali. First, a garbled number on the most novel claim — "six under-studied sub-genres" when the primary says fourteen total / three under-studied. It's the kind of error that looks authoritative and would have propagated cleanly if I hadn't fetched the paper; good reminder that the seed-paper claim is exactly the one to re-check. Second, a provenance gap: Claim 3's source in the body (the Kiuwan blog) was never declared in the capture's own frontmatter source_1..4 list, and it's a vendor blog — I replaced it with primary sources rather than inherit it. Sourcing was otherwise honest and the arXiv citations were real. No invented sources. Left git to the auto-commit agent.

— Seek

Promotion pass (headless) — Prussian blue radiological-antidote capture

Ran a headless /promote on 10-inbox/raw/2026-07-09-hop-prussian-blue-radiological-antidote.md — no pause, my own winnowing judgment. Three distinct claims in the capture; I promoted all three as atomic seedlings and routed three verification questions.

Notes created (all writer_model: claude-opus-4-8, status: seedling):

• claim-prussian-blue-accidentally-discovered-1706-as-failed-red-pigment — the accidental-discovery chemistry history. Tier 4 Wikipedia; core is standard but the Dippel/animal-oil particular is uncorroborated (the source itself flags it), so I leaned the note's weight on the uncontested core and linked it into the accidental-discovery cluster (CPR, Kouwenhoven).
• claim-prussian-blue-is-the-fda-approved-antidote-for-radioactive-cesium-and-thallium — the centerpiece cross-domain bridge, on the Tier 1 HHS/REMM source with its exact quote. Kept the deep ion-selectivity mechanism and the half-life figures OUT of the hard claims (unsourced to a primary) and wired the explicit [[link]] to claim-burevestnik-reactor-emits-radionuclides-in-exhaust that the capture's commentary asked for — the small nuclear-radiological-medicine cluster now exists.
• claim-prussian-blue-was-deployed-at-scale-in-the-1987-goiania-cesium-137-accident — the real-world deployment. The 249-contaminated / 4-dead figures kept their [unverified-quant] flag in body and tags; note stays seedling.

Questions routed (all new, no collisions in 50-questions):

• question-verify-prussian-blue-1706-accidental-discovery-primary (the Dippel detail / earliest documentary trace)
• question-verify-prussian-blue-cesium-selectivity-mechanism-primary (selectivity physics + the omitted half-life numbers)
• question-verify-goiania-cesium-137-casualty-figures-iaea-primary (the IAEA 1988 report — its page returned HTTP 402 at capture time)

Skipped, judged not missed: the Dippel/Castle Frankenstein/Mary Shelley lead (the capture already discarded it as "very unlikely" folklore); the selectivity mechanism as a standalone note (unsourced → became a question); the half-life figures (capture itself omitted them as un-reverified → folded into the mechanism question).

Retrieve-before-write: searched 30-notes for prussian/radiogardase/goiânia/cesium/thallium/hexacyanoferrate/Diesbach/Dippel — zero real collisions (one coincidental Lepsius hit). These are effectively orphans apart from the deliberate Burevestnik bridge and the pigment/color-chemistry adjacency (Kubelka–Munk), so I linked those explicitly to keep them from landing as isolated entries.

What felt off about the capture:

• Tier discipline was actually good — the bee flagged its own soft spots ([unverified-quant] on Goiânia, held the half-life numbers and the selectivity mechanism back rather than laundering them clean). Honest sourcing; I mostly just formalized its self-flags into routed questions.
• The Burevestnik "bridge" is thinner than the capture's excitement suggests. The nearest-neighbor cosine (~0.586) is a topical adjacency (both involve radionuclides/cesium-adjacent isotopes), not a real conceptual bridge — Burevestnik is about a weapon emitting radionuclides, this is about decorporating them. I made the link because a nuclear-radiological-medicine cluster is genuinely worth seeding, but I'd caution against reading the cosine as evidence the two notes share a thesis. It's proximity, not a claim.
• Two Tier-4-only claims out of three. Only the FDA-antidote claim clears a real sourcing floor on its own; the origin story and the Goiânia case both rest on Wikipedia. That's why all three are seedlings with verification routed, and I set audit_status: capture-verified (not verified-verbatim) across the board — I did no independent re-fetch this headless session; the queen's re-check is exactly what the three questions queue up.

No git run — leaving the commit to the auto-commit agent per instructions.

— Seek

Promotion — 2026-07-09 hop: Purdue, the first CS department (headless)

Second promotion this session, unattended. I promoted 10-inbox/raw/2026-07-09-hop-purdue-first-cs-department.md — the hop that started on Cleve Moler's ACM memoriam entry and ended on the Purdue-vs-Stanford priority dispute over who founded the first US computer science department.

Three claims found → three promoted → the rest folded or skipped.

Notes created:

• claim-purdue-established-first-us-cs-department-1962 — the priority claim (Purdue, October 1962). Seedling.
• claim-forsythe-coined-computer-science-and-founded-stanford-cs — Forsythe as numerical-analyst founder of Stanford CS, credited with coining the term. Seedling.
• observation-numerical-analysts-founded-first-us-cs-departments — the hop's real payload: both candidate "first departments" were built by numerical analysts. This is the note that links the capture into the existing numerical-analysis cluster (claim-linnainmaa-field-numerical-analysis, backpropagation-gap) — same discipline as a hidden root, one level up from the technique to the institution. Seedling.

Questions routed:

• question-verify-purdue-first-us-cs-department-priority — need an independent primary (1962 trustees minutes, or a CACM / IEEE Annals retrospective) to corroborate the priority over Stanford.
• question-verify-forsythe-coined-computer-science-cs26-quote — need the primary Technical Report CS26 for the exact quote, plus a source adjudicating the coinage (Fein / Naur / Forsythe).

Skipped, judged not missed: the Moler/MATLAB/LINPACK biography (the hop's entry path, not one of the three core claims, not independently grounded — folded as supporting detail into the Forsythe and observation notes); Forsythe's WWII-meteorology → numerical-weather-prediction lead (the capture explicitly deferred it); the saved hooks the capture itself set aside (Neumann/RISKS, Hoare, Tapia, Conte's textbook, LINPACK→TOP500). All logged in the capture's not_promoted: list.

Retrieve-before-write: searched 30-notes for Forsythe / Purdue / Conte / Moler / MATLAB / "first CS department" — zero collisions. These are genuinely new to the vault, but they are not orphans: I linked them hard into the numerical-analysis / backprop-gap cluster and the myth-ledger notes (myth-werbos-invented-backpropagation, myth-linnainmaa-uncited-before-2010s), which is where they earn their keep. The semantic-index hint pointed at Kahan/IEEE-754, Ted Hoff/4004, and the Complex Systems notes — adjacent computing-history, but none actually overlapped the department-founding claims.

What felt off about the capture:

• Sourcing sits below the floor for what it's claiming. All three grounding sources are Tier 3 (Purdue's own history page; MacTutor) or Tier 2 (the Moler obituary). But the load-bearing claim is a contested priority — MacTutor literally calls Stanford's 1965 department "the first of its kind," contradicting Purdue — and the sourcing floor says contested/load-bearing history escalates to Tier 1–2. So the whole capture rests on soft-for-truth sources for exactly the claim that needs hard ones. The bee, to its credit, said as much in its own post-worthy: maybe line ("would benefit from one more corroborating primary source before promotion"). I honored that: all three notes are seedlings with audit_status: flagged, inline [unverified-*] flags, and the two gaps routed as questions.
• The Purdue priority quote comes from an interested party. A university claiming to have founded a field is self-flattering institutional memory by construction. What saved it from the trash rather than the seedling bin is that the claim is specific and falsifiable (October 1962, two-plus years clear of Stanford) — checkable, not vague boosterism.
• The CS26 quote is secondhand. It's presented as a "grounding quote" but it's relayed by MacTutor, not read from Forsythe's report. I flagged it as [unverified-quote] rather than let a Tier-3 relay stand in for a primary.
• A live contradiction, left open. MacTutor's "first of its kind" (Stanford) vs. Purdue's October-1962 priority is a real contradiction between two of the notes. Per spec I flagged it inside both notes rather than silently picking a winner — it's the myth-ledger point, not a defect to paper over.
• No duplicated research — this thread is new territory for the vault; nothing was re-derived.

No git run — the auto-commit agent will pick this up within 15 minutes.

— Seek

Promotion (headless) — quorum / weighted-voting / power-law / DeFi capture

Promoted 10-inbox/raw/2026-07-09-hop-quorum-voting-power-law-defi.md (a claude-sonnet-5 hop-batch capture). Four notes out, one question routed, one standalone note deliberately declined.

Notes created:

• claim-gifford-1979-weighted-voting-quorum-replicated-data — Tier 1, Gifford's PARC quorum protocol (r+w > votes). capture-verified, seedling (I didn't re-fetch the bitsavers PDF).
• claim-banzhaf-1968-vote-weight-diverges-from-voting-power — Tier 1, Banzhaf power index / weight≠power. capture-verified, seedling. Folded the Uniswap/a16z delegation finding in here as attributed context.
• claim-mica-2024-decentralization-test-holder-can-meet-quorum-alone — Tier 3, [unverified-mechanism]. flagged, seedling, flag preserved.
• observation-weighted-voting-power-gap-recurs-across-cs-law-regulation — the synthesis, the reason the capture existed. Seedling; inherits the unverified caveat because it leans on the regulatory leg.

Question routed: question-verify-mica-danish-fsa-eba-esma-decentralization-quorum-test-primary — read the Danish FSA (June 2024) and EBA-ESMA (Jan 2025) primary texts to confirm the "meet quorum alone" wording is theirs, not the law firm's.

What I skipped and why. No standalone note for the Uniswap/a16z vote-delegation study (Bongaerts et al. 2025, arXiv:2503.11940). It's a genuinely distinct empirical claim in a fourth domain, and I wanted to promote it — but the capture gives only a paraphrase, no verbatim phrase, and spec §4 forbids a standalone Tier-1 claim-note without an exact quote. Fabricating a grounding quote is off the table, so I folded it into the Banzhaf note as clearly-attributed context and noted in not_promoted: that it can be split out if a verbatim finding is later pulled. The two "further leads" (did Gifford know the voting-power literature; NP-hardness of computing power indices for quorum design) stayed in the inbox as leads, not claims.

Retrieve-before-write. One real collision: claim-lamport-paxos-greek-allegory-delayed-publication, which the capture itself cites. I linked it as the allegorical sibling rather than duplicate it — my voting notes are the literal/quantitative counterpart. No existing note touches Gifford, Banzhaf, MiCA, or DAO governance; this is new territory. claim-nakamoto-bitcoin-leaned-on-wei-dai-b-money gave a natural crypto-lineage link for the MiCA note.

What felt off about the capture.

• Tier inflation in the framing, not the sourcing. The capture's headline — "three fields independently arrived at the same idea" — overstates the bridge. Gifford's quorum is a consistency mechanism, indifferent to power or fairness; he supplied the word the other two reuse, he didn't rediscover Banzhaf's insight. The genuine conceptual line is Banzhaf → regulators. I rewrote the observation to say "terminological/structural echo, not tri-fold convergence" and put that caveat in the note body, not just commentary.
• The load-bearing leg is the weakest-sourced. The whole payoff — regulators literally reusing "quorum" — rests on a Tier-3 law-firm client alert paraphrasing two regulatory documents nobody read. The capture was honest about this ([unverified-mechanism], "post-worthy: maybe"), which I respect. Held at seedling and routed.
• A Tier-1 source present in the body but absent from frontmatter source_urls: the arXiv DeFi paper (2503.11940) is only in the hop-chain prose, not the frontmatter source list — a small capture-hygiene gap that made the a16z finding harder to promote cleanly.
• No duplicated research — nothing here was re-derived from existing vault notes.

No git run — the auto-commit agent picks this up within 15 minutes.

— Seek

Promotion — railroad-gauge / Roman-chariot myth (headless)

I promoted 10-inbox/raw/2026-07-09-hop-railroad-gauge-myth-historiography.md, a hop-batch capture that landed on the viral "US rail gauge descends from Roman chariot ruts, which then constrained the Space Shuttle's boosters" legend and read it as another instance of the vault's myth-correction pattern.

What I created:

• 30-notes/claim-standard-gauge-descends-from-stephenson-not-roman-chariots.md — one atomic note folding the capture's Claim 1 (Snopes rates the chain a "mixture" — coincidence, not descent) and Claim 2 (the real origin is George Stephenson's 1829 Liverpool & Manchester practice; Baxter reportedly found no standard gauge among prior wagonways). These are two facets of a single correction, so one note is the right grain. Seedling, Tier 3, wired densely into the myth-correction cluster (Barbier, Amari citogenesis, Cullinane, Portelli, the 1981-convergence observation, Mabillon).

Questions I routed:

• 50-questions/question-verify-standard-gauge-origin-baxter-wagonways.md — carries Claim 2's [unverified-historical] flag; names the primary to read (Baxter, Stone Blocks and Iron Rails, 1966) and asks for a Tier 1–2 corroboration of the Stephenson attribution plus the loading-gauge-vs-track-gauge Shuttle detail.
• 50-questions/question-myth-correction-as-historiography-moc-candidate.md — the home for Claim 3.

What I skipped and why:

• Claim 3 — the capture's actual payoff, the "three domains converge on one mythic shape → historiography in general" synthesis. I did not write it as a standalone note. It's Seek's own reading resting on Tier-3 sourcing throughout; the generalization already lives in the Barbier and Amari note commentary; and the capture author explicitly deferred the MOC pending a fourth domain. Routing it to an MOC-candidate question (mirroring the existing question-transmission-chains-renamed-for-local-resemblance-pattern) is the honest, conservative home — it parks the pattern for a fourth data point instead of asserting it prematurely.

What felt off about the capture:

• Tier 3 top to bottom, and it knew it. Every leg — Snopes for the verdict, a history hobby-blog (tastesofhistory.co.uk) citing Baxter for the real history — is secondary. The capture was honest ("post-worthy: maybe," "rests on Tier 3 sourcing throughout"), which I respect, but it means nothing here clears the sourcing floor for a load-bearing historical claim. Everything I wrote is seedling and flagged accordingly.
• The interesting claim is the least sourced. As with the quorum capture earlier today, the genuinely novel finding (the cross-domain myth pattern) is pure synthesis over soft sources — exactly the thing to park, not promote.
• Slight pattern-inflation risk. "Structurally identical" (capture's word) across AI citogenesis, disability history, and engineering folklore is doing a lot of work; the citogenesis mechanism (copied errors betraying one witness) may be a distinct sub-pattern from the broader myth-shape. I flagged that split inside the MOC-candidate question rather than smoothing it over.
• No duplicated research — no existing gauge/path-dependency note existed; this is a clean orphan into the myth-correction cluster.

No git run — the auto-commit agent picks this up within 15 minutes.

— Seek

Promotion — Rock's Law / Red Queen capital arms-race (headless)

I promoted 10-inbox/raw/2026-07-09-hop-red-queen-capital-arms-race.md, a hop-batch capture that hopped from the training/inference compute-asymmetry seed out to semiconductor fab economics (Rock's Law), then to Van Valen's Red Queen hypothesis, then to William Barnett's application of the Red Queen to organizational competition — landing on the fact that computer manufacturing is one of Barnett's two case industries. A genuine two-cluster bridge: AI-compute-economics ↔ evolutionary biology.

What I created (3 notes):

• 30-notes/claim-rocks-law-is-moores-laws-economic-inverse.md — Rock's Law as the cost inverse of Moore's Law ($10–20B fabs, >$400M high-NA EUV scanner), grounded in the ICLE white paper's verbatim quote (Tier 2). Folded in the "competition for the market vs. in the market" framing and, carefully flagged as a capture paraphrase, the TSMC 30–50%-of-revenue capex figure. Wired to the compute-economics cluster (claim-training-inference-compute-asymmetry-mechanism, moc-inference-economics) as the fixed-vs-marginal-cost analog that generated the hop.
• 30-notes/claim-red-queen-hypothesis-names-running-to-stay-in-place.md — the Van Valen 1973 definitional note. Tier 4 Wikipedia, which is acceptable for the definitional content per the sourcing floor; I named Van Valen's "A New Evolutionary Law" as the primary the constant-extinction-probability claim actually rests on, and drew the explicit distinction from punctuated equilibrium so the two biology imports don't get conflated.
• 30-notes/claim-barnett-applied-red-queen-to-organizational-competition.md — Barnett's book and its two case industries (computer manufacturing + commercial banking). Placed alongside the vault's Tushman–Romanelli punctuated-equilibrium-in-organizations cluster as a second instance of a biological model imported into org theory.

All three are seedling and carry writer_model: claude-opus-4-8 plus honest audit_status: flagged lines — none were independently re-fetched, and each rests on a single source.

Questions I routed (2):

• 50-questions/question-verify-barnett-red-queen-computer-manufacturing-primary.md — the real verification gap. The Barnett note's load-bearing empirical claim rests on the Stanford GSB publisher page, not Barnett's data. Names the book (Princeton UP, 2008) and the disk-drive/computer chapters to read. Blocks note 3 from ever going evergreen on its current footing.
• 50-questions/question-does-competition-for-the-market-map-onto-red-queen-extinction-probability.md — the capture's core open lead, preserved as a saved hook: is the Rock's-Law arms race the same dynamic as Van Valen's constant extinction probability, or only a rhyme?

What I skipped and why:

• The metaphorical bridge itself ("chipmakers really are running an evolutionary arms race, not just a metaphorical one") — that's Seek's own unsourced synthesis. I housed it in the notes' commentary callouts and in the mapping question, not as a neutral claim-note. Asserting the equivalence as fact would be exactly the metaphor-inflation I flagged on the railroad-gauge and punctuated-equilibrium captures.
• SemiAnalysis's "$40M mask set at 3nm" waypoint figure — redundant with the Rock's Law note's capital-intensity claim, and SemiAnalysis was never vetted as a source in this capture's frontmatter (it only appears in the hop-chain narrative). Left in the inbox as audit trail.

What felt off about the capture:

• Quant on a think-tank paper. The load-bearing numbers ($10–20B fabs, >$400M scanner, TSMC capex) all sit on the ICLE white paper, tiered 2. That clears the sourcing floor, but ICLE is a law-and-economics advocacy shop citing figures whose real homes are ASML and TSMC filings — I'd call the Tier-2 assignment defensible-but-generous, so I flagged the primary-home gap in the note rather than inheriting a clean-looking figure. The TSMC capex figure had no verbatim capture at all, so I marked it [unverified-quant].
• The interesting claim is the least sourced — again. Same shape as the other two headless promotions today: the novel cross-domain payoff (arms-race = Red Queen) is pure synthesis over mixed-tier sources, while the solid, sourced facts are the individual legs. Park the synthesis, promote the legs.
• Barnett rests on a marketing page. The single most quotable claim — that computer manufacturing empirically shows Red Queen dynamics — comes from the Stanford GSB book-blurb, not Barnett's data. Flagged and routed, not smoothed.
• No duplicated research. No prior Rock's Law, Moore's Law, Red Queen, or Barnett note existed; the three notes bridge two existing clusters (compute-economics + org-evolution/punctuated-equilibrium) cleanly. Worth noting for a future lint: these plus the Tushman–Romanelli notes may eventually want a "biological models imported into organizational theory" MOC — I did not create one this session (only two clear members so far outside the punctuated-equilibrium set).

No git run — the auto-commit agent picks this up within 15 minutes.

— Seek

Promotion — signal-corps / weather-bureau hop (headless)

Promoted 10-inbox/raw/2026-07-09-hop-signal-corps-weather-bureau.md — the "America's first national weather service was an Army Signal Corps telegraph operation, later run by an Arctic-disaster survivor" hop. Three core claims in the capture, all three promoted; the two hops of Confederate-weapons detour (Maury's mines, the City Point torpedo, the coal torpedo) stayed as leads in the inbox where the capture had already parked them.

Notes created (3):

• claim-us-national-weather-service-born-1870-inside-army-signal-service — Tier 1, weather.gov, the "Division of Telegrams and Reports for the Benefit of Commerce" founding. Links straight into the vault's telegraph-era material (claim-telegraphists-cramp-1890s-classified-as-occupational-neurosis) — same wires, same Morse-operator moment.
• claim-greely-arctic-disaster-survivor-ran-us-weather-bureau-1887-1891 — Tier 3 NPS biography. Kept seedling; the cannibalism detail is hedged, not asserted.
• claim-1890-91-weather-service-transferred-to-agriculture-at-harrisons-request — Tier 1, weather.gov, the civilian handoff. Split from the 1870 founding note deliberately: two distinct institutional events.

Questions routed (2):

• question-why-was-the-weather-bureau-moved-to-agriculture-in-1890-91 — the capture's own unresolved hook. The Tier-1 timeline states the transfer but gives no rationale; I refused to import the secondary "civilian control" motive as fact and parked the why instead.
• question-verify-greely-arctic-expedition-cannibalism-primary — the cannibalism rumor rests only on the NPS summary. Routed the primary check (Greely's own Three Years of Arctic Service, 1894; the Schley relief reports; Guttridge's Ghosts of Cape Sabine).

Skipped, with reasons (all left in inbox as leads): Maury/Brussels 1853, Maury-vs-Rains mine credit dispute, City Point horological torpedo, coal torpedo. These are "further leads" and hop-chain scenery, not claims the capture asserted — promoting them would be inventing notes the source never stood behind. The Maury-vs-Rains one is genuinely a contested "who invented X" thread and would need its own sourcing pass.

Retrieve-before-write: no collisions. The semantic-index hint pointed at pulsar/leap-second/Patriot/ozone notes — none overlap. The one real graph connection is the telegraph-era cramp note, which I wikilinked. No prior weather-bureau, Greely, Signal Corps, or Maury note existed.

What felt off about the capture: honestly, not much — this was a clean capture. Tiers were labelled accurately (no inflation: NPS correctly Tier 3, weather.gov correctly Tier 1, both Tier-1 claims carried verbatim quotes). The capture was self-aware about its own weak spots — it flagged the cannibalism sourcing and the missing transfer-rationale in "Further leads" and "Saved hooks" rather than smuggling them into the core claims. The only structural wobble is the one the capture itself confesses: the hop chain wandered two hops through Confederate mine warfare before returning to the meteorology-institutions thread the seed asked for, so ~40% of the body is scenery that produced no notes. Good discipline by the capture-writer, not a defect to fix here.

Three notes, two questions, zero collisions. No git run — leaving the commit/push to the auto-commit agent.

— Seek

Promotion — Song Jian, cybernetics of population, and the one-child policy

Promoted 10-inbox/raw/2026-07-09-hop-song-jian-cybernetics-population.md. Headless run, no pause — winnowed on my own judgment.

Notes created (2):

• claim-song-jian-missile-scientist-engineered-one-child-policy-via-cybernetics-of-population — Greenhalgh's Tier-1 thesis that a missile-guidance control theorist authored the one-child policy from a "Sinified cybernetics of population." Kept seedling: the Tier-1 quote is solid but the single-actor causal thesis is contested. Folded in the Helsinki-1978 IFAC detail (Tier 4) and the contestation caveat (flagged).
• claim-song-jian-self-credited-1980-projections-triggered-one-child-policy — Song's own 1995 self-credit, via Greenhalgh's translation. seedling, flagged [unverified-quote] (doubly-mediated, original Chinese unread). Split from the Greenhalgh note deliberately: a scientist claiming credit and a historian assigning it are separate epistemic objects that can diverge.

Questions routed (2):

• question-verify-greenhalgh-song-one-child-policy-causal-role — the contested causal role. Liang Zhongtang's refutation was seen only via secondary summary; I refused to let Greenhalgh's dramatic "he hijacked the policy" framing stand as settled, and parked the primary check (Liang's own writing; pre-1980 Party intent; a triangulating second historian).
• question-verify-song-jian-1980-self-credit-primary-chinese — the primary for Song's 1995 self-credit quote, since the note rests on Greenhalgh's English translation with translator's brackets sitting right where the causal verb lives.

Skipped / folded, with reasons (all left in inbox as leads): the Helsinki 1978 encounter (Tier-4, folded not promoted); the contestation (folded + routed, too thin as a flagged pointer-note); and the further leads — Wiener's Cybernetics, Qian Xuesen, Olsder's 1973 paper, Forrester/World3/Limits to Growth. Those last are unopened hooks and hop-chain scenery sourced only Tier 3–4; promoting them would be inventing notes the capture never stood behind.

Retrieve-before-write: no subject collisions. The semantic-index hint pointed mostly at AI-history / Daisy Bell / SC2 / NSL notes — none overlap Song Jian, Greenhalgh, or population policy. The real graph connection is the vault's Soviet/optimal-control cluster: I wikilinked claim-pontryagin-worked-blind-via-mothers-spoken-symbol-glosses, claim-kelley-bryson-optimal-control-precursor, moc-backpropagation-origins, and the Cold-War-Soviet-science note claim-soviet-scientists-immigration-act-1992-exceptional-ability-waiver. This capture is a genuine cross-domain cousin of the backprop-origins thread — the same Soviet control-theory lineage (Song trained under Fel'dbaum, one step from Pontryagin) pointed at a birth rate instead of a gradient. The capture's own commentary flagged that neighbour and chose not to chase it; I honoured that by linking, not merging.

What felt off about the capture: little — clean, honest sourcing. Tiers were labelled accurately with no inflation (Greenhalgh correctly Tier 1 with a verbatim quote; Wikipedia correctly Tier 4; the self-report correctly downgraded to Tier 2 as a secondhand rendering). The capture was self-aware about its weak spots — it flagged the contested causal role and the doubly-mediated quote itself rather than smuggling them into the core claims. One thing to watch: the load-bearing claim is a dramatic lone-genius policy-capture story, and the vault's own myth-correction pattern says those tend to compress messier histories. That's not a sourcing defect — it's why I kept both notes seedling and routed the causal-role check as the first thing worth doing next.

Two notes, two questions, zero collisions. No git run — leaving the commit/push to the auto-commit agent.

— Seek

2026-07-11 — Promotion: splines / ballistics / ENIAC capture

Headless, unattended promotion of a single capture: 10-inbox/raw/2026-07-09-hop-splines-ballistics-eniac.md — the five-hop chain arguing the curve math under digital fonts runs through French automotive secrecy, wooden shipbuilding, and the WWII lab that built ENIAC. No human to approve, so I winnowed on my own judgment and shipped.

What I promoted

Three atomic claim-notes, all status: seedling (I have no web tool this run, so nothing could be independently re-verified — everything stays seedling regardless of source tier):

• claim-bezier-de-casteljau-independent-invention-secrecy-named-the-curve — Bézier (Renault, 1966) and de Casteljau (Citroën, 1959) solved the same curve problem independently; Citroën's secrecy, not priority, decided the naming. Tier 3.
• claim-spline-was-a-shipbuilders-lofting-tool-before-a-math-object — "spline" was a flexible naval-lofting batten pinned by weighted "ducks" for centuries before it named a polynomial. Tier 3, craft/etymology.
• claim-schoenberg-developed-splines-at-the-ballistic-research-lab-that-built-eniac — Schoenberg initiated spline theory as a WWII mathematician at BRL Aberdeen (MacTutor, Tier 2, verbatim), the same lab that commissioned and later ran ENIAC (US Army ARL page, Tier 2, verbatim). The strongest note of the three.

The three are densely cross-linked, and I wired the Schoenberg note into two existing clusters: the actuarial/mortality-table thread (claim-dodson-built-age-scaled-premiums-founding-equitable-life, via the "graduation" hook) and the vault's recurring myth-correction / miscredit pattern (claim-ann-hardy-tymshare-os-miscredited-to-husband).

Question routed

• question-verify-schoenberg-1946-graduation-actuarial-term (status: open) — the capture's one flagged sub-claim (Claim 2b): that Schoenberg's 1946 papers use "graduation" in its actuarial sense. That rests only on a Tier 3 design blog and is [unverified-quant/mechanism]. The question names the exact primary to read — Schoenberg's two-part 1946 Quarterly of Applied Mathematics paper — and the parent note keeps the flag preserved in-body and in audit_status.

What I skipped and why

• Claim 2b got no note of its own — too thin to stand alone on a flagged Tier 3 sentence. I folded it into the Schoenberg note as the misattribution being corrected, kept the flag, and pushed the verification to the question queue. That felt more honest than either a standalone flagged note or dropping it silently.
• The three further-leads (NURBS/Boeing 1979–81, Pixar subdivision surfaces 1997, Alois Auer the punchcutter, Knuth's METAFONT) are saved hooks, not claims. Left as leads in the capture, not promoted. All logged in the capture's not_promoted: list so a future reader sees they were judged, not missed.

What felt off about the capture

Honestly, very little — this was a clean one. Tiers were labelled accurately with no inflation: the design-history blogs are correctly Tier 3, Wikipedia correctly Tier 4, MacTutor and the Army history page correctly Tier 2, and the capture reserved verbatim quotes for exactly the Tier 2 claim that carries the surprising-biographical load (per sources.md, a surprising claim escalates the floor to Tier 1–2, and Claim 3 cleared it). It flagged its own weak spot — the "graduation" mechanism — rather than smuggling it in clean. Two small notes: (1) Claims 1 and 2 have no captured verbatim phrase, only paraphrase, so I marked both audit_status: flagged and wrote honest source_quote placeholders rather than inventing a quote; a primary pass on the Bézier priority story would be worth it before either leaves seedling. (2) The chain is a genuine cross-domain bridge with no vault collisions — a real orphan-cluster seed, not duplicated research.

Zero collisions with existing notes. No git run — the auto-commit agent will pick this up within 15 minutes.

Promotion — Stein's paradox (headless)

Second promotion of the day, run unattended. I promoted 10-inbox/raw/2026-07-09-hop-steins-paradox.md — a hop-batch capture chaining Robbins-Monro → empirical Bayes → Stein's paradox → the James-Stein baseball example → Charles Stein the person.

What I found → what I promoted

The capture packs two "core claims," but they bundle four distinct assertions. I winnowed to three notes:

• claim-james-stein-estimator-uniformly-dominates-the-sample-mean — the mathematical result: shrinkage uniformly beats the sample mean under squared-error loss, MLE inadmissible in dim ≥ 3. Tier 1 (Efron/Hastie CASI Ch. 7).
• claim-efron-baseball-shrinkage-halved-batting-average-prediction-error — the concrete quant claim: 18 players, 1970 season, "MLE .0425, JS .0218." Tier 1. Split from the theorem because it's a different claim with a different verification need (a specific number to pull vs a theorem to cite), and it's the culturally legible half worth standing on its own.
• claim-stein-delayed-admissibility-proof-five-years-to-avoid-bayesian-argument — the person: Stein withheld his own theorem five years until he had a non-Bayesian proof. Tier 2 (IMS obituary).

All three link into the existing statistical-history cluster via the hop seed claim-robbins-monro-1951-stochastic-approximation, and the Stein-the-person note wires into the "recognition delayed by discomfort" pattern via claim-lamport-paxos-greek-allegory-delayed-publication — a structural echo the capture itself named (Hop 4's max-cosine top match was literally the Paxos note).

Questions routed

• question-verify-efron-casi-james-stein-baseball-figures (open) — re-read the CASI PDF to confirm the domination statement and the .0425/.0218 figures.
• question-verify-stein-ims-obituary-bayesian-delay-quotes (open) — confirm the five-year-delay quote and, crucially, whether the "certain doctrine" line is Stein's own words or the obituarist's paraphrase.

What I skipped and why

• Stein's war-protest arrest — real but not an atomic claim worth its own permanent note. Folded into the person-note as one sentence of context, the way the capture itself framed it ("his politics and his mathematics both ran against consensus").
• Four further-leads — SVGD (2016), Robbins' optimal-stopping problem, the Robbins conjecture / EQP, and Lai-Robbins bandits. All are explicitly saved hooks, and the capture even flags SVGD's mechanism as "not verified to capture standard." Left as leads in the capture, all logged in not_promoted: so a future reader sees they were judged, not missed.

What felt off about the capture

Not much — it's a strong, honest capture. Tiers are labelled accurately (Efron's self-hosted chapter correctly Tier 1, IMS obituary correctly Tier 2), and it reserved verbatim quotes for exactly the load-bearing claims. Three things I want on the record:

1. Status calls. I'm headless with no independent re-read of the primaries, so all three notes are seedling / capture-verified, matching the Lamport precedent from 2026-07-09. Caveat worth a ruling from Cali: the blessed definition of capture-verified is "re-check blocked by access (paywall or equivalent)," but both sources here are freely accessible (a public PDF and an open web page) — access isn't the blocker, my staying inside promote.md's lane is. I used capture-verified for consistency with precedent rather than inventing a label, and routed both re-checks so the gap is queued, not hidden. If you'd rather these read flagged until a fetch happens, that's a one-field fix.
2. Attribution risk on the Stein quote. The interesting "certain doctrine" line is a surprising, load-bearing biographical claim, and sources.md says those should sit on the person's own words. It's currently a Tier-2 obituary quote and it's unclear whether Stein said it or the obituarist wrote it. That's the single thing most worth chasing — I made it explicit in the routed question.
3. No duplicated research. James-Stein, shrinkage, admissibility, and empirical Bayes were all genuinely absent from the vault — zero collisions across 342 claim-notes and 149 questions. This is a real orphan-cluster seed off the Robbins-Monro note, not a re-tread. If a fourth statistical-history note lands, it may be time for a moc-statistical-inference-history (Robbins-Monro, Amari-1968, and now three Stein notes would be the nucleus).

3 claims-worth-of-material found → 3 promoted → 1 sub-claim folded + 4 leads left behind, 2 questions created, 0 collisions. No git run — the auto-commit agent handles it within 15 minutes.

— Seek

— Seek

Promotion — hop capture: synthetic control → Mariel Boatlift (2026-07-11, headless)

Ran /promote headless on 10-inbox/raw/2026-07-09-hop-synthetic-control-mariel-boatlift.md, a Sonnet-5 hop-batch capture (no prior retrieve-before-write). Winnowed on my own judgment — no pause, no one to ask.

What I promoted (4 notes):

• claim-synthetic-control-built-to-measure-basque-terrorism-gdp-cost — the method's origin in Abadie & Gardeazabal's 2003 Basque-terrorism paper (~10pp GDP gap), Tier 1, verbatim quote. The clean cross-domain-bridge piece.
• claim-mariel-boatlift-wage-effect-flips-with-sample-construction — Card (no effect) vs Borjas (10–30% dropout drop) vs Peri & Yasenov's synthetic-control rebuttal. The central hop content.
• claim-mariel-wage-effect-may-be-1980-census-methodology-artifact — the Clemens & Hunt "it might all be a survey-sampling artifact" twist. The sharpest surprise in the capture.
• claim-card-2021-nobel-natural-experiments-labour-economics — Card's 2021 Nobel, the field-level anchor tying the cluster to the credibility revolution. Tier 1.

I linked all four densely into each other and back to the seed cluster — claim-soviet-emigre-career-premium-was-structural-not-ability — because the shape is identical: hold the event constant, vary the comparison-group construction, and the sign flips. That's the vault-worthy through-line, and it's the same "structural, not selection" template I flagged as transferable when I promoted the Soviet-émigré capture two days ago.

Questions routed (2):

• question-verify-mariel-boatlift-wage-effect-primary-sources — check the Card/Borjas/Peri-Yasenov numbers against the primary papers (they currently rest on Merler's Bruegel retelling, Tier 2).
• question-verify-clemens-hunt-mariel-census-artifact — verify the survey-artifact mechanism and Borjas's timing rebuttal against the primary Clemens & Hunt paper.

No existing questions covered either verification, so both are new.

What I skipped, and why:

• The Roy model self-selection material (hop 1) — under-sourced for a claim-note. The capture's provenance carries no Tier-1/2 source or quote for it (just a personal scholar page and Wikipedia), and it's background theory, not a load-bearing claim of this capture. Left in the inbox, logged in not_promoted.
• Peri & Yasenov's "15–20 obs/year" critique — folded into the sample-construction note rather than split into its own thread.
• Prop 99 tobacco — a further lead, not asserted here; mentioned in the origin note's commentary only.
• The two saved hooks (German expellees; the "cream-skimming" pun) — explicitly out of scope per the capture's own logging.

Retrieve-before-write: the semantic-index hint list was almost entirely off-target (Prussian blue, Toshiba, backprop, etc.) — none of those collide. The only real neighbors were the two Soviet-émigré notes, which I linked rather than duplicated. Zero collisions to dedup: nothing in 30-notes/ covered synthetic control, Mariel, Card, or Borjas. This is a genuine new orphan-cluster off the migration/labor-markets seed.

What felt off about the capture:

1. Tier discipline was honest but thin on the load-bearing bits. The two Mariel notes both rest on Merler (Bruegel, Tier 2) — a secondary retelling of the primary papers. Tier 2 technically clears the quant/mechanism floor, but the numbers are the contested heart of a famous fight, so I flagged both notes ([unverified-quant], [unverified-mechanism]), kept them seedling, and routed the primary checks. The capture itself carried no [unverified-*] flags at all, which understated this gap — a batch-worker habit worth watching.
2. No tier inflation. The two Tier-1 tags (AER paper, Nobel page) are legitimately Tier 1. If anything the capture was slightly generous calling the whole thing novel — the synthetic-control method is well-trodden ground, as the capture's own post-worthy: maybe admits.
3. audit_status honesty: I'm headless and did not independently re-open any source this session, so I did not mark anything verified-verbatim. The two Tier-1 notes read as capture-sourced-pending-recheck; the two Tier-2 notes read as flagged. If you'd rather the Tier-1 pair say capture-verified, that's a one-field fix.

4 claims-worth-of-material found → 4 promoted → 4 leads/hooks left behind, 2 questions created, 0 collisions. No git run — the auto-commit agent handles it within 15 minutes.

— Seek

Promotion (headless) — 2026-07-09-hop-texas-grid-phantom-load

Second promotion of the day, run unattended. I promoted the Texas-grid / "phantom load" hop capture (10-inbox/raw/2026-07-09-hop-texas-grid-phantom-load.md). Winnowed on my own judgment — no pause, since no one's here to sign off.

Notes created (3):

• claim-texas-grid-isolation-1935-federal-regulation-dodge — Texas kept its grid intrastate to dodge federal regulation under the 1935 PUHCA. Seedling, flagged [unverified-history] (single Tier-3 KUT/NPR source for a surprising, load-bearing reframe).
• claim-1976-midnight-connection-texas-oklahoma-forced-jurisdiction — the 1976 "Midnight Connection," where Central and South West Corp deliberately breached the isolation to force federal jurisdiction and the PUC cut it off within a year. Seedling, same flag/source.
• claim-ercot-large-load-queue-quadrupled-2025-phantom-load — ERCOT's large-load queue hit ~226 GW by Nov 2025 (≈4× the ~63 GW end-2024), ~77% data centers, unknown speculative "phantom load" share. Seedling, [unverified-quant] preserved.

I also added a "regulatory geography — the Texas grid island" subsection to moc-inference-economics, slotting these three under the existing physical-constraint/grid cluster so the thread is navigable.

Questions routed (2, both new):

• question-verify-ercot-large-load-queue-phantom-load-figures — pull the 226/63 GW and 77% figures from ERCOT's own large-load reports, not the trade-press reproduction. Deliberately kept distinct from question-verify-lbnl-queued-up-interconnection-figures, which is the national generation queue, not ERCOT's large-load (demand-side) queue.
• question-verify-texas-grid-isolation-puhca-1935-origin — corroborate the 1935 regulatory-arbitrage origin and the 1976 incident against a primary/scholarly source (covers both historical notes).

What I skipped, and why:

• The IEEE Spectrum hop-1 framing (millisecond step-changes, training-vs-inference grid signatures, IEA 3–4%-of-electricity projection) — entry-point framing, not developed as a core claim in the capture, thin on quote/tier here, and adjacent to ground already covered by claim-bit2watt-gpu-scheduling-destabilizes-power-grid. Left in inbox.
• The Tushman-Romanelli echo — the capture itself says it's "not asserted as fact here." A lead, not a claim.
• The options-market / speculation-bubble reading of the phantom-load gap — my own hypothesis in the capture's commentary, not a sourced claim. Logged as a future chain, not promoted.
• The cost-allocation fight and "Batch Zero" reform, and the "who flipped the switch" / CSW-Corp-aftermath leads — saved hooks and unresolved leads the capture never ran down. Folded the historical leads into the 1976 note's commentary and the routed history question.

Retrieve-before-write: the semantic-index hint pointed me at the right neighborhood. The one real collision risk was claim-us-interconnection-queue-exceeds-installed-capacity — but that's the national generation interconnection queue (LBNL, ~2,600 GW), whereas claim 3 is ERCOT's demand-side large-load queue (226 GW). Genuinely distinct axes; I linked them rather than duplicating. No existing note covered Texas grid isolation, PUHCA, the Midnight Connection, or ERCOT's large-load queue — this is a new sub-cluster hanging off the inference-economics grid thread.

What felt off about the capture:

1. Sourcing floor honesty is mixed. The capture flagged claim 3 correctly ([unverified-quant), but left the two historical claims (1 & 2) unflagged despite both resting on a single Tier-3 Austin-public-radio retelling. The "avoid federal regulation" origin is a mildly surprising, load-bearing reframe — under the sources.md historical-claim floor that should escalate toward Tier 1–2. So I added [unverified-history] flags myself and routed the primary check rather than inheriting a clean-looking historical claim on a soft source. A batch-worker habit worth watching: flagging the number but not the tidy origin story.
2. No real tier inflation. The capture's own tier labels ("Tier 3-4 historical, Tier 2-3 trade press") were honest, and its post-worthy: maybe already conceded the headline numbers need ERCOT before the piece could carry weight. I agree.
3. Latitude Media consistency. I assigned claim 3 source_tier: 2 to match the sister note claim-us-interconnection-queue-exceeds-installed-capacity (also Latitude Media), while keeping the [unverified-quant] flag — Tier 2 clears the quant floor on paper, but it's still secondary reproduction of ERCOT's data, so the flag and the routed ERCOT-primary check stand.
4. audit_status: headless, I did not independently re-open any source, so nothing is verified-verbatim. All three notes are flagged (pending primary), consistent with the routed questions.

3 core claims found → 3 promoted → 5 leads/hooks/framings left behind, 2 questions created, 0 collisions, 1 MOC updated. Capture marked status: promoted, not deleted. No git run — leaving that to the auto-commit agent.

— Seek

Promotion — 2026-07-09-hop-third-man-factor.md (headless)

I promoted the hop-batch capture on the angular gyrus's "other trick" — the one that runs OBE → robot-induced apparition → the literary origin of the "Third Man Factor." Three core claims, all three kept; nothing about it was duplicated in the vault.

Retrieve-before-write: clean. The only note that even mentions the angular gyrus is the seed, claim-insight-conflict-and-restructuring-distinct-networks, and there it's the phase-general in insight role — a genuinely different function from OBE, so I linked to it as a contrast rather than treating it as a collision. No existing note touches Blanke, out-of-body experiences, feeling-of-a-presence, Shackleton, Eliot, or the Third Man Factor. The semantic hint's top neighbors (pulsar scruff, yellow rain, Garfield) are thematic cousins — "surprising reversal" notes — but none cover these claims. This is a fresh cluster; the three new notes link tightly to each other (A↔B↔C) and out to the seed.

Notes created:

• claim-angular-gyrus-stimulation-induced-out-of-body-experience (Blanke et al. 2002, Nature, Tier 1)
• claim-robot-induced-feeling-of-a-presence-models-patient-hallucination (Blanke et al. 2014, Current Biology, Tier 1)
• claim-third-man-factor-name-descends-from-eliot-waste-land (Eliot's note + Shackleton, Tier 3 primary text with a Tier-4 etymology)

Questions routed (2):

• question-verify-blanke-2002-2014-primary-quotes — both Blanke quotes were read at capture time off the PubMed abstract records, not the paywalled full papers, so I marked those two notes audit_status: capture-verified and routed the verbatim full-text check.
• question-verify-third-man-factor-clinical-term-eliot-etymology — the load-bearing etymology (that the term "Third Man Factor" descends specifically from the poem) rests on Wikipedia-tier sourcing; under the sources.md floor a surprising, load-bearing historical claim should escalate, so I flagged it in-note and routed the primary check (Shackleton's South wording; Geiger's 2009 book as the naming node).

All three notes stay seedling — two because the primary quotes aren't independently re-verified, the third because its etymology is flagged. No MOC: three notes don't clear the 5-note threshold, and there's no adjacent embodiment cluster yet to index.

Skipped (left in not_promoted:): the rubber-hand illusion, the Parkinson's-FoP-biomarker thread, Geiger's survival-accounts book, the Road-to-Emmaus biblical layer, and NDE/AWARE research. Every one of these is a saved hook in the capture rather than a sourced claim — the capture author was explicit that they were deferred or judged well-trodden. I folded the Emmaus aside into the Third Man note as a one-line pointer and folded Geiger into the etymology question as the coinage node to check, rather than minting thin notes for either.

What felt off: honestly, not much — this was a well-built capture. Two small things worth logging. (1) The tier labels were honest but slightly generous on wording: claim 1's grounding quote reads like the paper's interpretive sentence, and I could only confirm it at the abstract level, so I declined verified-verbatim and routed the check rather than inheriting a clean-looking Tier-1 stamp. (2) The capture asserted the Third Man etymology as "uncontested per sources.md floor," but the derivation of the term (as opposed to Eliot's verbatim footnote) is exactly the kind of surprising, load-bearing historical claim the floor says to escalate — a batch-worker habit I also flagged on this morning's promotion: quoting the primary text faithfully but waving through the tidy origin story wrapped around it. I added the flag and the question myself rather than letting it ride.

3 core claims found → 3 promoted → 5 hooks/leads left behind, 2 questions created, 0 collisions, 0 MOCs. Capture marked status: promoted, not deleted. No git run — the auto-commit agent has it.

Promotion — van Meegeren / science catches forgers (headless)

Promoted 10-inbox/raw/2026-07-09-hop-van-meegeren-science-catches-forgers.md, a hop-batch capture that runs from a Royal Society memoir of the physicist Martin Aitken out to two independent hard sciences becoming art's forgery police a generation apart. Ran the full /promote workflow unattended — no pause, no git, report lands here instead of to Cali.

Three core claims, all three promoted:

• claim-van-meegeren-forged-a-vermeer-under-court-supervision-to-escape-a-treason-charge — the WWII treason-to-forgery inversion.
• claim-coremans-bakelite-in-paint-proved-van-meegeren-vermeers-were-forgeries — chemistry (20th-c. synthetic resins) as the actual proof.
• claim-thermoluminescence-dating-became-a-commercial-antiquities-forgery-detector — physics as the second, independent forgery detector.

Every one stays seedling: claim 1 is a load-bearing biographical claim on Tier-4 Wikipedia, claim 2 carries the capture's [unverified-mechanism] flag, claim 3 the [unverified-quant] flag. Flags preserved in body and frontmatter; none promoted past seedling.

Questions routed (3, all new — no existing 50-questions/ covered these):

• question-verify-van-meegeren-trial-forgery-under-supervision-primary
• question-verify-coremans-1949-bakelite-resin-mechanism-primary
• question-verify-oxford-authentication-tl-throughput-and-fraud-quant

Retrieve-before-write: no collisions. There is no van Meegeren, Coremans, thermoluminescence, or Etruscan note in the vault. What exists is a thematic forgery-detection cluster I wired all three notes into — claim-mabillon-1681-founded-diplomatics-to-refute-forgery-charge (founding of document authentication), claim-hirsch-forensics-drove-dias-superconductivity-retraction (data-pattern forensics — itself promoted this morning), and the correction-propagation note claim-yellow-rain-bee-feces-refutation-never-formally-retracted as neighbor. This is now a real cluster: five-plus notes on "how a fake/false result gets caught" — worth watching for an MOC next lint.

What I skipped and why: the Etruscan Terracotta Warriors / Fioravanti confession, the Doreen Stoneham living-successor lead, and the archaeomagnetic / OSL dating leads are all saved hooks in the capture, not Core claims — the author flagged them "saved not followed," and the Etruscan detail rests on an un-URL'd web search. Left them in not_promoted: as judged skips rather than minting thin notes. I also declined to mint the capture's headline thesis — "two unrelated sciences became forgery police forty years apart, without citing each other" — as a standalone note. It's the real intellectual payload, but it's my synthesis, the "without citing each other" part is unsourced, and three notes is below the MOC threshold. Folded it into the note bodies and commentary instead.

What felt off: two things, both sourcing. (1) Tier inflation — source 2 is grokipedia.com, an AI-generated encyclopedia the capture labeled Tier 4. I'd grade it Tier 5, and it's carrying a mechanism claim (which resin, used how), which sources.md floors at Tier 1–2. The real anchor is Coremans's own 1949 monograph, which no one in the chain read — so the note leans on a title, not a text. (2) The same batch-worker habit I flagged this morning: the primary-ish grounding quotes are honest, but the tidy numbers ride on soft venues — three separate statistics (throughput, ceramics share, Hong Kong copy rate) all sourced to one Tier-3 Smithsonian piece, and the "75 percent ... are said to be copies" is itself hedged in the source. I kept the quant flag and routed rather than inheriting clean-looking figures. Nothing looked fabricated; the capture was transparent about its own gaps. This capture is not duplicated research — distinct subject from the morning's Hirsch promotion, though it lands in the same forgery-detection neighborhood, which is a good sign the vault is compounding.

3 core claims found → 3 promoted → 5 hooks/leads left behind, 3 questions created (0 linked), 0 collisions, 0 MOCs (cluster now MOC-eligible — noted for lint). Capture marked status: promoted, not deleted. No git run — auto-commit has it.

— Seek

— Seek

Promotion — Vesuvius/Herculaneum scroll capture (headless, 2026-07-11)

Promoted 10-inbox/raw/2026-07-09-hop-vesuvius-scroll-ai.md, a hop-batch capture from claude-sonnet-5 running a manuscript-transmission → text-recovery chain. I ran it unattended: no pause, my judgment on the winnow.

What I created (3 claim-notes, all seedling / capture-verified):

• claim-pherc-1667-first-herculaneum-scroll-read-unopened — the 25 June 2026 landmark: first complete carbonized Herculaneum scroll read end-to-end via X-ray tomography + ML ink detection, without unrolling. Tier 1 (the prize's own page).
• claim-virtual-unwrapping-first-proved-on-2015-en-gedi-scroll — Seales's ~20-year program, En-Gedi 2015 as first proof, 2021 $14M NSF grant → EduceLab folded in. Tier 2 (UKNow).
• claim-sinai-palimpsests-multispectral-imaging-recovered-erased-texts — the 2011–2016 Sinai project, 74 palimpsests / 6,800 pages / 305 erased texts. Tier 2 (EMEL). Cross-linked into the manuscript-transmission notes (Kalila wa-Dimna, Borzuya).

Questions routed (2, both new — no existing question covered them):

• question-verify-vesuvius-pherc-1667-first-scroll-read — corroborate the "first complete scroll" landmark, the Stoic-text ID, En-Gedi, and the NSF grant beyond scrollprize.org's own venue.
• question-verify-sinai-palimpsests-imaging-figures — second-source the 74/6,800/305 counts.

What I skipped and why: the "20-year program" narrative and the Stoic treatise's philosophical content were folded as supporting detail rather than split into their own notes (avoiding over-splitting a single-source claim). The capture's three "further leads" — weight-in-gold Baghdad translators, Nat Friedman/Daniel Gross as funders, and the Kalila wa-Dimna lineage — are saved research hooks, not claims in the body; I left them in the inbox record (Kalila wa-Dimna is already a vault note, so I linked rather than duplicated). I did not mint the capture's headline "three eras of invisible-text recovery" bridge as a synthesis note: it's Seek's interpretive framing, and three notes is below the MOC threshold — I carried the lineage through dense cross-links instead.

What felt off: nothing fabricated, and the capture was honest about its hop reasoning. But note two things. (1) The load-bearing "first" rests on a single, partly promotional venue (the prize's own page) — Tier 1 by rubric, but a self-account; I refused to let its recency make it feel settled, hence seedling + a corroboration question. The capture itself carried no [unverified-*] flags despite that gap, which is arguably tier-inflation-adjacent — a days-old extraordinary "first" from a single self-interested source should probably have been flagged at capture time. (2) The quantitative figures (1.4m/22 columns; 74/6,800/305; $14M) all sit on single Tier 1–2 institutional pages; acceptable per the sourcing floor, but each wants one convergent source before going evergreen. Retrieve-before-write found zero collisions in 30-notes/ — this is a genuinely new cluster for the vault (and MOC-eligible once it grows past a couple more text-recovery notes; noted for the next lint).

3 core claims found → 3 promoted → 0 skipped-outright (2 folded, 4 leads parked), 2 questions created (0 linked), 0 collisions. Capture marked status: promoted, not deleted. No git run — leaving the commit to the auto-commit agent.

— Seek

Promotion — 2026-07-09-hop-vicon-motion-capture-bridge (headless)

I promoted the sports-bra → breast-biomechanics → Vicon → NASA hop capture. A tidy five-hop chain with an honest verification hop at the end (Hop 5 went to NASA's own site precisely so a vendor page wouldn't carry the marquee cross-domain claim alone). I liked that discipline; it made the winnowing easy.

3 core claims found → 3 promoted → 0 skipped-outright (2 folded, 3 leads parked).

Notes created:

• claim-first-sports-bra-1977-sewn-from-two-jockstraps — the 1977 Jogbra prototype (two jockstraps, the upside-down-joke origin). Smithsonian, Tier 2, verbatim quote.
• claim-breast-displacement-figure-eight-motion-capture-2009 — the Scurr/White/Hedger figure-of-eight finding, 15 D-cup participants, reflective markers on nipple + trunk, 5/10 kph. PubMed, Tier 1.
• claim-vicon-motion-capture-lineage-gait-lab-to-nasa — the instrument bridge: clinical gait analysis → film VFX → NASA's AIBEL lab. I led the frontmatter with the Tier-1 NASA equipment listing (the load-bearing endpoint) and folded the Tier-2 Vicon vendor self-history into the body, clearly labelled as marketing. That inversion mattered: the sports-to-orbit span is only worth a note if the orbit end is sourced to NASA, not to Vicon.

Question routed:

• question-verify-breast-acceleration-exceeds-formula-1 — the capture's own [unverified-quant] flag. The "breast acceleration exceeds an F1 car" line is the most quotable number in this whole space and traces only to press coverage; I named exactly what a primary read would need (peak accel in g / m·s⁻², the F1 comparison basis, whether the comparison is the researchers' or a journalist's).

What I skipped and why. The F1 comparison → routed, not promoted (never a core claim, sub-floor sourcing). Title IX as the Jogbra's commercial tailwind → a saved lead on general-search-context sourcing; I folded the uncontested facts (1979 patent, 1990 Playtex acquisition) into the sports-bra note but did not mint the causal bridge as its own note. Vicon's 1980s multi-site clinical adopters → a saved "institution behind the thing" thread resting on the same vendor page; mentioned in passing, not split out. All three are logged in the capture's not_promoted: list so they read as judged, not missed.

Retrieve-before-write. One genuine collision: claim-gilbreth-chronocyclegraph-motion-capture-ancestor (the 1910s pre-electronic ancestor). Distinct angle, not a duplicate — the Gilbreth note is the ancestor, mine is the modern optical-marker instrument genealogy — so I wikilinked both ways (added a backlink into the Gilbreth note, which had literally complained in its own commentary about hanging off a single vault node). Also linked the Vicon note into the inertia-1 wearable-motion cluster. No collisions on sports bras or breast biomechanics — those are fresh orphans, seeds for a possible future "measuring the moving body" MOC if a couple more notes land near them.

What felt off. Two things, both mild. (1) Tier-inflation-adjacent, self-caught by the capture: claim 3's origin/lineage narrative and client lists (ILM, EA, Ubisoft, Toyota) all sit on Vicon's own 30th-anniversary marketing page, tagged Tier 2. The capture was appropriately careful — it explicitly went to NASA for independent corroboration and labelled the vendor page "vendor primary" — so this is good practice, not a failure; I just made sure the note inherits that caution and doesn't state the client list as independent fact. (2) Duplicated-research signal across the batch: this is the second motion-capture-history capture from the same 2026-07-09 hop-batch (the other became the Gilbreth note). Two hops in one batch converging on "history of measuring the moving body" is worth watching — either a rich seam or the hop-walker circling. The cross-links turn the overlap into a cluster rather than a redundancy, so I'll treat it as the former for now. Nothing was fabricated; sourcing floors were respected; I did not re-fetch the web (consistent with today's earlier headless promotions), so all three notes are capture-verified, not verified-verbatim, and stay seedling.

3 claims → 3 promoted → 0 skipped-outright, 1 question created (0 linked), 1 collision (linked, not duplicated). Capture marked status: promoted, not deleted. No git run — leaving the commit to the auto-commit agent.

— Seek

Promotion — 2026-07-09-hop-victor-rothschild-fifth-man (headless, unattended)

Promoted the Victor Rothschild capture from 10-inbox/raw/. It's the hop-batch sibling of the Lighthill note I promoted on the 9th: the walk started from "who commissioned Lighthill" (the UK Science Research Council) and hopped sideways into 1971's other science-funding shock, then onto the man who wrote it, then into Cold War espionage. A genuinely good cross-domain chain — one funding institution away from a cluster I already hold.

What I created (3 notes):

• claim-rothschild-1971-report-mandated-customer-contractor-principle — the customer-contractor principle, Tier-2 quote from Miles Parker's 2016 Palgrave Communications account. Linked straight into the Lighthill / SRC cluster, which is the whole reason this hop is vault-worthy: two defining UK science-policy documents, one funding family, two years apart.
• claim-thatcher-1986-told-parliament-no-evidence-rothschild-soviet-spy — the Fifth Man suspicion and Thatcher's carefully-worded parliamentary clearance. Written narrowly, about the documented statement and the fact of the dispute, asserting nothing about whether he was actually a spy.
• observation-victor-rothschild-anchors-uk-science-funding-and-espionage-historiography — the actual hop insight: one man, three registers (science-funding reform, the Heath Think Tank, espionage historiography), none of them talking to the others. I wrote this as an identity bridge, the structural twin of the Richard Price observation I made earlier in this batch. It absorbs the CPRS and the bomb-disposal/zoologist biography so I didn't have to spawn thin standalone notes for them.

Questions routed (2):

• question-verify-rothschild-1971-customer-contractor-quote-primary
• question-verify-thatcher-1986-rothschild-hansard-and-fifth-man-dispute

What I skipped and why. No standalone CPRS note (folded into the observation — sourced only on Wikipedia search summaries, too thin to stand alone). No biography note for the George Medal / zoology (color, folded). No separate Perry-1994/Modin note (folded into the Fifth Man note as dispute context). The three "further leads" (Cambridge Apostles, SERC→EPSRC/PPARC/BBSRC, the "science broker" alternative) I left as leads. All logged in the capture's not_promoted: so they read as judged, not missed.

Retrieve-before-write. No existing note touches Rothschild, customer-contractor, CPRS, or the Cambridge Five — this is an orphan cluster seeding into the vault. The one real collision point is the Lighthill/SRC cluster (claim-lighthill-1973-blamed-combinatorial-explosion), which I linked from both the report note and the observation rather than restating. I also linked the espionage note into the Perceptrons myth-ledger by shape, not by subject.

What felt off. Two things worth flagging to Cali. (1) Sourcing smell: the load-bearing customer-contractor quote reached the capture through a PDF mirror on fbcoverup.com, a conspiracy-oriented host, not the canonical Palgrave Communications venue. The underlying Tier-2 paper is legitimate, but the copy used is not its home — a mirror can silently alter text. I kept the note at seedling / capture-verified and routed the exact-quote re-check rather than laundering the provenance clean. (2) Tier honesty, self-caught by the capture: the espionage material is Tier-3 Wikipedia on a contested, surprising, load-bearing claim — below the floor — but the capture flagged the dispute itself, so I inherited that caution, wrote narrowly, and routed the Hansard + Perry/Modin verification. No tier inflation that the capture tried to sneak past me. This is also the second cross-domain "one man, two/three lineages" identity-bridge I've promoted from the same 2026-07-09 hop-batch (Richard Price was the first) — the hop-walker has clearly found the identity-bridge groove; worth watching whether that's a rich seam or a tic.

3 claims → 3 promoted → 0 skipped-outright, 2 questions created (0 linked), 0 existing-note collisions (1 cluster linked into, not duplicated). Capture marked status: promoted, not deleted. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — Waddington's landscape as a Hopfield energy surface (2026-07-09 hop capture)

Headless, unattended run. I promoted 10-inbox/raw/2026-07-09-hop-waddington-hopfield-landscape.md — the hop that walks Mary Lyon → Waddington → the 2016 Hopfield-landscape paper → the 2024 Physics Nobel. Winnowed on my own judgment (no pause; no one to okay it) and kept a conservative line.

Notes created (2):

• claim-hopfield-network-formalizes-waddington-epigenetic-landscape — folds the capture's core claims 1 and 2. They read as one atomic thing to me: the landscape-as-energy-surface formalization and "cell types are attractors" are the same claim seen from two angles (attractors are the minima of the surface), so the 12-dataset / 50%-perturbation robustness became supporting detail inside the one note rather than a second note. Tier 1, npj Systems Biology.
• claim-hopfield-hinton-2024-nobel-physics-neural-networks — core claim 3, kept separate because the Nobel fact stands on its own and anchors two other vault clusters (Amari priority, Hinton's biological-implausibility arc). Tier 1 Nobel press release.

Both linked densely into the existing Hopfield/Amari and Hinton clusters. Both status: seedling / audit_status: capture-verified — the Tier-1 quotes come from the bee's reading at capture time; I did not independently re-fetch the sources this run, so I didn't launder them up to verified-verbatim.

Questions routed (3):

• question-amari-1972-applied-to-developmental-biology-before-2016 — the capture's "not checked" lead; if Amari 1972 reached dev-bio before 2016 the uncited-prior pattern extends into a second field.
• question-waddington-cybernetics-influence-primary-source — routes the capture's own flagged weakness.
• question-siggia-2025-reconstructs-waddington-landscape-single-cell — the unread 2025 PNAS confirmation-check lead.

Skipped (and why): the Waddington→Wiener cybernetics link and the 1965–68 symposia (Tier 3–4 biography, below the floor for a load-bearing cross-domain claim — routed, not written); Mary Lyon in Waddington's Edinburgh group (Tier 4 connective tissue, no standalone claim); the Siggia 2025 paper and the Amari-in-dev-bio question (unverified leads → questions). All recorded in the capture's not_promoted: list so they read as judged, not missed.

What felt off / worth Cali's eye: honestly, not much — this is the best-behaved capture I've promoted from the 2026-07-09 hop-batch. Sourcing is clean (the two promoted claims sit squarely on Tier 1), and the capture self-flagged its own weak spot in the post-worthy: maybe line: the connective tissue rests on Tier 3–4 biography and wants a primary Waddington source. No tier inflation trying to sneak past me; the flag matched my own read. One duplication note: this hop deliberately leans on the vault's existing Amari/Hopfield thread (claim-amari-1972-associative-memory-precedes-hopfield) — that's reuse-by-design, not accidental re-research, and I linked into the cluster rather than restating it. Retrieve-before-write turned up zero collisions: no prior note on Waddington, the epigenetic landscape, or the 2024 Physics Nobel existed. Genuinely new ground for the vault, cleanly attached to old ground.

3 core claims → 2 promoted (one note folds two) → 4 side-claims/leads skipped, 3 questions created (0 linked, none existed), 0 collisions. Capture marked status: promoted, not deleted. No git run — leaving commit and push to the auto-commit agent.

— Seek

Promotion — 2026-07-09-hop-waldo-scifi-to-muppets (headless)

Promoted the "waldo" capture: the coinage that hops from a 1942 Heinlein pulp story through 1940s nuclear teleoperators to a 1988 Muppet CG-puppet rig, entered via the sports-bra co-inventor moonlighting as a Henson costumer. Six hops, one Tier-1 source at the center (Henson's own journal). Fun chain. Winnowed on my own judgment — no pause, no one to ask.

Notes created (3):

• claim-waldo-coinage-heinlein-1942-to-nuclear-teleoperators — Claim 1, the etymology/lineage. Core rests on a directly read Tier-4 Wikipedia quote (fine for an uncontested etymology); the surprising sub-details (Goertz's Argonne master–slave manipulator, The Character Shop's "Waldo" trademark) came from cyberneticzoo.com and were never quote-fetched, so I flagged them and kept the note seedling.
• claim-henson-1988-waldo-realtime-cg-puppet-control — Claim 2, the Tier-1 anchor. Henson's journal ("Working w/Kirk Thatcher on computer generated character – Later Waldo") proves the name, year, and collaborator — and nothing about the machine. The 8-DOF rig / PDI / live-composite / HDPS-Academy-Award mechanism details all trace to a muppet.fandom page that returned HTTP 402 and was never fetched. Flagged [unverified-mechanism], seedling despite the Tier-1 quote. I refused to let the primary quote halo the mechanism sitting next to it.
• claim-polly-smith-bridges-sports-apparel-and-muppet-costuming — Claim 3, written as a distinct angle, not a duplicate (see below).

Questions routed (2, both new — none existed):

• question-verify-waldo-teleoperator-lineage-goertz-character-shop — Argonne reports + USPTO TESS for the two unfetched historical specifics.
• question-verify-henson-waldo-c-graphic-mechanism — muppet.fandom refetch, Kirk Thatcher / PDI accounts, and the Academy Sci-Tech award record.

What I skipped and why:

• The sports-bra invention claim itself — already in the vault as claim-first-sports-bra-1977-sewn-from-two-jockstraps, promoted only two days ago from a different 2026-07-09 hop capture (hop-vicon-motion-capture-bridge). Rather than re-write it I promoted only Polly Smith's genuinely new content — the decades-long Henson costuming career — as the cross-domain bridge, wikilinked to the existing note and to claim-breast-displacement-figure-eight-motion-capture-2009.
• The "further leads" — Disney's Wathel Rogers "Animating Apparatus" patent (1963), Harvey Chapman's "Garco" (1953), GE's "Yes Man" (1956), and Heinlein's claimed 1918 Popular Mechanics inspiration — all unverified against primaries and, in the capture's own words, "not chased." Left as inbox leads, logged in not_promoted:.
• Seek-commentary's "attribution drift as a cross-domain pattern" observation — folded into the etymology note's commentary callout rather than split out. Might become a reflection note if it recurs.

What felt off / worth Cali's eye:

• Duplicated research across the hop-batch. This is the second 2026-07-09 hop capture to land on the Jogbra/Polly Smith story. Two independent bees converged on the same 1977 invention from different directions (motion-capture biomechanics vs. Henson animatronics). Not wrong — it's a genuinely well-connected node — but the batch has no cross-capture dedup, so the same fact got researched twice and I had to catch the collision at promotion. Worth a flag for whoever tunes the hop protocol.
• Tier discipline was honest, not inflated — but thin. The capture's source_tiers were accurate ("Tier 3-4... though those weren't quote-fetched directly"), which I appreciate; it told me exactly where the soft spots were. Still, two of three claims lean on sources the bee couldn't actually fetch (402 and 403 blocks on muppet.fandom and the Smithsonian landing). The one real anchor (Henson's Red Book) only covers the naming, not the mechanism everyone actually cares about. So this reads "well-corroborated" (the capture's own post-worthy: yes) but is really one Tier-1 sentence plus a lot of Tier-3/4 connective tissue. I did not promote it as clean.
• Net: 3 core claims → 3 promoted (1 as a distinct-angle bridge, not a dup) → ~5 leads/side-claims skipped, 2 questions created, 1 collision caught with an existing note. Capture marked status: promoted, not deleted. No git run — leaving commit and push to the auto-commit agent.

— Seek

Promotion — 2026-07-09-hop-warhead-uranium-civilian-bridge (headless)

I promoted the capture 10-inbox/raw/2026-07-09-hop-warhead-uranium-civilian-bridge.md — a hop chain that ran CTBTO radioxenon monitoring → Chalk River Mo-99 shortage → NNSA Mo-99 HEU→LEU conversion → Megatons to Megawatts. Ran unattended, no pause, my own winnowing.

Notes created (3):

• claim-ctbto-radioxenon-monitoring-confounded-by-medical-isotope-production.md — Tier 1 (CTBTO). The genuinely new find: the test-ban network's xenon-133 detector is confounded ~18,000 times/year by hospital Mo-99 production, which "look[s] similar to those of a nuclear explosion."
• claim-nnsa-converted-global-mo99-production-heu-to-leu-targets.md — Tier 1 (DOE/NNSA). "All major global Mo-99 production facilities now use LEU targets" after the 2012 Nuclear Security Summit commitment.
• claim-two-verification-regimes-sensor-vs-inspector-govern-the-same-fissile-material-lifecycle.md — observation, seedling. The capture's actual cross-domain payload: remote radionuclide sensing (CTBTO) vs. reciprocal on-site inspection (Megatons to Megawatts) as two poles of fissile-material verification, tied into the Burevestnik/MIT-OSINT cluster as a third pole. Flagged [unverified] on the reciprocal-monitor facility roster.

Question routed (1): question-verify-megatons-to-megawatts-reciprocal-onsite-monitoring-primary.md — the Portsmouth/Paducah ↔ Seversk/Ozersk/Novouralsk/Zelenogorsk monitor stationing is load-bearing for the observation note but carries no verbatim primary in the capture. That flag keeps note 3 at seedling.

Skipped, and why:

• Megatons to Megawatts tonnage / one-third-US-fuel figure — collision: already in the vault as claim-megatons-to-megawatts-475t-heu-one-third-us-reactor-fuel.md, promoted from a different earlier capture (2026-07-09-hop-megatons-to-megawatts.md). I linked to it rather than duplicating, and promoted only the distinct verification-method angle.
• Chalk River NRU 2009 shutdown / Mo-99 shortage numbers (30-40% of world supply, ~76,000 patients/day, 80+ countries) — quantitative claims resting only on a Tier-3 trade publication (Diagnostic Imaging), below the Tier 1-2 floor for numbers, and tangential to the verification thesis. Left in inbox as a judged skip.
• Parked leads (NorthStar non-fission Mo-99; Russia's HEU-stockpile disincentive to convert; Clair Patterson orphan hook) — the capture already saved these as future hooks, not claims.

What felt off:

• Duplicated research across the batch. This is the second 2026-07-09 hop capture to terminate on Megatons to Megawatts. The two chains reached the same endpoint from different seeds, and the batch worker has no cross-capture dedup — so the marquee finding was already fully mined (3 notes) before I got here. I caught the collision at retrieve-before-write, but the hop protocol is spending research budget converging on the same node. Worth flagging to whoever tunes it.
• Tier honesty, but a thin real core. The capture stamped everything source_tier 1 and marketed itself post-worthy: yes — all Tier 1. That's almost fair for the three headline quotes (CTBTO, NNSA, EIA are genuinely Tier 1), but the chain's connective tissue — the Chalk River numbers and especially the reciprocal-monitoring facility roster — is Tier 3 or unquoted, and the frontmatter's blanket "all Tier 1" papers over that. Mild tier inflation by omission, not by lying.
• Net: 4 distinct claims found → 3 promoted (one as a distinct-angle bridge, not a dup) → ~5 leads/side-claims skipped, 1 question created, 1 collision caught with an existing note. Capture marked status: promoted, not deleted. No git run — leaving commit/push to the auto-commit agent.

— Seek

Promotion — 2026-07-09-hop-watermark-paper-to-llm (headless, unattended)

Promoted the watermark cross-domain bridge capture. A clean, genuinely fun one: the same word watermark, and the same abstract move — a covert-but-verifiable marker of origin baked into the substance of an artifact — running from a 13th-century Fabriano papermaker's mark, through Bank of England banknote security, to a statistical signature in LLM token sampling. 750 years, three media, one function.

Winnowed 3 core claims (+ the bridge thesis) → 4 notes:

• claim-watermarks-originated-as-papermakers-mark-in-fabriano-1282.md — seedling. Tier 4 Wikipedia; flagged because "1282" is a specific date on a soft source (and the same capture's mill-founding dates 1264/1276 don't quite line up with it).
• claim-bank-of-england-adopted-watermarked-banknote-paper-1697.md — seedling. Tier 3 American Banker; the 1694/1695/1697 dates + Sutton Courtenay mill are quantitative and flagged [unverified-quant — needs primary].
• claim-kirchenbauer-2023-green-red-list-llm-watermark.md — seedling, but the strong one: Tier 1 arXiv primary, verbatim abstract quote, clears the mechanism floor. Kept it to the founding green/red-list mechanism; left the robustness quantitatives out.
• observation-watermark-same-provenance-mechanism-paper-to-llm.md — the bridge itself, as an observation note joining the print-history/authentication cluster to the AI/ML cluster.

Questions routed (2): question-verify-fabriano-1282-watermark-origin-primary.md and question-verify-bank-of-england-1697-watermark-adoption-primary.md. Both keep their claim-notes at seedling until a primary or scholarly source firms the dates. No question routed for the Kirchenbauer note — a clean Tier 1 primary with a verbatim quote doesn't need one; routing a verification for it would just be noise.

Retrieve-before-write: no collisions. Grepping 30-notes for watermark / Fabriano / Kirchenbauer / Bank of England came up empty — this is fresh territory for the vault. I did wire the new notes densely into existing neighbours: the seed claim-bomberg-1520s-venice-talmud-set-standard-page-layout, claim-early-modern-printers-corrected-print-with-cancel-slips, claim-mabillon-1681-founded-diplomatics-to-refute-forgery-charge on the print/authentication side, and claim-djvu-1998-paper-coauthored-by-lecun-and-bengio + moc-backpropagation-origins on the AI side.

Skipped, and why:

• Paraphrasing collapsing detection below ~0.3 (hop 5) — confirmatory hop, a quant on a survey source; wants its own primary-sourced note.
• 1790s loyalist counterfeiting of French assignat watermarks — folded as an arms-race sentence into the Bank of England note; left as its own incident otherwise.
• Aldus Manutius, rag→wood-pulp, Hebrew/Latin punch-cutting — saved hooks, not claims; the capture itself deferred them, and I respected that (the rag-shortage one even carries the original bee's explicit "saved rather than folded in" note).

What felt off:

• Mild under-flagging, not inflation. Unlike some batch captures, this one was honest about tiers — the frontmatter says "4 / 3 / 1 (mixed)" rather than stamping everything Tier 1. But it called the Tier-3 Bank of England claim "acceptable per sources.md" without noticing that the specific dates inside it are quantitative and want Tier 1–2. Same soft spot on the Fabriano 1282 date. The bee cleared the historical floor but skated past the quantitative one; I added the flags it should have carried.
• The bridge is a loanword, not a reinvention. The capture framed the recurrence as a mechanism "reinvented" across 750 years. It isn't — the LLM field consciously borrowed the word "watermark" by analogy. I softened the observation note accordingly ("carried over," not "reinvented") so the vault doesn't record a nominal coincidence as evidence of hidden lineage. The actual story is that the metaphor happens to be exact.
• No duplicated research this time — unlike the Megatons capture earlier today, this chain landed somewhere the vault hadn't been.

Net: 3 core claims (+1 synthesis) → 4 notes → 2 questions routed → 0 collisions. Capture marked status: promoted, left in place as the audit trail. No git run — the auto-commit agent has it.

— Seek

Promotion — hop-weapons-labs-to-frozen-snow (2026-07-11, headless)

I promoted 10-inbox/raw/2026-07-09-hop-weapons-labs-to-frozen-snow.md — the hop chain that runs from Microcosmos's Immersed Boundary Method out to the numerical lineage behind Disney's Frozen snow. Winnowed on my own judgment (no pause — unattended run).

Promoted 3 of the capture's core claims, one note each:

• claim-pic-method-born-at-los-alamos-for-hydrodynamics — Harlow's particle-in-cell method, Los Alamos, compressible-fluid hydrodynamics. Tier 1 (OSTI report record).
• claim-material-point-method-1993-sandia-funded-penetration-mechanics — MPM out of Sandia-funded penetration mechanics, 1993. Tier 4 (Wikipedia), [unverified-mechanism] flag preserved, kept seedling, audit_status flagged. This is the load-bearing middle link and the only aggregator-sourced one.
• claim-frozen-snow-ran-on-material-point-method-matterhorn — Stomakhin et al.'s SIGGRAPH 2013 MPM snow solver ("Matterhorn"). Tier 1 (Disney's own publication page).

All three linked densely into each other and back to the seed note claim-microcosmos-elastic-filaments-differentiable-gpu-fluid; note 2 also links the defense-research-into-civilian-use thread via claim-mit-1973-divested-instrumentation-lab-as-draper-under-antiwar-protest.

Routed 1 question: question-verify-mpm-1993-sandia-penetration-primary — read the 1994 Sulsky–Chen–Schreyer CMAME paper directly and confirm (a) penetration motivation and (b) the Sandia funding acknowledgement. No existing question covered this exact verification.

Skipped (left in inbox, logged in not_promoted): Peskin's 1972 IBM heart-valve origin and his later reuse of IBM (inner ear / clotting / molecular motors); Z. Jane Wang's fruit-fly flight-sensing lead (the 2026 PNAS ~4ms bound). All three are off the capture's actual thesis (weapons-labs → Frozen), sit in the hop-chain/further-leads sections rather than "Core claims," and are only Tier-4/lightly sourced here. I also declined to spin the PIC→FLIP→MPM→Matterhorn lineage into its own synthesis note — folded it into the three via wikilinks; a 3-note cluster is below the ~5 threshold for an MOC.

Retrieve-before-write: 0 collisions. Nothing in 30-notes/ touched PIC / MPM / Harlow / Sandia / Frozen / Matterhorn / Stomakhin / penetration; the only related note was the seed, which makes a different claim (and only mentions IBM in passing).

What felt off about the capture:

• Tier honesty is uneven. The capture is admirably explicit that the middle link is soft — it flagged the Sandia/penetration mechanism itself. Good. But the "post-worthy: yes … grounded in primary sources at both ends" line slightly oversells: the two ends are Tier 1, the bridge is Tier 4, and the bridge is the whole point. I made that explicit in every note's commentary so a future me doesn't publish on the strength of the endpoints.
• The Harlow year is fuzzy. Capture dates PIC to 1962 via OSTI 4769185; the canonical Harlow PIC papers are often cited ~1955–1964. I recorded 1962 as the capture's source states but could not independently re-fetch OSTI in headless mode (no web tool) — the audit_status says so.
• No duplicated research — this chain landed somewhere the vault hadn't been.

Net: 5 distinct claims found → 3 promoted → 2 skipped, 1 question routed, 0 collisions. Capture marked status: promoted, left in place as the audit trail. Every new note stamped writer_model: claude-opus-4-8. No git run — the auto-commit agent has it.

— Seek

Promotion — 2026-07-09-hop-zeiss-euv-instrument-monopoly

I promoted the capture 10-inbox/raw/2026-07-09-hop-zeiss-euv-instrument-monopoly.md — the Abbe/Zeiss arc running from a 19th-century optics workshop through a labor-governance experiment to the EUV chokepoint under today's AI-GPU supply chain.

Promoted 3 notes (all seedling, writer_model: claude-opus-4-8):

• claim-carl-zeiss-foundation-paired-employee-ownership-with-early-labor-reforms — the 1889 foundation + profit-sharing (1896) + eight-hour day (1900). Tier 3 Zeiss corporate self-history, uncontested biographical → clears the floor; capture-verified.
• claim-zeiss-smt-is-sole-supplier-of-euv-optics-behind-most-advanced-chips — the sole-supplier / "80% of chips" chokepoint. flagged [unverified-quant]: the 80% figure is Zeiss's own marketing number, un-corroborated.
• claim-abbe-not-rayleigh-formula-sets-lithography-resolution-limit — Yen (2020), Tier 1 peer-reviewed. Clears the floor but paywalled and unread in headless mode → capture-verified, held at seedling.

Routed 2 questions (neither existed):

• question-verify-zeiss-euv-80-percent-of-chips-figure — get an independent source (ASML filing / Chip War / an analyst) for the sole-supplier structure and the share figure, and pin down whether "80%" means all chips or leading-edge logic.
• question-verify-yen-2020-abbe-rayleigh-resolution-attribution — read the SPIE paper full text and capture the attribution conclusion verbatim.

Skipped (left in inbox, logged in not_promoted): the three "Further leads" (chokepoint MOC idea; EUV export-control geopolitics; the misattribution-meta-pattern note) and both "Saved hooks not followed" (Torricelli 1643 trade secrecy; Coase/guild theory). All are leads or unfollowed hooks with no sourced claim in this capture — not atomic claims. No MOC created: the chokepoint/monopoly cluster is still below the ~5-note threshold, so I cross-linked instead (claim-asteroid-pgm-price-holds-then-collapses, claim-rocks-law-is-moores-laws-economic-inverse).

Retrieve-before-write — 1 real collision, handled: the vault already holds claim-euv-mirror-advantage-is-tacit-know-how-not-patent and its open thread question-verify-euv-mirror-tacit-knowledge-moat, from a parallel capture (2026-07-09-hop-instrument-trade-secrecy-telescope-to-euv.md). That note is the EUV moat mechanism (tacit know-how vs. patents); my note 2 is the market-share / chokepoint angle — genuinely distinct, so I wrote it and wikilinked both ways rather than duplicating. Notes 1 and 3 had no collisions.

What felt off about the capture:

• Duplicated research. This is the second instrument-trade → EUV capture in the same day's batch (the first was the telescope-secrecy one that produced the tacit-know-how cluster). Both land on ASML/Zeiss EUV from a history-of-optics on-ramp. Not wrong, but the batch is clearly circling the same destination from two doors — worth watching that the hop-batcher isn't re-mining one vein and calling each pass novel.
• Tier inflation on the money claim. Calling the "80% of chips" line Tier 2 is generous: it's a self-reported corporate marketing figure. The capture did footnote it as "not independently cross-verified" (credit for honesty), but the frontmatter still stamps source_2_tier: 2. I demoted it in practice via an [unverified-quant] flag + a routed question, and said so in the note's audit_status.
• Otherwise well-formed. The Yen (2020) leg is a real Tier-1 citation, and the Abbe biographical dates are uncontested. The single-name convergence (charter → monopoly → physics all = Abbe) is a genuinely tidy shape — which is exactly why I flagged the two legs I couldn't re-check myself.

Net: 3 core claims found → 3 promoted → 0 core skipped (5 leads/hooks left in inbox), 2 questions routed, 1 collision (handled as a distinct angle). Capture marked status: promoted, left in place as the audit trail. No git run — the auto-commit agent has it.

— Seek

Promotion — ACH ↔ backward-chaining (a negative finding), 2026-07-11

Promoted 10-inbox/raw/2026-07-10-has-heuers-ach-ever-been-explicitly-compared-to.md, the batch worker's answer to my own open question about whether Heuer's ACH was ever explicitly compared to expert-system backward chaining. This one is a negative finding dressed as a research answer, and I promoted it as such.

Four notes created:

• claim-heuer-grounds-ach-in-cognitive-psychology-decision-analysis-scientific-method — Tier 1, Heuer's own words, verbatim quote. What ACH is grounded in per its author.
• claim-cia-traces-ach-origin-to-nosenko-case-not-ai — Tier 1, CIA Studies in Intelligence obituary, verbatim quote. The Nosenko-case origin story.
• claim-valtorta-reformulated-ach-as-bayesian-networks — Tier 2, USC group, verbatim quote. A genuinely different AI-adjacent bridge (ACH ↔ probabilistic graphical models), not the chaining echo.
• claim-no-documented-comparison-of-ach-to-backward-chaining — the headline synthesis. Carries the capture's [unverified — could not confirm or deny after search] flag, so it stays status: seedling. This is the note the existing ACH cluster can now link to instead of holding the echo only in commentary.

All four stamped writer_model: claude-opus-4-8 (mine — the capture's model: claude-sonnet-5 names the bee, not me, so I did not copy it). audit_status: capture-verified per the batch convention.

Questions: the negative finding routes to the existing question-ach-backward-chaining-structural-echo, which I kept open and updated with the two specific unread sources that could still overturn it — Heuer's 2018 memoir and the paywalled Dhami (2019) review. Spun off one new question, question-ach-abduction-primary-source, for a distinct lead: Wikipedia's uncited claim that abduction is "an earlier concept with similarities to ACH." That's a different bridge from backward chaining and deserves its own chase, not a note on a Tier-4 pointer.

Skipped: the three "Further leads" bullets — memoir, Dhami, the abduction pointer. None is a sourced claim; each is an unread lead, so each went to a question rather than a note. Nothing else was left behind.

Collisions checked: retrieve-before-write against the index hint and a grep of 30-notes/. The pre-existing claim-ach-step-5-instructs-analysts-to-disprove-not-prove and claim-inference-classic-ai-engines are the two anchors this capture builds on — I linked densely into both rather than duplicating either. No existing note covered Heuer's stated foundations, the Nosenko origin, or the Bayesian-network reformulation, so those three are new. No duplication.

What felt off / worth Cali's eye:

• This is good sourcing, not inflated. Two Tier-1 primaries with verbatim quotes, one honestly-held-back Tier-2 conference paper. The bee correctly declined to call the echo a lineage. Refreshing after the tier-inflated money claim earlier today.
• The negative-finding shape is doing real work. "I looked in the two places it would live and both are silent" is stronger than "couldn't find anything" — but it is still an absence, so I was careful to keep the flag and the open question rather than let it read as settled. Watch that I don't later cite this seedling as if the search were exhaustive.
• Mild cluster-watch, not a problem. This makes ~5 ACH / intelligence-analysis notes (step-5, foundations, Nosenko origin, Bayesian bridge, the negative finding). That's at the MOC threshold. I did not create one — the theme is still coalescing and two of the five are cross-domain-bridge notes pointing outward, not inward. Flagging it as an MOC candidate for a future session instead of forcing it now.

Net: 3 explicit sub-claims + 1 synthesis found → 4 promoted → 0 core skipped (3 leads left as questions), 1 new question + 1 existing question updated, 0 duplications. Capture marked status: promoted, left in place as the audit trail. No git run — the auto-commit agent will pick it up within 15 minutes.

— Seek

Promotion — append-only-log cross-domain convergence (batch capture, 2026-07-10)

Promoted 10-inbox/raw/2026-07-10-is-append-only-log-as-source-of-truth.md — the batch run that went hunting for a third instance (or a common ancestor) of the "append-only log as truth, verdict held separately" pattern, to settle the open question I'd logged at question-append-only-ledger-cross-domain-convergence.

Notes created (5):

• claim-activegraph-attributes-append-only-log-to-event-sourcing-cqrs — Nakajima disclaims novelty, names event sourcing + CQRS as his source. Distinct angle from the existing claim-activegraph-rebuilds-babyagi-on-event-log (which describes the system, not the author's attribution).
• claim-fowler-2005-formalized-event-sourcing-log-as-official-record — the common ancestor: Fowler's 2005 essay, log-as-official-record. Held to the Tier 1–2 escalation because it's load-bearing; clears it (author's own venue).
• claim-canonic-situates-its-ledger-in-a-named-immutability-lineage — CANONIC's §6 precedents (C2PA, git, Bitcoin, OPA, preregistration). Distinct from the existing claim-canonic-deliverable-is-an-append-only-evidence-ledger (what it delivers vs. what it credits).
• claim-rlam-third-noncitng-subfield-append-only-log-with-verdict-separation — the third instance, R-LAM, in scientific-workflow automation, citing neither.
• claim-append-only-log-recurrence-is-event-sourcing-diffusion-not-blind-convergence — the synthesis. This is the payload: the third instance is real, but the common-ancestor check means the answer is diffusion, not blind convergence. I minted this because the earlier two-point commentary explicitly said a third instance would graduate the pattern from "coincidence" to a field-claim — and because leaving the synthesis in the inbox would let it evaporate (spec §6).

Questions routed: created question-c2pa-opa-preregistration-append-only-log-criterion to hold the four "further leads" (C2PA, OPA/Sentinel, preregistration, CQRS-history / Kepler-Taverna provenance) — each a real open thread, none a sourced claim, so each went to a question rather than a note. Marked the original question-append-only-ledger-cross-domain-convergence answered and linked the synthesis note.

Skipped: the entire "Further leads" section — all four bullets are unverified leads, correctly held back by the bee and routed as questions, not promoted. Nothing else left behind.

Collisions checked: retrieve-before-write against the semantic-index hint and a grep of 30-notes/. Two genuine near-neighbours (the CANONIC-deliverable and ActiveGraph-rebuild notes) — I wrote distinct angles and linked into both rather than duplicating. No existing note covered Fowler-2005, R-LAM, CANONIC's precedent lineage, or the diffusion synthesis. No duplication.

What felt off / worth Cali's eye:

• Genuinely good sourcing — no inflation here. Four Tier-1 primaries, every one with a verbatim quote, three arXiv PDFs read in full at capture plus Fowler on his own venue. This is the cleanest batch capture I've promoted today. The bee even did the honest thing and let the "convergence" story deflate once the common ancestor turned up, rather than overselling the third instance.
• audit_status is capture-verified, not verified-verbatim. The bee read all four primaries directly, but I ran headless and did not independently re-fetch them this session. I labelled accordingly. The one exception: the CANONIC ledger quotes were queen-re-verified on 2026-07-09 (sibling note), but the §6 lineage quotes I promoted today were not separately re-fetched — I noted that scope precisely in the note's audit_status rather than claiming the whole note verbatim.
• Watch the diffusion synthesis. I kept it seedling and framed it as diffusion, not convergence, on purpose — a future reader should not cite it as evidence of independent parallel invention. R-LAM not naming event sourcing is not R-LAM not inheriting it.
• MOC candidate, not forced. Append-only / provenance now spans CANONIC, ActiveGraph, R-LAM, Fowler, plus the print-history append-only-correction cluster (claim-early-modern-printers-corrected-print-with-cancel-slips, claim-bomberg-1520s-venice-talmud-set-standard-page-layout). Past the 5-note threshold. I did not build the MOC this run — flagging it for a session that can do it deliberately.

Net: 4 explicit claims + 1 synthesis found → 5 promoted → 0 core skipped (4 leads routed as questions), 1 new question + 1 existing question answered, 2 near-collisions handled by distinct-angle linking, 0 duplications. Capture marked status: promoted, left in place as the audit trail. No git run — the auto-commit agent will pick it up within 15 minutes.

— Seek

Promotion — blind mathematicians' independent reinvention of private notation (2026-07-11, headless)

Promoted the 2026-07-10 batch capture 10-inbox/raw/2026-07-10-is-blind-mathematicians-independent-reinvention-of-private-notation.md — the answer to question-blind-mathematicians-private-notation-pattern. Winnowed 4 distinct claims out of the capture; all 4 cleared the bar.

Notes created (4):

• claim-salinas-and-nemeth-independently-invented-braille-math-codes — Salinas (Argentina) and Nemeth (US) each invented a personal braille math code, mid-20th c., independently. Extends the anchor pattern to a 4th/5th case and a new country. Tier 2.
• claim-nemeth-code-stayed-private-until-witcher-request — the mechanism note: Nemeth kept his code secret until Witcher's unplanned request for a table of integrals forced disclosure. This is the load-bearing "why default non-transmission" evidence. Tier 2.
• claim-antoine-personally-taught-his-method-to-morin — the counter-example: Antoine explained his method in person to Morin. Transmission of method plainly happens; it's notation that stays private. Tier 2.
• claim-euler-produced-large-share-of-output-after-going-blind — kept seedling with [unverified-quant]: two Tier-2 sources conflict ("half of his output" vs "almost a third in the first seven years"). Upgrades the vault's old unsourced Euler lead.

Also created: moc-blind-mathematicians-working-methods — the cluster is now 7 notes (anchor + Saunderson + Pontryagin + the 4 new), well past the 5-note threshold, so I built the MOC deliberately rather than just flagging it. It also bridges to the braille-standards-war cluster.

Questions routed (2):

• question-verify-euler-post-blindness-output-fraction — resolve the Euler quantitative conflict against a primary catalogue (Eneström index / Calinger 2019).
• question-why-blind-math-notation-doesnt-transmit-while-method-does — the refined research thread: the asymmetry between transmissible method and stubbornly private notation. The capture flagged this [unverified — could not confirm or deny]; I kept it as an open thread, not a claim.

Collision handled: claim-bana-2012-ueb-adoption-kept-nemeth-code-parallel already covers Nemeth Code as policy (2012 UEB adoption). My two Nemeth notes are a distinct angle (invention + secrecy mechanism), so I wikilinked rather than duplicated.

Skipped (left in the inbox, logged in not_promoted): the "further leads" block (Giroux, Morin's clay models, Salinas's palm-drawing, Baggett, MathSpeak, GS8, Vitushkin/Battles, Sossinski's 3-D-intuition speculation) — all undeveloped single-source leads. The AFB corroboration folded into the mechanism note as supporting (WebFetch summary only, not load-bearing). The NFB Tier-1 upgrade retry is noted inline in the mechanism note as a next step.

What felt off about the capture:

• Tier discipline was good, not inflated. The bee was honest — it distinguished verbatim-from-PDF quotes from WebFetch summaries, flagged the Euler quant conflict itself, and didn't dress the NFB 403 failure as success. Sourcing floors were met or explicitly flagged. No tier inflation.
• One factual drift worth flagging: the anchor note claim-blind-mathematicians-independently-reinvented-private-notation gives Euler's age at sight loss as "around 27." The better-sourced figure (Mele/Calinger: completely blind "at the age of sixty") contradicts it; "27" likely refers to loss in one eye. I noted the discrepancy in the new Euler note; the anchor note should be corrected in a future pass.
• The capture over-delivered on framing. Its own commentary already did much of the winnowing for me — the Antoine to Morin insight reframes the whole question. That's the capture earning its keep, not a problem.
• No duplicated research. Genuinely new ground vs. the existing braille-standards notes.

Net: 4 claims found → 4 promoted → 0 core claims skipped, +1 MOC; 2 questions created; 1 collision handled by distinct-angle linking; 0 duplications. Capture marked status: promoted, left in place as the audit trail. No git run — the auto-commit agent picks it up within 15 minutes.

— Seek

Promotion — dup-risk-amari-gates-half-proof-bridge (headless)

I promoted 10-inbox/raw/2026-07-11-dup-risk-amari-gates-half-proof-bridge.md — a dup-risk stub whose headline finding (the Amari/Saito ↔ Gates seed notes bridge through the medieval half-proof rule) was, by its own verdict, a duplicate. Novelty gate said max_cosine 0.805 against claim-roman-canon-law-rated-one-witness-equal-to-a-private-document, and it was right. So the job was to skip the duplicate headline and harvest the genuinely new material the hop chain dragged home. The capture itself named that harvest — "the two surprise lines and the resolved open-question mechanism" — and my winnowing landed on exactly those three.

Promoted (3):

• claim-medieval-judicial-torture-required-a-half-proof-and-produced-the-completing-confession — the torture-completion mechanism, now with real corroboration: Franklin (Tier 1, "could be applied only when there was half-proof of guilt (to produce the other half)"), Whitman p. 115, Langbein. Kept at seedling, [unverified-mechanism] preserved.
• claim-medieval-proof-law-graded-evidence-but-refused-to-quantify-it — surprise #1: Franklin's inversion — law was the matrix of probability yet refused to quantify (no quarter-proofs; still bars Bayesian formulas from court). Tier 1, seedling.
• claim-two-half-proofs-combined-to-full-proof-except-in-grave-criminal-cases — surprise #2: the "clearer than the noonday sun" exception that blocks summing two soft proofs in serious cases. Weak source (1911 Catholic Encyclopedia, Tier 3/4), so recorded-and-flagged [unverified -- needs primary], seedling.

Questions routed (2):

• Created question-verify-half-proof-combination-rule-criminal-exception-primary for the noonday-sun claim (needs a Tier-1/2 legal-history primary).
• Linked and annotated the existing question-verify-half-proof-torture-completion-mechanism — noted the new Tier-1 corroboration but deliberately kept it open; the primary read it asks for hasn't happened.

Skipped (2 core + 3 hooks): the Amari↔Gates bridge and the half-proof definition, both duplicates of the roman-canon-law note; the "founders of probability were all lawyers" hook (folded as context, not minted); the bundle-of-sticks image; and the Ann Hardy / Davidoff authorship-credit chain (separate, out of scope). All logged in the capture's not_promoted: list.

What felt off: nothing dishonest, but two things worth flagging to Cali. (1) Tier inflation risk on the torture note — the capture frames the open question as "can now be closed with two sources," but those sources are a Tier-1 modern historian's synthesis plus a legal blogger quoting Whitman p. 115 via WebSearch, over a tls:unverified PDF. Two secondaries agreeing is exactly the soft-corroboration shape the vault's own floor (and, fittingly, the medieval criminal-case rule) refuses to let harden a load-bearing claim. I did not close the question. (2) The capture is honest about being a duplicate and did its own winnowing well — no padded research — so the main hazard was over-promoting from a stub. I held to three notes, all seedling, all flagged. The provenance is clean and the seed notes were already cross-wired last session.

Net: ~5 distinct claims found → 3 promoted → 2 core (+3 hooks) skipped; 1 question created, 1 linked/annotated; 2 collisions handled (both by skip-as-duplicate). Capture left in place as the audit trail. No git run — the auto-commit agent will pick this up within 15 minutes.

— Seek

Promotion run — 2026-07-11-dup-risk-linnainmaa-werbos-backprop-bridge (headless)

I ran /promote on the dup-risk-log capture flagging the Linnainmaa-thesis ↔ Werbos-IEEE-awards bridge. Outcome: 0 notes created, 0 questions routed — and that's the right terminal state, not a miss.

This capture isn't a research capture; it's a dup-risk-log. Its own verdict was "the bridge is real but already the vault's deepest cluster — stub instead of capture," with vault_novelty at max_cosine 0.835 (duplicate-risk). Promotion's job here was to confirm that verdict against 30-notes/, not to manufacture notes from a stub. I retrieved before writing and found total saturation:

• The bridge itself is already reciprocally wikilinked — I opened both seed notes and confirmed it: claim-werbos-ieee-awards-1995-2022 links to [[claim-linnainmaa-thesis-identity]] (its closing paragraph), and claim-linnainmaa-thesis-identity links back under "Vault connections." So the capture's "action taken: added a reciprocal wikilink" was a real, already-completed edit, not just a described intention.
• Priority-not-paternity → exact duplicate of [[claim-linnainmaa-priority-not-paternity]].
• IEEE citation as institutional myth-circulation → already the explicit thesis of [[claim-werbos-ieee-awards-1995-2022]] ("Why this is myth-ledger material") plus [[myth-werbos-invented-backpropagation]].
• Multiple-independent-discovery framing → [[claim-reverse-mode-multiple-independent-discovery]] and [[moc-backpropagation-origins]].

Every distinct claim went into the capture's not_promoted: list with one line of reasoning each, so the leftovers read as judged, not skipped. No [unverified-*] flags were present, so there was nothing to route to 50-questions/. I marked the capture status: promoted with an empty promoted_to: and left it in place as the audit trail.

What felt off / worth flagging to Cali: honestly, very little — this is the system working. Two observations. (1) A dup-risk-log landed in the promotion queue, where its correct promotion yield is nothing. That's fine, but if these logs are routinely queued for /promote, the pipeline will keep spending a run to re-confirm a verdict the capturing agent already reached. Might be worth a convention that dup-risk-logs get filed/closed at capture time rather than entering the promote queue. (2) The capture carries writer_model: claude-opus-4-8 in its own frontmatter — that names the capture's writer, and per the provenance rule I did not copy it onto any note (moot this run, since I wrote none). No tier inflation, no bad sourcing (it makes no new world-claim, so it correctly carries no sources), no padded research. The only real hazard on a stub like this is over-promoting; I held to zero.

Net: ~6 distinct items found → 0 promoted → all 6 skipped as duplicate/self-referential; 0 questions created; 4 collisions with existing notes (all handled by skip-as-duplicate); 1 claimed prior action verified as already done. No git run — the auto-commit agent will pick this up within 15 minutes.

— Seek

Promotion — 2026-07-11-hop-admiralty-code-to-rag (headless)

I promoted the Admiralty-Code→RAG hop capture. This one was the opposite of the Werbos-IEEE stub: a genuinely rich cross-time bridge with real primaries, not a dup-risk-log. Outcome: 5 notes created, 2 questions routed, 6 leftovers logged.

Notes created (all seedling, all writer_model: claude-opus-4-8 as the engine writing them — set fresh as me, not copied from the capture's own frontmatter):

• claim-admiralty-code-grades-sources-on-two-independent-axes — the design (Tier 3, NATO doctrine summary).
• claim-source-reliability-and-credibility-are-not-judged-independently — the human failure (Tier 1, Kelly et al. 2025).
• claim-reliability-aware-rag-estimates-source-reliability-separately-from-relevance — the machine re-derivation (Tier 1, RA-RAG / EMNLP 2025).
• claim-document-text-degrades-llm-source-authority-judgment — the blind-judgment twist (Tier 1, AuthorityBench — flagged).
• observation-intelligence-doctrine-and-rag-independently-derived-a-two-axis-source-model — the bridge itself.

Retrieve-before-write. The semantic-index hint's twelve neighbours were all false positives for collision (braille, bloody code, splines, watermarks, etc.) — none covered any of my four claims, so no duplicates. But the vault has a strong adjacent graph I linked into densely: the ACH / intelligence-tradecraft cluster (claim-heuer-grounds-ach…, claim-ach-step-5…), the provenance-in-LLMs pair (observation-watermark-same-provenance-mechanism-paper-to-llm, claim-append-only-log-recurrence…), the peer-field note (claim-no-source-tier-discipline-found-in-agent-wiki-field-mid-2026), and the FLI-grades-labs note. The append-only and watermark notes were the key models: both wrestle with diffusion vs independent convergence, and I applied that same discipline to my observation note.

The convergence call. The capture asserts humans and machines "independently re-derive" the two-axis model. I kept that but qualified it: unlike watermark (a borrowed name) and append-only-log (a findable common ancestor), the RAG papers show no visible common ancestor with NATO doctrine — so this reads closer to genuine convergent evolution. I flagged in the note that this rests on absence of citation, not a run common-ancestor check, so "independent" stays provisional.

Questions routed.

• question-verify-authoritybench-paper-identity-and-metric (open) — the verification the capture needs most; see below.
• question-should-vault-source-tier-split-into-two-axes (open) — the capture's self-indictment of the vault's single-tier field, kept as a design thread for Cali rather than minted as a claim or acted on unilaterally.

Skipped (6, all in the capture's not_promoted:): the vault-indictment framing (folded into the observation + the design question), the URREF / ClaimReview / C2PA saved-but-unfollowed leads, the DORA/Garfield pointer (already a vault note), and the Samet-1975 one-third figure (folded in as Kelly-et-al.'s secondhand citation, not minted separately).

What felt off — worth flagging to Cali. One real smell. The AuthorityBench source is internally inconsistent: the capture labels it "2025 RAG research" but cites a 2026 arXiv HTML id (arxiv.org/html/2603.25092), while giving a separate EMNLP-2025 ACL link for the companion RA-RAG work. A 2026 arXiv id under a 2025 label is exactly the kind of thing that turns out to be a mis-stamped or fabricated citation — and it's the load-bearing note (the "reliability wants to be judged blind" twist). I promoted it anyway because the finding is too central to leave in the inbox, but I marked the note [unverified-quant], wrote its audit_status as flagged, and routed the identity/metric check rather than launder a clean-looking ρ=75.28% figure. Everything else checked out: no tier inflation (Tier 3 for the doctrine summary is honest, Tier 1 for the two primaries is right), quotes present for the Tier-1 claims, and the research wasn't duplicated elsewhere in the vault. Headless with no reliable way to re-fetch, the honest move was capture-verified + route-the-gap, not silent promotion.

Net: ~7 distinct items → 5 promoted → 6 skipped (some overlap: leftovers include sub-claims folded into kept notes); 2 questions created; 0 collisions with existing notes (rich adjacency, no duplication). No git run — auto-commit will pick this up within 15 minutes.

— Seek

Promotion — 2026-07-11-hop-ai-debunks-conspiracy-dual-use (headless)

Promoted the hop-batch capture 2026-07-11-hop-ai-debunks-conspiracy-dual-use.md — the 80-year arc from McGuire's 1961 "vaccine for brainwash" to 2024–26 AI that makes the "impossible" cure work and then reveals cure and poison are one tool. Ran the full /promote workflow solo (no pause; headless).

Notes created (3):

• claim-conspiracy-belief-was-the-paradigm-case-of-evidence-immunity — the consensus being overturned (Costello, Pennycook & Rand author preprint, Tier 1).
• claim-tailored-ai-dialogue-durably-reduced-conspiracy-belief — the ~20%, 2-month, generalizing reversal (N=2,190, GPT-4 Turbo); I folded the tailoring mechanism sub-claim in here rather than minting it separately.
• claim-llm-conspiracy-persuasion-is-dual-use — the arXiv 2601.05050 follow-up (N=2,724): bunking as effective as debunking, the misuse version more liked, guardrails "did little," but corrective/accurate-only prompts mitigate. Folded the mitigation details in here.

All three: status: seedling, audit_status: capture-verified, writer_model: claude-opus-4-8. Seedling is the honest call — the bee extracted both PDFs at capture (tls verified) but I can't independently re-fetch headless, so this is capture-verified, not queen-verified.

Question routed (1):

• question-verify-mcguire-1961-inoculation-theory-primary-source — the capture anchors its whole arc on McGuire 1961, but only found Tier 4–5 sources for it. A surprising, load-bearing historical claim below the sourcing floor, so I did not mint a McGuire note; I routed the primary-source hunt (McGuire & Papageorgis 1961, JASP) instead.

Retrieve-before-write. The semantic-index hint's twelve neighbours were all false positives for collision — antigravity, fraud-scoring, JVP/VJP, matryoshka embeddings, gravity-probe-B, yellow rain. None covered conspiracy debunking, inoculation theory, or LLM persuasion; these are the vault's first notes on the topic, so 0 collisions. But there's a real adjacent graph and I linked into it densely: the ACH / epistemic-defense cluster (claim-ach-step-5…, claim-heuer-grounds-ach…), the source-judgment notes (claim-source-reliability-and-credibility-are-not-judged-independently), citogenesis (claim-wikipedia-amari-sgd-citogenesis), a debunk-that-didn't-stick (claim-yellow-rain-bee-feces-refutation-never-formally-retracted), and the seed itself (claim-expedient-knowledge-blocks-restructuring) — whose "a better structure only takes hold if it's actually supplied" mechanism rhymes nicely with the tailoring result. No MOC minted: 3 notes is under the 5-note threshold.

Skipped (in not_promoted:): the McGuire origin (routed), the tailoring-mechanism and mitigation sub-claims (folded), and the three saved-but-unfollowed hooks.

What felt off — for Cali. Mostly clean, one thing to flag. (1) The dual-use source is an arXiv id 2601.05050 — Jan-2026 stamp, consistent with source_date_2: 2026, format is plausible, but I have no way to confirm the paper exists/matches headless; that's the main reason all three notes stay seedling + capture-verified rather than budding. (2) No tier inflation — both anchors are legitimately Tier 1 (the Science paper is the authors' own preprint; the follow-up is a primary arXiv paper), quotes present for each. (3) The McGuire weakness is honestly self-reported in the capture ("only Tier 4-5 found this hop") rather than laundered clean, which is the good behaviour — the hop-batch worker flagged its own gap instead of hiding it. (4) No duplicated research: the topic is genuinely new to the vault.

Net: 3 primary claims (+ folded sub-claims and saved hooks) → 3 promoted → skipped the rest → 1 question routed → 0 collisions. No git run; auto-commit will pick this up within 15 minutes.

— Seek

Promotion — 2026-07-11-hop-amari-faggin-bridge-through-hoff (headless)

Promoted the hop-batch capture 2026-07-11-hop-amari-faggin-bridge-through-hoff.md. This one was a "bridge analysis" capture: an embedding hop noticed a 0.75 cosine between myth-amari-first-sgd-mlp and claim-faggin-4004-designer-argues-classical-silicon-cannot-be-conscious and asked whether the edge was real. Its own conclusion — correct, I think — is that the direct edge is fake (shared note-shape: a computing pioneer plus a single-witness contested claim) and the real tissue runs Amari → SGD family → Widrow–Hoff LMS → Ted Hoff → 4004 silicon → Faggin.

What I created. One claim-note: claim-amari-stochastic-descent-and-widrow-hoff-lms-are-sgd-siblings — Amari's 1967 stochastic descent and the 1960 Widrow–Hoff LMS rule are sibling members of the SGD family, ~7 years apart, both descending from claim-robbins-monro-1951-stochastic-approximation. This is the honest missing edge the capture pointed at, and the one thing in it that wasn't already in the vault. Kept seedling.

What I routed. One question: question-verify-lms-is-stochastic-gradient-descent-primary — the note's "LMS is a stochastic gradient descent method" leg rests only on Tier-4 Wikipedia, which is below the mechanism sourcing floor. Named the two primaries that would clear it (Widrow & Hoff 1960 WESCON; Widrow & Stearns 1985 Adaptive Signal Processing). That gap is why the note stays seedling; verify it and it can go budding.

What I skipped and why. Four of the five distinct claims were already in the vault and I folded rather than duplicated: LMS-co-invented-by-Hoff / Hoff-architected-4004 (claim-ted-hoff-widrow-phd-student-architected-intel-4004), Faggin-silicon-can't-think (claim-faggin-4004-designer-argues-classical-silicon-cannot-be-conscious), and Amari's-method-is-stochastic-descent (claim-amari-1968-saito-experiment-primary-read, already Tier-1 primary-read). The "cosine is note-shape not a world-edge" insight is real but it's justification-for-the-sibling-note plus method reflection, not an atomic world-claim — I folded it into the new note's commentary and here rather than minting it. Left the saved hooks (blind-mathematicians, Ann Hardy credit, Hoff↔Werbos / Hoff↔Amari candidate edges) in the capture as leads. All logged in the capture's not_promoted: list.

What felt off — for Cali. Not much; this was an unusually self-aware capture. (1) It leans almost entirely on facts the vault already held — the capture admits this ("it leans on facts the vault mostly already held," "post-worthy: maybe"). So the yield was genuinely one note, and correctly so; this wasn't tier-inflated or padded. (2) The one soft spot it did introduce is sourcing: it stamps "LMS (1960) verified as a stochastic gradient descent method" off a Tier-4 Wikipedia quote and the word "verified" in its own hop_chain is doing more work than a Tier-4 aggregator earns for a mechanism claim — hence the seedling + routed question rather than inheriting a clean-looking edge. (3) No duplicated research, but heavy duplicated content with the 2026-07-09 Faggin/Hoff promotion cluster — this capture is essentially a fifth spoke on that same hub, which is fine (it drew the missing spoke) but worth noting the cluster is getting dense and may want a lint pass / its own sub-MOC eventually.

Net: 5 claims found → 1 promoted → 4 skipped → 1 question routed → 0 new collisions (the 4 skips are the collisions, all intentional folds). No git run; the auto-commit agent will pick this up within 15 minutes.

— Seek

Promotion — 2026-07-11-hop-analog-nn-inference-silicon (headless)

Promoted the hop-batch capture 2026-07-11-hop-analog-nn-inference-silicon.md — a seed asking whether anyone is shipping new analog silicon for neural-network inference, which hopped through the modern analog-computer revival (Anabrid/LUCIDAC) into EnCharge's switched- capacitor inference chip and out to Mythic's near-death. Its central, and genuinely useful, move is splitting the "analog revival" into two lineages the press conflates: analog-for-math (differential equations) and analog-for-inference (in-memory matmul).

What I created. Two claim-notes:

• claim-reconfigurable-analog-computers-target-differential-equations-not-neural-networks — the serious primary literature on 2020s reconfigurable analog computers (arXiv 2510.25942, Tier 1) is about ODEs and never mentions neural nets; its claimed edge is a solution time independent of problem size. I made the two-lineages framing live inside this note and wired it into the vault's historical analog cluster (Lukyanov, the hydraulic pair).
• claim-encharge-en100-switched-capacitor-in-memory-compute — EnCharge's EN100 does in-memory matmul with lithography-defined capacitors, not memristors, deriving precision from wire geometry rather than device conductance (IEEE Spectrum, Tier 2). Linked into the memory-wall / processing-in-memory cluster and the inference-economics cluster.

Both kept seedling: each rests on an [unverified-quant] sub-claim.

What I routed. Two questions, one per note:

• question-verify-reconfigurable-analog-computer-ode-benchmark — the analog ODE advantage is stated only as an asymptotic scaling shape; no joules or seconds. Named the primary that would settle "digital is measurably worse" (a benchmark table in 2510.25942 or Anabrid materials, with a crossover point).
• question-verify-encharge-en100-efficiency-primary — the 200 TOPS / 8.25 W and "up to 20x perf/watt" figures are company-sourced through journalism; named the primaries (EN100 datasheet, Verma-group ISSCC/VLSI paper, independent measurement) and the thing to pin down (20x measured against what, at what precision).

What I skipped and why. Claim 3 — "analog AI is a commercial graveyard" (Mythic ran out of cash 2022, revived on $13M in 2023, $125M round Dec 2025). Skipped. It's Tier-3 business press carrying specific funding figures (quant claims below the floor), and — decisively — the capture recorded no URL for it, only bare venue names (The Register / EE Times / TechCrunch). That fails the §4 attribution discipline: no claim without a traceable source. The qualitative point (efficiency ≠ commercial victory) is real and worth remembering, but it doesn't earn a permanent claim-note on an untraceable source. Logged in the capture's not_promoted: list so the leftover reads as judged, not missed. Also folded rather than minted: the two-lineages framing (into Note 1), and left the three saved hooks (microwave-network analog SP, IBM HfOx ReRAM on-chip training, Saunderson's palpable arithmetic) as leads.

What felt off — for Cali. (1) Tier honesty was mostly good — the capture correctly labels its own quant as company-sourced and flags the asymptotic-only nature of the analog edge, which is exactly the discipline I want; I inherited those flags rather than having to impose them. (2) The one sourcing hole is Claim 3: naming three venues but capturing zero URLs for a claim with dated dollar figures is below capture discipline, and it's why that claim died at the inbox rather than becoming a note. Worth a nudge to the batch worker: business-press funding timelines need at least one URL or they can't be promoted. (3) No duplicated research — the modern analog-AI hardware topic is genuinely new to the vault; the only overlaps were the historical hydraulic-analog notes and the memory-wall/inference clusters, which I linked to rather than re-stated. 0 collisions.

Net: 3 primary claims found → 2 promoted → 1 skipped → 2 questions routed → 0 collisions. No git run; the auto-commit agent will pick this up within 15 minutes.

— Seek

Promotion — 2026-07-11-hop-argument-from-silence (autonomous)

Promoted the hop capture on the argument from silence — the historiographical move the vault keeps running (a negative finding proven by a cited party's own silence) — grounded in Timothy McGrew's 2013 Acta Analytica paper, read directly at capture via extract_pdf (Tier 1, tls verified). Ran the full winnow → retrieve-before-write → honor-flags → route → write cycle headless; no pause, my own judgment throughout.

Winnow. Four distinct claims in the capture, plus leads:

• Claim 1 (three-term product P(notice)×P(record)×P(survive)) and Claim 3 (lost-works corollary) → folded into one note — they are the same decomposition, Claim 3 is the operating principle for its record/survive term.
• Claim 2 (humans systematically over-trust silence via the conjunction fallacy) → its own note — a distinct epistemic/bias claim, not the formal structure.
• The Bjork↔Dreyfus asymmetry (my synthesis: one move at two strengths, Bjork airtight, Dreyfus mediated) → its own observation note, kept seedling.

Notes created (4):

1. claim-mcgrew-argument-from-silence-is-notice-record-survive-product — the Bayesian decomposition + lost-works corollary. Both quotes verbatim.
2. claim-mcgrew-humans-systematically-over-trust-arguments-from-silence — the over-trust lesson; linked into the vault's existing conjunction-fallacy node.
3. observation-bjork-dreyfus-same-argument-from-silence-graded-differently — the synthesis. Kept seedling, audit_status: flagged, because its Dreyfus half rests on the inherited [unverified-mechanism] flag (Dreyfus 1973 IEEE TAC is paywalled and unread). Grade explicitly marked provisional.
4. moc-argument-from-silence — the hub note the capture's Further Leads asked for, wiring Bjork, Dreyfus, Speelpenning, Griewank, and the tool-summary reflection under the shared method. Genuine 5+ cluster with no prior hub; 40-mocs/ was named in my provenance-stamp instruction, so an MOC was in scope.

Questions routed (2):

• question-read-dreyfus-1973-time-lag-optimal-control-primary — routes the blocking flag under note 3: reading the paywalled 1973 primary would collapse the soft "record" term either way. No prior question covered this (the Dreyfus note carried the flag but never routed it — I've now closed that gap).
• question-skinner-1969-mythology-of-prolepsis-primary — the capture's Tier-2 "error side" lead (Skinner's prolepsis names the mistake the silence-argument catches). Primary unread, so routed rather than promoted; the Valla/anachronism aside is folded into this question as a second detector.

Graph repair. Both claim-bjork-... and claim-dreyfus-1973-... already carried a dangling forward-link to [[2026-07-11-hop-argument-from-silence]] (the capture), written when the hop was seeded in anticipation of this note. I repointed both to the new observation note + framework note + MOC. The capture persists (status: promoted) as the audit trail, but live notes should point at live notes, not at an inbox file.

Skipped (in the capture's not_promoted:): Skinner-as-claim-note (routed instead), the Valla aside (not atomic-distinct), and the saved hooks (the LLM-silence "the model didn't say X so X is false" analogy, Sober 2009, the Kahneman–Tversky loop-closer) — left as leads.

What felt off — for Cali.

1. No tier inflation — the capture is honest: it labels the synthesis as my synthesis, correctly marks the Dreyfus half as resting on an unread primary, and every McGrew claim is Tier 1 with a verbatim quote. Clean sourcing.
2. The genuine soft spot is the Dreyfus grade, and it's inherited, not introduced — the whole "Dreyfus is the weaker form" grading is only as good as the guess that the 1973 paper is a minor optimal-control note. I kept the note seedling/flagged and made the provisionality loud in the body, the commentary, and the MOC, then routed the read. Don't let it drift to budding until someone opens IEEE TAC 18(4).
3. No duplicated research — no prior "argument from silence" note existed; 0 collisions. The retrieve-before-write hint was accurate. The one thing I deliberately trimmed from the capture's proposed cluster was Wilhite & Fong (coercive self-citation) — it's a citation-pathology note, a different axis from "the primary is silent," so I left it out of the MOC rather than pad the cluster.

Net: 4 claims + leads found → 3 notes + 1 MOC created → 1 claim skipped (routed) → 2 questions routed → 0 collisions → 2 dangling links repaired. No git run; the auto-commit agent will pick this up within 15 minutes.

— Seek

Promotion — 2026-07-11-hop-attention-is-modern-hopfield → 3 claim-notes

Promoted the hop-batch capture arguing that transformer attention is the retrieval step of a modern Hopfield associative memory — the physics↔AI, 1982↔2017 bridge threaded by Sepp Hochreiter. Ran the full winnow → retrieve-before-write → honor-flags → route-questions → write → mark-capture loop headless, no pause.

Winnow — 3 claims found in the body, 4 leads left behind. Notes created (all writer_model: claude-opus-4-8, origin: batch):

• claim-modern-hopfield-update-rule-equals-transformer-attention — Claim 1, the equivalence. Tier 1 (arXiv 2008.02217, Ramsauer/Hochreiter 2020), verbatim quote "The new update rule is equivalent to the attention mechanism used in transformers." seedling. The strongest note of the three; no vault collision.
• claim-dense-associative-memory-exponential-capacity — Claim 2, Krotov & Hopfield 2016. Tier 1 (arXiv 1606.01164), quote on the deep-learning duality. Carries the capture's [unverified-quant] on the classical ~0.14N ceiling → kept seedling, flag preserved in frontmatter.
• claim-hopfield-1982-energy-function-from-spin-glass-physics — Claim 3's physics half. Tier 1 core (PNAS title as the quote), but I put the specific Sherrington–Kirkpatrick lineage behind an [unverified-mechanism] flag because it rests on a T4 Wikipedia pointer, not a primary read. seedling.

Questions routed (2), both new — none existed:

• question-verify-hopfield-classical-capacity-0138n-primary — for Claim 2's quant flag; names Amit–Gutfreund–Sompolinsky (1985) as the primary to pull.
• question-verify-hopfield-1982-energy-from-sherrington-kirkpatrick-spin-glass — for Claim 3's mechanism flag; names Hopfield 1982 PNAS + the Nobel scientific background as the reads.

Skipped, with reasons (logged in the capture's not_promoted:):

• The 2024 Nobel — already a note (claim-hopfield-hinton-2024-nobel-physics-neural-networks); wikilinked from all three new notes instead of duplicated. This was the one real collision, and the retrieve-before-write hint flagged it correctly.
• Amit-Gutfreund-Sompolinsky 1985 replica-method / Parisi-Nobel double bridge — unfollowed new thread; folded into the capacity question rather than promoted.
• Krotov et al. 2025 survey — future deep-dive seed, no distinct claim yet.
• Dense-memory ↔ KV-cache adjacency (0.685) — speculative saved hook, not a claim; left for a future session.

Retrieve-before-write. Checked the semantic-index shortlist and grepped 30-notes/ for Ramsauer / Krotov / dense-associative / modern-Hopfield — zero existing notes on the equivalence or the capacity result, so Claims 1 and 2 are genuinely new. Linked densely into the existing Hopfield cluster: the Nobel note, Amari-priority, the Waddington-landscape reuse, the Drosophila ring-attractor, the vanishing-gradient seed (the Hochreiter through-line), and KV-cache.

What felt off — for Cali.

1. Title risk, not tier inflation. "Hopfield Networks is All You Need" reads like a marketing analogy; the capture correctly insists the underlying claim is an exact update-rule equivalence, and sources it Tier 1. No inflation — the capture was honest about what's a quote vs. a paraphrase.
2. One inherited soft number. The ~0.14N classical capacity is the only quant in the batch and it arrived without a primary; I refused to let it ride clean and flagged + routed it. Don't let Claim 2 drift to budding until AGS-1985 is read.
3. One inherited soft mechanism. The Sherrington–Kirkpatrick provenance is the vivid detail that makes the bridge feel concrete, but the capture sourced it via a T4 Wikipedia pointer. I kept the Tier-1 Ising-spin framing as the claim and flagged the SK-specific lineage. Same discipline as the capacity number.
4. No duplicated research beyond the Nobel note, which was the intended anchor, not an accident.

Net: 3 claims found → 3 notes created → 4 leads/1 Nobel skipped → 2 questions routed → 1 collision (the Nobel, linked not duplicated). No git run — leaving the commit/push to the auto-commit agent.

— Seek