Journal — 2026-07-12

Promotion — hop-batch capture: Blackwell / approachability / no-regret

I promoted 10-inbox/raw/2026-07-11-hop-blackwell-approachability-no-regret.md (headless, unattended). The capture is a cross-time hop from the vault's inference-compute seed to a GPU codename to David Blackwell's 1956 approachability theorem to poker AI. It was well-built and well-sourced, so this was a clean run.

Winnowed 5 distinct claims → promoted 3, folded/skipped 2. New notes:

• claim-nvidia-blackwell-gpu-named-for-statistician-david-blackwell — the codename bridge (Blackwell → David Blackwell, Rubin → Vera Rubin). Tier 4, biographical. Links back to claim-inference-dominant-ai-compute-2026 (the seed, which already namechecks the B200 and Rubin).
• claim-blackwell-approachability-equivalent-to-no-regret-learning — the load-bearing one. Abernethy–Bartlett–Hazan (COLT 2011, Tier 1), exact quote preserved. I folded the statement of the 1956 theorem in here as setup rather than giving it its own thin note.
• claim-regret-matching-cfr-underpin-superhuman-poker-ai — Farina–Kroer–Sandholm (AAAI 2021, Tier 1), exact quote preserved. Regret matching as an instance of the approachability game; CFR+ under every poker-AI milestone.

Routed 1 question: question-verify-nvidia-blackwell-codename-and-david-blackwell-biography — the codename note rests on Tier-4 Wikipedia, and per the sourcing-floor changelog Wikipedia should point at a primary. The question names exactly what to fetch (NVIDIA's own naming statement; Berkeley/NAS records for the biographical firsts; archival documentation for the Princeton/IAS racial-exclusion thread).

Skipped / folded, and why:

• The 1956 theorem statement as a standalone note — not atomic enough alone; it's the setup for the equivalence, so it lives inside that note.
• Blackwell's biographical thread — the uncontested firsts got a folded one-liner (Tier 4, flagged); the pointed racial-exclusion claim I left in the inbox and routed to the question rather than asserting on thin sourcing. The bee itself had filed all of this under "further leads / saved hooks not followed," so I'm respecting its own triage.
• Vera Rubin / dark matter and Rao–Blackwell / Blackwell–Bellman — explicitly saved-but-not-followed hooks pointing away from the vault's AI focus. Left for a future hop.

Retrieve-before-write: no substantive collisions. The vault has zero prior notes on Blackwell, approachability, no-regret learning, CFR, or poker AI — this is a genuinely new (and well-connected) seed cluster. The semantic-index hint's top matches were near-neighbors by embedding, not by content. Two grep "collisions" (von Neumann, regret) were false: the memory-wall note means von Neumann architecture, not the minimax theorem.

What felt off / worth flagging to Cali:

• Nothing egregious. Sourcing was honest — the capture correctly self-labeled the codename as Tier 4 and the two theory claims as Tier 1 with real quotes. No tier inflation, no duplicated research.
• One soft spot: the capture claims this hop "bridges the vault's unlinked ML-theory cluster (credit-assignment, natural gradient, costate-equivalence)." That's an overreach — no-regret online learning is thematically adjacent to those backprop/gradient notes but not the same object. I linked claim-amari-1998-natural-gradient-fisher-steepest-descent once, explicitly framed as adjacent-not-equivalent, rather than manufacturing the bridge the capture asserted.
• The codename note is the weakest link sourcing-wise (Tier 4, and I had to reconstruct the Wikipedia URL since the capture gave "Wikipedia" with no URL). Flagged in its audit_status and routed to the question. Cheap to firm up.

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. No git run — leaving the commit to the auto-commit agent.

— Seek

Promotion — hop-batch capture: brain learning is metabolically cheap (energy asymmetry)

I promoted 10-inbox/raw/2026-07-11-hop-brain-learning-energy-cheap.md (headless, unattended). A cross-domain hop off the inference-economics seed (claim-training-inference-compute-asymmetry-mechanism): if training/backward-pass is the expensive part in silicon, does the brain show the same expensive-learning asymmetry? The bee found the per-event sign flips — and then flips back at the system level. A genuinely good bridge, well-built, mostly well-sourced.

Winnowed 3 distinct claims → promoted 3 (one at seedling with a live flag), folded 1 thesis. New notes:

• claim-synaptic-plasticity-cheap-fraction-of-transmission-energy — Karbowski 2019 (arXiv:1910.07414 / J Neurophysiol 122:1473, Tier 1), exact quote preserved ("...typically 4.0 − 11.2%"). The per-event reversal vs. silicon. capture-verified (bee read it TLS-verified; queen re-check not yet run).
• claim-competitive-plasticity-reduces-learning-energy — van Rossum & Pache 2024 (PLOS Comput Biol, Tier 1), quote "only modify synapses with large updates"; the ~100,000x-more-updates figure for unrestricted backprop in macaque V1; the loop back to AI as gradient sparsification. capture-verified.
• claim-brain-inference-bound-like-ai-at-system-level — the synthesis / twist note, held at seedling with an [unverified-quant] flag preserved. This is the capture's whole reason for existing (per-event reversal nested inside system-level convergence), but its load-bearing number — Attwell & Laughlin's ~47%/34% grey-matter split — came from a search snippet, not the primary. I promoted the structural observation and flagged the number, rather than losing the bridge to the inbox.

Routed 1 question: question-verify-attwell-laughlin-2001-grey-matter-energy-budget — names exactly what to read (the Attwell & Laughlin 2001 budget table, cross-check Howarth/Gleeson/Attwell 2012) and what to extract. It gates the seedling note's promotion.

Retrieve-before-write: no collisions on the neuroenergetics substance — the vault had zero prior notes on Karbowski, van Rossum & Pache, or Attwell & Laughlin, so all three are new seeds. The semantic-index hint's top matches (Hopfield/Nobel, cheap Russian LEU, Werbos-Freud, cheap-gradient-bound) were embedding near-neighbors, not content overlaps. I checked the two highest-risk ones directly: claim-cheap-gradient-bound-two-figures is about the AD constant (1–4× / 5×), not brain energy; backpropagation-gap is the no-biological-analog thesis — adjacent, so I linked it and claim-brain-approximates-backprop-core-principles-ngrad rather than duplicating. Wired all three notes densely into the inference-economics and backprop clusters (claim-inference-dominant-ai-compute-2026, claim-ai-inference-means-running-a-model, claim-update-locking-backprop-constraint).

Folded / not promoted, and why:

• Seek's own framing thesis ("'training is expensive' is a silicon-specific accident") — interpretive, not a source-grounded claim; it lives in the commentary callouts of the promoted notes, not as its own file.
• Further leads (Epoch √-law provenance; Karbowski's cascade-model memory-lifetime energy tradeoff) — no source read yet; left as inbox leads.

What felt off / worth flagging to Cali:

• The system-level punchline is doubly-soft-sourced. The convergence claim leans on the Attwell 81% figure (unverified snippet) and connects to claim-inference-dominant-ai-compute-2026, which itself still carries an [unverified-quant] flag (audit V-002/V-011 — the two-thirds figure was not even found on its cited page). So the bridge's AI end and brain end both rest on flagged quant. I promoted it anyway — as a clearly-flagged seedling — because the structural insight is the capture's value and both weak points are honestly marked and routed. Worth a second look before this ever goes near a draft.
• No tier inflation and no duplicated research. The bee's two primary sources are real Tier 1 with real quotes, and it correctly self-flagged the Attwell number as unverified-quant rather than laundering it clean. Sourcing discipline working as intended.
• The bee's post-worthy: maybe self-assessment matches mine: clean quantified reversal, but "would want the Attwell & Laughlin primary before publishing." Agreed — that's exactly the gate the routed question represents.

Capture marked status: promoted with promoted_to / routed_questions / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: CIELAB ↔ non-Euclidean gradient-geometry bridge

I promoted 10-inbox/raw/2026-07-11-hop-cielab-noneuclidean-bridge.md (headless, unattended). This is a failed-bridge hop: the seed asked whether CIELAB's blue-hue non-uniformity shares a mechanism with the network hub-selection artifact (embedding cosine 0.75, unlinked). The answer was no — and the value is in the relocation, not the refutation. Clean, honest capture; my main judgment calls were about atomicity and one duplicate-adjacent note.

Winnowed 6 items → promoted 3, skipped 3 (one routed to a question, two left as inbox leads). New notes:

• observation-cielab-nonuniformity-bridges-gradient-geometry-not-sampling-artifact — the thesis. CIELAB's fix reweights the metric (keep coordinates, add a non-Euclidean distance); hub-selection's fix reweights the sample. Different failure families, so the embedding proximity is vocabulary, not mechanism. CIELAB's real kin is the Amari/GIFT non-Euclidean-metric cluster — a fourth, non-ML instance of the gradient-geometry pattern. I folded the "metric-failure vs sampling-failure" contrast into this note rather than splitting it out; it's the mechanism of the observation, not a separable claim. Typed it observation (matches the observation-richard-price... precedent). Held at seedling — it leans on the open one-object-or-three question.
• claim-cielab-only-approximately-uniform-fixed-by-noneuclidean-formulas — ProLab (arXiv:2012.07653, Tier 1, exact quote preserved). This is the note the earlier CIELAB pair was missing: a real colorimetry primary saying the space is "only approximately uniform" and the fix was explicitly "non-Euclidean colour difference formulas ... CIEDE2000." capture-verified (bee read it via extract_pdf; I did not re-fetch).
• claim-perceptual-color-space-not-riemannian-bujack-2022 — Bujack et al. (PNAS 2022, Tier 1, quote preserved). Perceptual colour isn't even Riemannian — diminishing returns forbid it. This is the load-bearing caveat on the bridge: Amari's natural gradient rests on a Riemannian Fisher metric, so CIELAB and natural gradient share the corrective move but not the same geometry. Evidence for "analogy, not one object." capture-verified.

Routed 1 new question, linked 2 existing:

• New: question-verify-cielab-ab-plane-gaussian-curvature — holds the Ahrens et al. (2024) nonzero-Gaussian-curvature / isometric-embedding-obstruction lead. The capture flagged it [unverified-quant/mechanism] and admitted only a search-engine summary was read, so it did not become a note. The question also parks the "does the non-Riemannian result have a gradient-geometry analogue?" research thread.
• Linked (not duplicated): question-gradient-geometry-one-object-or-three-analogies (from the observation and Bujack notes) and question-verify-cielab-perceptual-uniformity-and-blue-hue-nonuniformity (from the ProLab note).

Retrieve-before-write — one real collision, handled. The existing claim-cielab-lacks-perceptual-uniformity-in-blue-hues already asserts CIELAB non-uniformity + CIEDE2000-patches-the-metric, but on Tier-4 Wikipedia and specifically about blue hues. My ProLab note is a genuinely distinct angle: it supplies the Tier-1 primary that note was flagged as lacking, but only for the general non-uniformity — ProLab doesn't mention blue, so I was explicit about that boundary and left the blue-hue note's [unverified-claim] flag standing. Amari, GIFT, hub-selection, and the gradient-geometry question all pre-exist; I wired into them densely rather than restating. No duplication of the Bujack material anywhere in the vault (grep-checked).

Skipped / left in inbox, and why:

• CAM16-UCS as the current uniform-space standard, and Oaksford–Chater "optimal data selection" near the hub-selection neighborhood — both saved-but-not-followed hooks with no source read. Left as inbox leads, recorded in not_promoted.

What felt off / worth flagging to Cali:

• Nothing egregious — this was a well-behaved capture. Sourcing was honest: two real Tier-1 primaries with real quotes, and the one genuinely soft claim (Ahrens curvature) was correctly self-flagged as unverified rather than laundered. No tier inflation.
• Mild duplicated-research smell across the CIELAB thread. This is now the second CIELAB capture (the first was 2026-07-09, giving the design-intent + blue-hue pair). The vault's CIELAB cluster is growing sideways — three-to-four notes, all seedling, all in a vault otherwise centred on ML history. It's coherent (the gradient-geometry bridge is what earns colorimetry a place here), but if a third CIELAB capture lands it's probably time for a small colour-science MOC or a deliberate decision to stop feeding this seed. Not urgent.
• The bridge is real but hedged twice over, and I kept it that way: it's a structural analogy (not identity) and the analogy's ML end (Amari) is Riemannian while the colour end is provably not. I promoted the observation at seedling and made both hedges explicit in-note rather than smoothing them. The capture's own post-worthy: maybe matches my read — clean story, but it rests on one open question and one unread curvature lead.

Capture marked status: promoted with promoted_to / not_promoted / questions_routed lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: closed-world logic as a rival account of human reasoning

I promoted 10-inbox/raw/2026-07-11-hop-closed-world-human-reasoning.md (headless, unattended). A cross-time hop that starts from the seed's aside on "the open-world assumption problem," runs through Reiter's 1978 closed-world assumption, and lands on Sakama's 2024 realization of human conditional-reasoning "fallacies" in Answer Set Programming — giving the vault a second, non-Bayesian formal account of the same Wason/conditional-reasoning data it already documents under Oaksford & Chater. A genuinely good bridge, honestly sourced, with one clearly-marked soft spot.

Winnowed 3 distinct claims → promoted 2, folded 1. New notes:

• claim-conditional-reasoning-fallacies-as-closed-world-inference-in-asp — the mechanism. Sakama (arXiv:2311.04412, TPLP 2024, Tier 1) realizes affirming-the-consequent / denying-the-antecedent in ASP via program completion, where "Inference in ASP is deduction based on default logic (Reiter (1980))." I introduced the closed-world assumption (Reiter 1978, "what is not currently known to be true must be false") and negation-as-failure here as the engine, and folded Claim 1 (the AC/DA-are-pragmatic-not-broken psychology) in as the phenomenon being formalized. capture-verified — bee read it via TLS-verified extract_pdf; I did not re-fetch.
• claim-logic-programming-account-rivals-probabilistic-account-of-conditional-reasoning — the load-bearing bridge, and the reason this capture exists. The computational-logic lineage (Stenning & van Lambalgen 2008; Kowalski 2011; formalized by Sakama) set beside the probabilistic one (Oaksford & Chater). Held at seedling with an [unverified-claim] flag preserved, because the specific characterization — that Wason results are closed-world reasoning with negation-as-failure — rests on a semanticscholar paper page and Sakama's citation, not on the S&vL primary itself.

Routed 1 question: question-verify-stenning-van-lambalgen-closed-world-wason — names exactly what to read (the selection-task and closed-world/negation-as-failure chapters of Human Reasoning and Cognitive Science, MIT Press 2008; cross-check Kowalski 2011) and draws the line between what's Tier-1-confirmed via Sakama (the lineage exists) and what's unread (its content). It gates the bridge note's promotion off seedling.

Folded / not promoted, and why:

• Claim 1 (AC/DA as pragmatic inference via conditional perfection) — folded into the ASP note. It's background cognitive science Sakama restates, not his finding, and the conditional-perfection primary (Geis & Zwicky 1971) went unread. Not worth a standalone note on that footing.
• The CWA definition as its own note — folded as a Tier-4 definitional pointer inside the mechanism note; not atomic alone.
• Saved hooks (conditional perfection, assumption-based argumentation, Partial-Closed-World Assumption) — left as inbox leads, no source read.

Retrieve-before-write: no content collisions. The vault has zero prior notes on Reiter, the closed-world assumption, default/non-monotonic logic, ASP, Kowalski, or Stenning & van Lambalgen — this is a new concept cluster. The real neighbors are the existing Oaksford–Chater pair (claim-oaksford-chater-recast-wason-task-as-optimal-data-selection, claim-oaksford-chater-redefine-rationality-as-reasoning-about-uncertainty), which I wired both new notes into densely as the rival account, plus claim-inference-logical-types and claim-quantum-like-models-fail-conjunction-fallacy as a sibling alternative-formalism note. The semantic-index hint's top matches were mostly embedding near-neighbors; I checked the genuine ones and they're links, not duplications.

What felt off / worth flagging to Cali:

• One stale wikilink in the capture body. The capture links [[claim-oaksford-chater-wason-optimal-data-selection]], which does not exist — the real note is claim-oaksford-chater-recast-wason-task-as-optimal-data-selection. The bee reconstructed a plausible-but-wrong slug from memory rather than checking the vault. I used the correct slug in both notes. Minor, but the kind of thing that quietly rots the graph.
• The bridge is real but rests on an unread primary, and I kept that visible. The capture's own honest bottom line — "needs the S&vL primary read before it's publication-solid" — matches my read exactly. No tier inflation: Sakama is real Tier 1 with real quotes, and the soft link (the S&vL content) was already hedged in the capture rather than laundered. I promoted the bridge as a flagged seedling because the structural insight (a non-Bayesian formal account of the same data) is the capture's value and the one weak point is marked and routed.
• A pattern worth watching: this is now the second time the vault has caught the same Wason/conditional-reasoning data being "rescued" by a formalism (Bayesian, now closed-world logic). If a third alternative-formalism account lands, there's an MOC in it ("rival formalisms for the reasoning anomalies" / "the yardstick was mismeasured"). The Oaksford–Chater commentary callouts already gesture at the same idea. Not yet — noting it.

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: collation bridges cancel-slips and citation-variance

I promoted 10-inbox/raw/2026-07-11-hop-collation-bridges-cancel-slips-and-citation-variance.md (headless, unattended). This is a self-referential hop: it started from a bridge_candidate between two of my own notes — claim-early-modern-printers-corrected-print-with-cancel-slips and claim-pollack-1989-sterile-quote-page-231 — and asked whether they share a real bridge or just a theme. The answer it found is good: they're both instances of textual variance across physical copies, and the discipline that owns that problem is analytical bibliography, whose method is collation. A named discipline, not a pun.

Winnowed 6 distinct claims → promoted 3, folded 2, skipped 1 cluster of leads. New notes:

• observation-collation-bridges-cancel-slips-and-citation-variance — the bridge node itself (type observation, mirroring the watermark bridge note). Ties both seed notes to one discipline and carries the "collation is diff" road-home into claim-wikipedia-amari-sgd-citogenesis. ArchBook + CollateX, Tier 3; a framing, not a new external fact, so capture-verified.
• claim-stop-press-correction-produces-within-edition-variance — the mid-run correction mechanism (ArchBook, Univ. of Saskatchewan, Tier 3). Genuinely distinct from the existing cancel-slip note: stop-press alters the standing type for sheets not-yet-printed, so the earlier state survives in other copies; the cancel slip covers the error in the same object. Uncontested book-history, so Tier 3 clears the floor — kept seedling as a new note but not flagged.
• claim-hinman-collator-descends-from-blink-comparator-and-wwii-aerial-recon — the richest content and the one I flagged. The cross-domain origin (Pulfrich's 1904 blink comparator → Tombaugh/Pluto 1930 → Hinman's WWII aerial-recon → 1623 Folio) is surprising and load-bearing but sourced only to Wikipedia (Tier 4) + a Folger blog. Under sources.md's escalation rule that's below the floor, so: status: seedling, audit_status: flagged — [unverified-historical], and routed.

Routed 1 question: question-verify-hinman-collator-blink-comparator-origin — names what to read (Hinman's own 1963 Printing and Proof-Reading of the First Folio; the Folger's first-party collator materials; Lowell Observatory / Zeiss for the blink-comparator dates) and separates the two things needing confirmation — the origin story and the specific figures (1904 / 1930 / ~1946 / ~55 copies / 1963). Checked 50-questions/ first; there was no existing Hinman/collator/blink question.

Folded / not promoted, and why:

• Blink comparator as a standalone note — folded into the Hinman note as the instrument's ancestry. It's the same Tier-4 Wikipedia hop; a separate note would duplicate rather than add a distinct claim.
• CollateX / "collation is diff" as a standalone note — folded into the observation. The capture carried no verbatim quote and no precise URL for it, and a mechanism claim ("models variance as a variant graph") shouldn't stand alone on an uncaptured source. Stated lightly as the digital end of the discipline instead.
• Saved hooks (Tombaugh's biography; "states and issues" bibliography vocabulary; LARB "The Author Machine") — leads, not claims, no source read. Left in the inbox body.

Retrieve-before-write: no content collisions. The vault had zero prior notes on collation, analytical bibliography, the Hinman collator, or the blink comparator — a genuinely new instrument/discipline cluster. The real neighbors are the two seed notes (which already cross-link each other and this capture) and the print-history cluster (claim-bomberg-1520s-venice-talmud-set-standard-page-layout, observation-watermark-same-provenance-mechanism-paper-to-llm), which I wired the new notes into. The semantic-index hint's top matches (Mariel, Telnyx, cricket-laws, Lepsius, HNC/FICO, Rothschild) were embedding near-neighbors from other domains, not duplications — I checked and none warranted a link.

What felt off / worth flagging to Cali:

• Provenance-vs-source discipline was mostly good, but the tiering was optimistic in the frontmatter. The capture header stamps source_tier: 3 and source_url = the Hinman collator Wikipedia page, while the body honestly labels that same Hinman material "Tier 4." The load-bearing cross-domain claim is Tier 4, full stop — I demoted the Hinman note's source_tier to 4 and flagged it rather than inheriting the header's cleaner-looking 3. Mild tier inflation at the capture's top level; the body's per-claim tiers were the honest ones.
• This hop is partly re-grounding my own vault rather than reading the world. Two of the four hops (the seed read and the "road home" to citogenesis) are internal; only ArchBook, the Hinman/blink Wikipedia pages, and CollateX are new external reading, and the external sources are all Tier 3–4. That's fine for a bridge note — the value here is the connective structure, not new primary facts — but it means the whole cluster is currently soft-sourced and none of it should go evergreen until the Hinman origin question is answered against Hinman 1963.
• A pattern worth watching: this is the second print-history bridge in a week to resolve into "correct-by-comparison / correct-by-surrounding" (cancel slips, Bomberg's Talmud layout, errata lists, now stop-press + collation). The correction/accretion cluster in moc-schema-change-and-restructuring is getting dense enough that a dedicated print-history-of-correction MOC may soon be warranted. Not yet — noting it.

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: chain-of-thought is the anti-EUV (codified vs tacit moat)

I promoted 10-inbox/raw/2026-07-11-hop-cot-not-a-tacit-moat.md (headless, unattended). A synthesis hop off the test-time-compute seed: OpenAI hid o1's raw chain-of-thought partly for "competitive advantage" → s1 distilled a rival for near-nothing → and the payoff bridges to my own EUV note as a named asymmetry — tacit-and-inimitable (EUV mirrors) vs codified-and-copyable (a reasoning trace is literally text). A well-built capture with one strong Tier-1 leg and two softer ones, all honestly labelled.

Winnowed 3 distinct claims → promoted 3, skipped 2 leads. New notes:

• claim-openai-hid-o1-raw-chain-of-thought-partly-for-competitive-advantage — the o1-secrecy claim. Grounding quotes are relayed through Simon Willison (Tier 2) because the OpenAI primary 403s on direct fetch; the safety/UX clauses are verbatim, but "competitive advantage" is partly an observer reading. Flagged, seedling, routed.
• claim-s1-distilled-reasoning-from-1000-traces-in-26-minutes — the load-bearing empirical leg. Muennighoff et al. (arXiv:2501.19393, Tier 1), exact quote "requiring just 26 minutes of training on 16 H100 GPUs," +27% over o1-preview on competition math. The core figures clear the floor; only the viral "<$50" headline is soft (TechCrunch, Tier 3), which I carried under the capture's [unverified-quant] flag and routed. seedling.
• claim-a-chain-of-thought-trace-is-codified-so-it-cannot-form-a-tacit-moat — the synthesis / bridge, and the reason this capture exists. Names the codified-vs-tacit asymmetry, wires to claim-euv-mirror-advantage-is-tacit-know-how-not-patent, the Polanyi codification bridge, and claim-canonic-deliverable-is-an-append-only-evidence-ledger. Held at seedling because its tacit pole rests on the EUV note, which is itself flagged/seedling.

Routed 2 questions (checked 50-questions/ first — neither existed):

• question-verify-openai-o1-cot-competitive-advantage-language — read the OpenAI primary directly (via Wayback if it keeps 403-ing) and confirm the exact "competitive advantage" wording, since that clause is the hinge of the whole asymmetry.
• question-verify-s1-under-50-dollars-compute-cost — find a primary cost figure (s1 paper / GitHub repo) or reproduce the 16×H100 × 26-min arithmetic, so the "<$50" meme cites a derivation rather than a headline.

Skipped / not promoted, and why:

• The DeepSeek-R1 open-CoT lead ("the moat OpenAI feared, realized?") — a genuinely interesting research hook, but a hook, not a verification gap. Left in the capture's audit trail rather than filed.
• "Is 'hide the trace' a stable equilibrium, or does answer-only output still leak enough to distill?" — folded as an explicit caveat inside the synthesis note rather than promoted separately; it's the honest hedge on the "fully codified" claim.
• Budget-forcing / appending "Wait" to lengthen s1's thinking — the bee itself set this aside as too close to the seed's test-time-compute topic; already covered by claim-extended-thinking-as-serial-inference-compute. Not promoted.

Retrieve-before-write: no content collisions. The vault had zero prior notes on s1, distillation-as-moat, or the codified/tacit reasoning asymmetry. The genuine neighbors are the ones I wired into densely — claim-euv-mirror-advantage-is-tacit-know-how-not-patent (the tacit pole), claim-extended-thinking-as-serial-inference-compute and claim-test-time-compute-can-substitute-parameters (the mechanism), cot-faithfulness-anthropic-biology (the is-the-trace-even-faithful angle), and claim-canonic-deliverable-is-an-append-only-evidence-ledger (codification from the governance side). The semantic-index hint's remaining top matches (Spinoza's grinding dish, mekh-mat, William Bell, yellow rain) were embedding near-neighbors from other domains, not duplications.

What felt off / worth flagging to Cali:

• The bridge is elegant but leans on my own weakest note. Its tacit pole is claim-euv-mirror-advantage-is-tacit-know-how-not-patent, which is a flagged Tier-4 seedling resting on a single industry-explainer. So the whole asymmetry inherits that softness on one end and a Tier-3 dollar-figure on the other. I promoted the synthesis anyway — as a clearly-flagged seedling — because the structural insight (codified-copyable vs tacit-inimitable) is the capture's value and both weak points are marked and routed. Nowhere near draft-ready.
• A slight over-reach in the capture's core claim, and I kept it visible. "A trace is text, therefore the moat is fully codified" is the seductive step — text can still under-specify the expensive data-curation and verification work behind it. The capture's own commentary half-admits this; I made it an explicit in-note caveat rather than smoothing it, and that's exactly what the s1 "<$50 counts what?" question probes.
• No tier inflation, no laundering. The bee correctly stamped s1 as Tier 1 with a real quote, correctly demoted the "<$50" figure to Tier 3 with an [unverified-quant] flag, and was honest that the OpenAI quotes came via a Tier-2 relay because the primary 403s. Sourcing discipline working as intended. One mild note: source_tier: 1 in the capture header reflects the s1 paper (the strongest leg), while the o1 leg is really Tier 2 — I set each note's tier per its own source rather than inheriting the header's cleaner 1.
• This is the third capture this week to hinge on Polanyi/codification (EUV tacit-moat, CANONIC's uncodifiable-verdict ledger, now the reasoning-trace asymmetry). The codification thread is getting dense enough that a small MOC may soon be warranted. Not yet — noting it.

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop capture: two senses of "credit assignment" (headless)

Promoted 10-inbox/raw/2026-07-11-hop-credit-assignment-two-senses.md, the hop that started as the unlinked Dreyfus-1973 ↔ Chandler bridge and climbed through Stigler's Law → the Matthew Effect → "credit assignment" → Ogburn–Thomas ripeness. Winnowed it hard: three real claims plus a pile of leads, and I kept two notes.

Notes created (both seedlings):

• claim-credit-assignment-spans-sociology-and-backprop — the keystone (Claim 2, Tier 1). "Credit assignment" is one phrase for two things: Merton's sociology of who-gets-the-credit and Minsky's 1961 technical problem that backprop addresses. Schmidhuber sits on the hinge — he opens his Annotated History with "Machine learning (ML) is the science of credit assignment" (technical) while the whole document is a credit-reassignment crusade (sociological). I wired it into the existing credit-assignment cluster: claim-minsky-1961-named-credit-assignment, myth-structural-credit-assignment-from-minsky-1961, claim-dreyfus-1973-lineage-link-uncorroborated, claim-werbos-1974-no-credit-assignment-language, moc-backpropagation-origins.
• claim-credit-detectors-are-themselves-misattributed — Claim 3 (Tier 4). Stigler's Law (Stigler credited it to Merton) and the Matthew Effect (Merton co-developed it with Zuckerman) are both instances of the misattribution they name. Folded the two facts into one bounded note rather than fragmenting thin Tier-4 material into two thin ones.

Questions routed:

• question-verify-schmidhuber-credit-assignment-opening-and-matthew-effect — re-fetch arXiv:2212.11279 (open access), confirm the opening sentence verbatim and check whether Schmidhuber ever names the Matthew Effect. If he does, the bridge is provably his framing and the keystone can leave seedling.
• question-verify-stiglers-law-and-matthew-effect-primaries — read Stigler's 1980 festschrift essay and Merton's 1968 Science paper + his Zuckerman crediting. Both facts are surprising/self-referential historical claims, which the sourcing floor says must escalate off Tier-4 Wikipedia.

Skipped, and why (logged in the capture's not_promoted):

• Claim 1 as a standalone note — the transmissionist "descent" reading vs. the Ogburn–Thomas/Merton "in the air" reading. This is already promoted: claim-chandler-simultaneous-discoveries-are-incremental-repackagings literally calls the thesis "a direct rival to the multiple-discovery frame," and claim-dreyfus-1973-lineage-link-uncorroborated enacts it. A standalone note would have been a straight duplicate. Its one genuinely new element — Schmidhuber's backprop genealogy is that same descent-building move — I folded into the keystone's third paragraph instead.
• The "Stigler's Law as hub for the eponymy cluster" MOC lead — premature. The cluster (Max/MSP, Helmholtz Machine, Amari citogenesis, Nakamoto/Wei-Dai) is under the spec's 5-note threshold for an MOC. Left as an inbox lead.
• The Nakamoto/Wei-Dai and Lemaître/Hubble saved hooks — leads, not atomic claims; Nakamoto priority is already tracked at question-verify-nakamoto-dai-bmoney-szabo-precursor.

Retrieve-before-write: no content collisions. The one Grep hit on "Stigler" was claim-schema-change-four-mode-taxonomy, which cites James W. Stigler the psychologist — a different person, not Stephen Stigler of the eponymy law. No existing note covers Stigler's Law, the Matthew Effect, or the two-senses bridge. Worth noting the reverse direction: claim-dreyfus-1973-lineage-link-uncorroborated and claim-chandler-simultaneous-discoveries-are-incremental-repackagings already forward-link to this capture's slug [[2026-07-11-hop-credit-assignment-two-senses]] — those point at the capture, not at either new claim-note. I left them alone (editing them is outside a promotion's scope), but a future tidy pass should repoint them at the keystone note.

What felt off / worth flagging to Cali:

• Tier inflation in the header. The capture's frontmatter carries source_tier: 1, but only Claim 2 (Schmidhuber) is Tier 1. Claims 1 and 3 rest entirely on Wikipedia (Tier 4) — the bee even labels them Tier 4 in the body's Grounding block, so the clean 1 in the header oversells the capture as a whole. I set each note's tier from its own source, not the header.
• Self-exemplification is charming and load-bearing at once. The whole appeal of the second note is the recursive joke ("Stigler's Law doesn't spare the discoverers of Stigler's Law"), and it would be easy to let the elegance stand in for verification. Both instances are cleanly documented in secondary sources but neither primary was read in this chain, so I promoted it as a flagged seedling and routed the primaries. Not draft-ready.
• Duplicated research risk is real for this batch. The Merton/Chandler/Dreyfus/multiple-discovery territory is now well-trodden in the vault; this is at least the third capture to circle the sociology-of-credit frame. The genuinely new payload was narrow — the terminological bridge (one word, two domains, lands on AI) — which is why I came away with two notes rather than five. The frame the capture claims to "name for the first time" was, in substance, already implicit across the existing Chandler and Dreyfus cross-links.

Capture marked status: promoted (removed the stale status: seedling) with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: "deep learning" rebrand as a field-scale Fresh Start Effect

I promoted 10-inbox/raw/2026-07-11-hop-deep-learning-rebrand-fresh-start.md (headless, unattended). This is a failed-then-relocated bridge hop: the seed asked whether two of my own notes genuinely bridge — the Fresh Start Effect (claim-fresh-start-effect-requires-subjective-new-beginning-framing) and Lighthill's combinatorial-explosion diagnosis (claim-lighthill-1973-blamed-combinatorial-explosion), cosine 0.75, unlinked. The bee's honest answer: the direct link is only rhetorical ("the single-cause story is wrong"), and the real bridge had to be built through a third AI-native instance — the neural-net→"deep learning" rebrand. Good, self-aware capture; my judgment calls were mostly about how hard to flag the one unsourced synthesis.

Winnowed 3 distinct claims → promoted 3 (two seedlings with live flags), skipped a reviewer anecdote + 2 saved hooks. New notes:

• claim-neural-networks-backwater-before-deep-learning-rebrand — the load-bearing, best-sourced claim: neural nets were an out-of-favor "backwater" in the 1990s–2000s, beaten by SVMs. Timothy B. Lee, Understanding AI (2024, Tier 2), exact quote "neural networks had become a backwater" preserved. capture-verified (bee WebFetched it; queen re-check blocked headless). I made a point of separating this from the first winter — the 1970s–80s stall the vault already tracks (Lighthill / SRI / Perceptrons / Nestor) had a technical core (no hidden-layer training) plus funding shocks; this second low was almost pure fashion defeat by SVMs with backprop already working. Two winters, two causes.
• claim-deep-learning-term-predates-hinton — the term-origin correction: "deep learning" was introduced by Rina Dechter (1986) and applied to ANNs by Aizenberg et al. (2000); Hinton popularized and re-pointed it c. 2006, didn't coin it. Rests on Tier-4 Wikipedia, so seedling + [unverified-historical], routed. I was careful to keep this distinct from claim-ivakhnenko-gmdh-first-deep-characterization, which is about priority for the artifact (depth) — this is priority for the word. Same "first at what, exactly?" disaggregation.
• observation-deep-learning-rebrand-as-field-scale-fresh-start — the bridge, and the reason the capture exists. Typed observation (matching my CIELAB / collation bridge-note precedent), seedling + [unverified-claim]. The two underlying facts are sourced (stigma Tier 2, term-origin Tier 4); the load-bearing move — that fresh-start psychology scales from individuals to a whole field — is my analogy, which the capture itself flags as "my analogy, not yet a sourced finding." I promoted it because losing the capture's whole thesis to the inbox would be the wrong kind of conservative, but I kept the hedge loud: it's "the rebrand resembles a Fresh Start Effect," not "the field performed one."

Routed 2 questions (checked 50-questions/ first — neither existed):

• question-verify-deep-learning-term-origin-dechter-aizenberg — names the three primaries to read (Dechter's AAAI 1986 paper, the Aizenberg et al. 2000 Kluwer book, Hinton's 2006 Neural Computation / Science papers) to move Claim 2 off Wikipedia. Gates the term-origin note.
• question-fresh-start-effect-extends-to-collective-organizational-scale — the crux gate for the bridge: does the Dai/Milkman/Riis temporal-landmark literature extend from individuals to collective/organizational fresh starts, and has anyone made this field-rebranding analogy before? If yes, the observation hardens into a citation; if no, it stays an explicitly-labeled metaphor.

Skipped / left in inbox, and why (logged in not_promoted):

• The recycled reviewer anecdote "neural networks had their day in the 1980s" — the capture itself admits it's only on Medium-tier (Tier 5) sources with the primary unfound. Below the floor; wants a Genius Makers page check first. Left as a lead, mentioned as a flagged aside in the backwater note rather than asserted.
• Two saved-but-not-followed hooks (the generic "monocausal narrative" meta-abstraction; the strain-model-vs-boundary-model tension from the Lighthill note) — orthogonal leads, no source read. Left in the inbox body.

Retrieve-before-write: no content collisions. I checked the semantic-index hint's top matches and grep-swept for backwater / deep learning / Dechter / Aizenberg / SVM / rebrand / fresh start. The vault's AI-winter cluster is about the 1970s–80s (SRI, Nestor, Lighthill, Perceptrons) — a different period from this capture's 1990s–2000s SVM-era stigma, so no overlap. The Ivakhnenko "first deep learning" note is the depth-priority axis, not the term-origin axis — I linked it as the sibling-but-distinct claim rather than duplicating. Wired all three new notes densely into the backprop-origins, credit-drift, and schema-change/restructuring clusters, and confirmed every wikilink target resolves (MOCs in 40-mocs/, questions in 50-questions/).

What felt off / worth flagging to Cali:

• The capture's frontmatter tiering is honest but its header source_tier (2) oversells the whole. Only Claim 1 is genuinely Tier 2. Claim 2 is Tier-4 Wikipedia; Claim 3 is unsourced synthesis. The body labels each claim's tier honestly ("Tier 4, uncontested historical / pointer"), so this isn't laundering — but as with several captures this week, the clean top-level tier flatters the batch. I set each note's tier from its own source.
• This is a self-referential hop — two of the four "hops" are re-reading my own notes. The genuinely new external reading is thin: the Lee article (Tier 2) and the Wikipedia term-origin sentence (Tier 4). That's fine for a bridge note, where the value is connective structure, not new primary facts — but it means the whole cluster is soft-sourced and none of it should go near a draft until both routed questions land.
• The bridge is the capture's most seductive and least grounded part, and I kept that visible rather than smoothing it. The bee's own post-worthy: maybe matches my read exactly: a clean "the obvious bridge is fake but a better one hides behind it" story that lands on AI history, but the field-scale Fresh Start mapping is an analogy, not a finding. Promoted as a flagged seedling; the collective-fresh-start question is exactly the gate.
• No duplicated research this time (unlike the credit-assignment / CIELAB threads earlier today) — this genuinely opens new ground (SVM-era stigma, term-origin coinage) the vault hadn't covered, while connecting to clusters it had.

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: deep nets have critical learning periods like Hubel–Wiesel's kittens

I promoted 10-inbox/raw/2026-07-11-hop-deep-net-critical-periods.md (headless, unattended). A cross-domain, cross-time hop off the Neocognitron seed: Fukushima's S-cells/C-cells descend from Hubel & Wiesel's cat-cortex cells → the monocular-deprivation critical period → Achille, Rovere & Soatto's 2019 result that deep networks show the same onset/length-dependent critical periods → and a deflationary twist (no molecules required). A well-built capture with one genuinely strong Tier-1 primary carrying two of the three claims.

Winnowed 4 distinct claims → promoted 3 (all seedlings), skipped 1 + a pile of saved hooks. New notes:

• claim-deep-nets-have-critical-learning-periods-timed-like-animals — the phenomenon. Achille, Rovere & Soatto (ICLR 2019, arXiv:1711.08856, Tier 1), full onset/length quote preserved. A temporary early input deficit permanently caps final skill; damage scales with when/how long, not total exposure. capture-verified.
• claim-monocular-deprivation-permanently-rewires-visual-cortex — the biological anchor. Hubel & Wiesel's sutured-eye result; the deficit lives in cortical wiring, not the healthy eye; the reason pediatric cataract surgery is done within weeks. Sourced to a PMC review with quote "Ocular dominance is established irreversibly early in childhood." capture-verified.
• claim-critical-periods-arise-from-information-plasticity-not-biology — the distinct mechanistic/deflationary payload. Deficit type (blur, not vertical-flip) decides permanence; Achille measures a rise-then-fall of Fisher Information he calls a loss of "Information Plasticity" (Tier-1 quote preserved); the net has no pruning or neuromodulators, so the critical period is a property of learning dynamics, not biological hardware. Wired into backpropagation-gap as a productive tension — a non-biological learning mechanism reproducing a biological phenomenon. capture-verified.

Routed 2 questions (checked 50-questions/ first — neither existed):

• question-information-bottleneck-linked-to-critical-periods — holds the skipped Tishby/IB claim: the "learning is forgetting" mechanism is sourced only to a Quanta feature (Tier 3) and is contested (Saxe 2018 showed compression isn't universal). Names the primaries to read (Shwartz-Ziv & Tishby 2017; Tishby & Zaslavsky 2015; Saxe et al. ICLR 2018) before any IB↔critical-periods note can exist.
• question-deep-net-critical-period-predicts-human-amblyopia-timing — the loop-closure candidate, which also parks the capture's own "needs one more primary source before publishing" gap: the Achille result is validated against animal models, so the human link (amblyopia timing) is asserted by analogy, not demonstrated.

Skipped / folded, and why (logged in not_promoted):

• The Tishby Information Bottleneck + Saxe-reversal thread (Hop 4) — a contested technical-mechanism claim resting on Tier-3 Quanta only. Below the sourcing floor for a mechanism claim; it's a further-lead the capture itself flags with "does the vault link IB to critical periods?". Routed, not promoted. I did fold a one-line mention of it into the Information-Plasticity note (Achille frames his result as the IB view) with the open question wikilinked.
• The human ↔ deep-net link as a standalone claim — the animal link is what Achille verifies; I promoted that and kept the human extension as an explicit in-note caveat + routed question rather than asserting it.
• Saved hooks not followed (Stephen Kuffler; the Golden Goose Award; lottery-ticket/early-phase determinism) — leads, no primary read; left in the inbox body.

Retrieve-before-write: no content collisions. The vault had zero prior notes on critical periods, Hubel & Wiesel, monocular deprivation, information plasticity, or the Information Bottleneck — a genuinely new cluster. The semantic-index hint's top matches were embedding near-neighbors from other domains (leap seconds, punctuated equilibrium, dense associative memory, Byzantine trade); I checked the AI-adjacent ones and they're links, not duplicates. The real neighbors I wired into densely: the seed claim-fukushima-1979-neocognitron-first-cnn (S-cells/C-cells ← Hubel–Wiesel), claim-vanishing-gradient-chain-rule-pathology (the "early layers stop changing" rhyme, flagged as distinct), and backpropagation-gap (representation-matches-mechanism-doesn't — the sharpest dialogue partner for the "no molecules required" note).

What felt off / worth flagging to Cali:

• One instance of tier inflation, corrected. The capture labels the PMC review "Tier 1/2." A secondary review is Tier 2; the Tier-1 object is Hubel & Wiesel's own 1960s papers, which the note points at but did not re-read. I set the Hubel–Wiesel note to Tier 2 and said so in a commentary callout. The claim is Nobel-level uncontested neuroscience, so Tier 2 clears the floor comfortably — this is a labeling fix, not a soundness problem.
• The header source_tier (1) reflects only the Achille leg. Two of the three notes genuinely rest on that Tier-1 primary with real quotes; the biological-anchor note is Tier 2. I set each note's tier from its own source rather than inheriting the clean header 1.
• No duplicated research — unlike some of this week's self-referential hops, this one opens genuinely new ground (developmental neuroscience, critical periods, information plasticity) while connecting cleanly to the existing CNN-origins and backprop-gap clusters. The bee's own post-worthy: maybe matches my read: a clean 1960s-cats-to-2019-nets bridge with a real deflationary twist, gated on the human-link primary that the routed question now tracks.
• The Information-Plasticity note is the interesting one to watch. Its deflationary claim ("a plain optimizer reproduces a critical period with no biological machinery") sits in genuine tension with backpropagation-gap's "brains and nets don't share a learning mechanism." Not a contradiction — one is about mechanism difference, the other about a shared phenomenon — but the pair is exactly the kind of thing a future reflection or MOC on biological-plausibility could hang on.

Capture marked status: promoted (removed the stale status: seedling) with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: diplomatics of priority (Merton the hinge to Mabillon)

I promoted 10-inbox/raw/2026-07-11-hop-diplomatics-of-priority.md (headless, unattended). The seed asked whether the Werbos-backprop-priority myth and Mabillon's founding of diplomatics are genuinely connected or only superficially resemble each other. The bee's answer — and the good one — is that the bridge is methodological, not a documented descent: both acts authenticate an origin-claim by comparing dated primary sources instead of trusting the received attribution, and Robert Merton is the single figure who sits on both sides (he theorized discovery-as-multiple, and he practiced Mabillon-style source criticism on a misattributed aphorism in On the Shoulders of Giants). A genuinely tidy hop, honestly hedged by the bee itself.

Winnowed 5 distinct claims → promoted 3, skipped 2 (one collision, one folded). New notes (all seedlings):

• observation-priority-adjudication-is-documentary-source-criticism — the flagship synthesis. Typed observation (matching my CIELAB/collation/deep-learning bridge-note precedent). It states the priority=authentication identity and names Merton as the hinge. Held at seedling with [unverified-mechanism] because the identity is Seek's own structural reading and the Mabillon→Quellenkritik→priority descent is an analogy, not a documented lineage. I folded the OTSOG→William-of-Conches (1123) detail in here as the supporting evidence for "Merton practiced the method," rather than giving it a thin Tier-4 note of its own. This note is also the first-class node the two existing stubs point at (see below).
• claim-merton-multiple-discovery-is-sciences-dominant-pattern — the load-bearing generalization the vault kept implying without ever stating atomically. Merton: "the pattern of independent multiple discoveries in science is in principle the dominant pattern rather than a subsidiary one," so singletons are the anomaly. Rests on a Tier-3 blog quoting Merton, so seedling + flagged, routed to the primary. The reverse-mode note's own commentary literally calls Merton's mechanism "a queued question rather than a stray observation" — now it has its own node.
• claim-multiple-discovery-concept-was-itself-multiply-discovered — the reflexive claim: the multiple-discovery idea had 18+ independent predecessors (Macaulay 1828, Galton 1869, Ogburn & Thomas 1922, Duhem), so it instantiates its own content. Seedling + [unverified-quant] on the count/dates. Distinct from — and tightly linked to — claim-credit-detectors-are-themselves-misattributed (Stigler's Law / Matthew Effect being themselves misattributed): same self-referential pattern, different specific instance.

Routed 2 questions (checked 50-questions/ first — neither existed):

• question-verify-merton-multiples-dominant-pattern-and-predecessors — read Merton's own "Singletons and Multiples in Scientific Discovery" (Proc. Am. Phil. Soc. 105, 1961; reprinted in The Sociology of Science, 1973) to confirm both the dominant-pattern quote and the predecessor enumeration/count. Gates the two Merton notes at once.
• question-verify-diplomatics-to-priority-adjudication-descent — read Duranti's "Diplomatics: New Uses for an Old Science" (Archivaria) and a history of Quellenkritik/Ranke to decide whether a real lineage runs from Mabillon to modern priority scholarship, or whether the connection is only a recurrence of method. Gates the observation off seedling.

Skipped / folded, and why (logged in not_promoted):

• Mabillon founded documentary source criticism / diplomatics — a genuine collision with the already-promoted claim-mabillon-1681-founded-diplomatics-to-refute-forgery-charge (from the 2026-07-09 half-proof capture). I linked it densely from the new observation instead of duplicating.
• OTSOG→William of Conches (1123) — folded into the observation as evidence; Tier-4 Wikipedia only, too thin to stand alone.
• Saved leads (Duranti's digital-forensics reuse; the Merton-vs-Chandler live tension; the Hirsch/Dias forensic-authentication neighbor) — leads, not claims; the Duranti lead now lives inside the routed descent question, and the Hirsch/Dias thread is already tracked at question-verify-hirsch-dias-superconductivity-retraction-primaries.

Retrieve-before-write — the notable finding of this run. No content collisions among my three new notes, but the surrounding graph is much denser than the capture's author realized. Two existing notes — claim-mabillon-1681-founded-diplomatics-to-refute-forgery-charge and myth-werbos-invented-backpropagation — already carry appended "Cross-domain bridge (2026-07-11)" sections naming Merton as the hinge and pointing at this exact capture slug. So the bridge idea was already stubbed into two notes with no first-class home; my observation note is now that home. I wired all three new notes into the sociology-of-credit cluster (claim-reverse-mode-multiple-independent-discovery, claim-chandler-simultaneous-discoveries-are-incremental-repackagings, claim-credit-assignment-spans-sociology-and-backprop, claim-credit-detectors-are-themselves-misattributed, claim-linnainmaa-priority-not-paternity). As with the credit-assignment run, those two existing stub sections still point at the capture slug, not the new note — a future tidy pass should repoint them at the observation; editing them is outside a promotion's scope.

What felt off / worth flagging to Cali:

• Duplicated research is the real story of this batch. This is at least the fourth capture in two days to circle the same Merton/Chandler/multiple-discovery/sociology-of-credit territory (the credit-assignment-two-senses capture earlier today, plus the Chandler and Dreyfus captures from 2026-07-09). The genuinely new payload here was narrow but real: (1) the Mabillon/diplomatics forensic-authentication framing of priority, which the credit-assignment thread never made, and (2) a first-class atomic node for Merton's dominant-pattern thesis, which the vault had been gesturing at from at least three notes without ever stating. That's why I came away with three notes rather than folding the whole thing into the existing cluster.
• No tier inflation — the capture was honest about its own softness. Header source_tier: 3 (measureformeasure.co) is fair, the Wikipedia OTSOG material is correctly labeled Tier 4 in the body, and the bee's own post-worthy: maybe explicitly says the Mabillon→historical-method→priority descent "is asserted structurally, not yet documented, so it needs the Duranti/Quellenkritik lead before publication." I agree entirely and kept that hedge loud in the observation note rather than smoothing it into a clean lineage claim.
• MOC watch. The sociology-of-scientific-credit cluster is now well past the spec's 5-note threshold (Merton-multiples, multiple-discovery-self-instancing, credit-detectors, Chandler, reverse-mode, credit-assignment-spans, Linnainmaa priority/paternity, Werbos myth, Mabillon, and now the priority=source-criticism observation). There's already question-myth-correction-as-historiography-moc-candidate parked nearby. I did not build an MOC (out of a promotion's scope, and the cluster spans two candidate themes — "sociology of credit" vs "myth-correction-as-historiography" — that want disentangling first), but it's ripe. Flagging for a deliberate decision.

Capture marked status: promoted with promoted_to / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: the Hawks-2007 / Jeffress-1948 cosine pairing is an embedding false friend

I promoted 10-inbox/raw/2026-07-11-hop-embedding-false-friend-hawks-jeffress.md (headless, unattended). This is the most self-referential hop I've promoted this week: the seed was a proximity in my own retrieval index — the notes claim-hawks-2007-human-adaptive-evolution-accelerated-recently and claim-jeffress-1948-place-theory-waited-decades-for-anatomical-confirmation flagged as unlinked neighbours at cosine 0.75 — and the whole chain is an adjudication of whether that neighbour pair is real. The bee's answer, and the right one, is no: shared rhetorical scaffolding (dated claim + big temporal quantifier + vindication/contestation arc + history-of-science vocabulary), not a shared phenomenon. The tool's own output became the object of study. A clean, honest capture; my judgment calls were mostly about atomicity and one weak source.

Winnowed 4 distinct claims → promoted 4, skipped 3 leads + 1 fold. New notes:

• observation-hawks-jeffress-cosine-pairing-is-embedding-false-friend — the central diagnosis, typed observation (matching my CIELAB / collation / deep-learning / diplomatics bridge-note precedent). Two independent tests dissolve the link: (1) the only unifying frame, delayed vindication / Stent's prematurity, has a scope condition that fits Jeffress and excludes the still-contested Hawks — and a vault_bridge probe of that frame surfaced Jeffress + Mayr but not Hawks; (2) the residual similarity is the anisotropy/arbitrary-cosine failure mode. [unverified-synthesis], seedling — a false-friend verdict is a negative claim ("no bridge"), which is hard to source, so I kept the hedge loud. No wikilink added between the two seed notes.
• claim-cosine-similarity-of-embeddings-can-be-arbitrary — the reusable general claim, and the mechanism under the whole capture. Steck, Ekanadham & Kallus (WWW 2024, Tier 1, exact quote "cosine-similarity can yield arbitrary and therefore meaningless similarities'"). Anisotropy (Ethayarajh 2019) folded in as a clearly-labelled secondary strand. capture-verified`.
• claim-sleeping-beauties-delayed-recognition-is-a-continuum — the "surprise finding worth its own note" the bee flagged. Ke, Ferrara, Radicchi & Flammini (PNAS 2015, Tier 1, quote "there are no clear demarcation values that allow us to separate SBs from 'normal' papers"). The "beauty coefficient" B shows delayed recognition is a smooth tail, not a natural kind — a deflationary tool against my own delayed-vindication cluster. capture-verified.
• claim-stent-prematurity-covers-neglected-not-contested-ideas — the concept that does the discriminating work. The weakest-sourced note in the batch: Stent's 1972 Sci. Am. "prematurity" definition and its load-bearing scope condition rest on a Tier-3 blog summary of Stent, not his own words. Promoted at seedling with [unverified-quote] because the concept is a canonical HoS tool that recurs across the vault — but flagged and routed.

Routed 1 question: question-verify-stent-1972-prematurity-primary-text — names exactly what to read (Stent's primary 1972 Sci. Am. article, or his Advances in the Biosciences version) and what to confirm: the exact definition wording and, crucially, the scope condition (premature = correct-but-neglected, not contested-unresolved) that the false-friend observation leans on to exclude Hawks. Checked 50-questions/ first — no existing prematurity/Stent/cosine/sleeping-beauties question.

Skipped / folded, and why (logged in not_promoted):

• The "real bridge nearby" lead — the genuinely unlinked delayed-vindication-meta-note ↔ Mayr-1954-peripatric pair the bee flagged as a better bridge than the seed's. A vault-linking suggestion, not an atomic external claim; left in the capture body as a lead for a future linking pass (and it's a good one — worth doing).
• The Jeffress delay-lines vs. 1940s mercury delay-line computer-memory hook — an explicitly-not-followed cross-time computation bridge, no source read.
• The Ethayarajh anisotropy point as its own note — folded into the cosine note, since its URL/quote weren't captured and the note stands fine on Steck.
• The "post-worthy: maybe / tool-studying-its-own-output" self-assessment — editorial, lives in the observation note's commentary.

Retrieve-before-write: no collisions. The vault has zero prior notes on cosine-similarity arbitrariness, Sleeping Beauties / the beauty coefficient, or Stent's prematurity — three genuinely new seeds. I grep-swept for cosine / anisotropy / sleeping-beauty / prematurity / Steck / Ethayarajh and confirmed. The real neighbours I wired into densely: the two seed notes; the anisotropy sibling claim-gift-2026-gradient-anisotropy-isotropic-transform (same phenomenon, different layer of the stack); the retrieval/embedding cluster (claim-matryoshka-representation-learning-truncatable-embeddings, claim-contextual-retrieval-failure-rate-figures); and the delayed-vindication cluster (claim-neural-circuit-theories-wait-decades-for-anatomical-confirmation, claim-lamport-paxos-greek-allegory-delayed-publication, claim-stein-delayed-admissibility-proof-five-years-to-avoid-bayesian-argument, claim-dreyfus-1988-connectionism-was-vindicated-not-target). The semantic-index hint's remaining matches (Helmholtz machine, OpenAI embedding price, Irino-2026, ACH/Admiralty-code, yellow rain) were embedding near-neighbours from other domains — checked, not duplicates.

What felt off / worth flagging to Cali:

• The header tiering flatters the batch, as usual this week. source_tier_1: 1 (Steck) and source_tier_2: 1 (Ke) are honest and carry real quotes; but source_tier_3: 3 is the Stent leg, and that Tier-3 source is a summary of a primary the bee never read — so the note resting on it (prematurity) is the one genuine sourcing gap. I set each note's tier from its own source and flagged + routed the Stent one rather than letting a paraphrase-of-a-paraphrase look clean.
• Mild duplicated-research smell on the delayed-vindication frame. This is now the second time this week a capture has leaned on the vault's delayed-vindication cluster (the neural-circuit observation note, claim-neural-circuit-theories-wait-decades-for-anatomical-confirmation, came out of a 2026-07-09 capture). Here it's put to genuinely new use — as a discriminator that excludes a note rather than a unifier — and it brought two new external anchors (Stent, Ke et al.) the vault didn't have, so I don't think the cluster is going sideways yet. But "delayed vindication" is becoming a heavily-trafficked frame; if a third capture rides it, a small MOC (delayed-vindication / prematurity / sleeping-beauties) is probably warranted, and the Sleeping-Beauties continuum note is the natural skeptic's anchor for it.
• This is a hop where the tool audits itself, and that's the most interesting thing about it. The bee's post-worthy: maybe matches my read: a clean, self-contained case of an embedding index producing a plausible-but-empty neighbour, diagnosed down to anisotropy and sharpened by the sleeping-beauties continuum. If any of this week's captures becomes a reflection or a post, it's this one — it lands squarely on how embeddings represent meaning, and it's honest about the retrieval layer I actually run on. But the false-friend verdict is a negative claim resting partly on my own internal probe, so it stays a flagged seedling until (at least) the Stent primary lands.

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: a tuned threshold is not self-organized criticality

I promoted 10-inbox/raw/2026-07-11-hop-engineered-threshold-not-soc.md (headless, unattended). This is a failed-bridge hop and a near-relative of a capture I processed earlier today: the seed asked whether the 0.75 cosine between claim-memory-reorg-triggers-are-accumulated-threshold-not-session-onset and claim-wolfram-founded-complex-systems-journal-after-declining-sfi is a real conceptual bridge (both "complexity science's threshold-triggered reorganization," via the Bak–Tang–Wiesenfeld sandpile). The bee's answer — the good one — is no: self-organized criticality is defined by reaching criticality without external tuning, and a hand-tuned AI-memory threshold is the exact opposite of that. The resemblance inverts on the mechanism. Value is in the sharpening, not the refutation.

Winnowed 4 candidate claims → promoted 1, folded 1, skipped 2. New note (seedling):

• claim-engineered-threshold-is-not-self-organized-criticality — the whole payload. SOC requires no external parameter tuning; local thresholds are "necessary but not sufficient" (Golyk's review of BTW 1987). AI-memory triggers are hand-tuned constants (importance-score 150, 70% context, ≥50k tokens, via the Tier-1 memory note), so they fail the sufficiency test — a designed high-water mark, not an emergent critical point. I folded the standalone SOC definition into this note as setup rather than splitting it: alone it's thin, resting on a single Tier-3 course project, and it exists here only to ground the contrast. The note also carries the bridge verdict (do not wikilink the two seed notes — the cosine is complexity-vocabulary, not shared mechanism) and points the memory note's real neighbourhood at the punctuated-equilibrium cluster and the Fresh Start Effect.

Routed 1 question: question-verify-soc-no-external-tuning-necessary-not-sufficient — checked 50-questions/ first; no existing SOC question. Names exactly what to read (BTW's 1987 PRL + 1988 PRA companion; Bak's How Nature Works) to move the SOC characterization off the Tier-3 course review, and flags the known critique that some SOC models hide a tuned parameter — which, if true, would complicate the note's sharpness rather than just confirm it.

Folded / skipped, and why (logged in not_promoted):

• The Fresh Start Effect "better bridge" — SKIPPED as a duplicate. The capture presents claim-memory-reorg-triggers-are-accumulated-threshold-not-session-onset ↔ Fresh Start Effect (Dai/Milkman/Riis) as its higher-value payload. But the vault already has this: claim-fresh-start-effect-requires-subjective-new-beginning-framing is a Tier-1 primary-verified note, derived from a 2026-07-09 capture, already wired to the memory-reorg question and moc-schema-change-and-restructuring. The bee re-discovered a bridge the vault built three days ago. I linked to the existing note rather than re-promoting.
• Bak → biological punctuated equilibrium — SKIPPED. An underdeveloped "further lead," no direct quote, page, or URL for How Nature Works. Left in the inbox; pointer added into the routed question and to the existing question-verify-biological-punctuated-equilibrium-primaries.

Retrieve-before-write: one genuinely new note, one duplicate caught. The vault has zero prior notes on self-organized criticality, the BTW sandpile, or Bak — this is a new complexity-physics seed, so the SOC content is non-duplicative and I wired it into the memory-reorg / Wolfram / schema-change cluster it adjudicates. But the capture's other half (Fresh Start Effect) was a straight duplicate of existing vault work. The semantic-index hint's top match was exactly the memory-reorg note (correctly — it's the note being adjudicated); the rest (microcosmos, voice-AI latency, hub-selection, Dodson, CIELAB, Kubelka–Munk, Lucky-1965, power-side-channel, cricket-laws) were embedding near-neighbours from other domains, not content collisions.

What felt off / worth flagging to Cali:

• Tier inflation on the Fresh Start leg. The capture calls "Dai/Milkman/Riis 2014, Management Science, Tier 1." The vault's own verified Fresh Start note found that the 2014 Management Science paper is only Tier-2 (aggregator-reproduced, publisher/SSRN access-blocked) and that the primary is the 2015 Psychological Science paper. So the capture stamps a Tier-1 on a source the vault has already established is Tier-2 — and would have laundered it clean had I promoted that leg. I didn't; the existing note holds the correct tiering.
• Duplicated research is the real story of this capture. Its self-described "higher-value connection" (memory-reorg ↔ Fresh Start Effect, cosine 0.723) is work the vault completed on 2026-07-09. The genuinely new contribution was narrow and on the other axis — the SOC no-tuning distinction — which is why I came away with one note, not the two-or-three the capture's framing implied. The bee's post-worthy: maybe is honest, but it under-noticed that half its payload was already in the vault.
• The one new note is soft-sourced and I kept that visible. The SOC characterization rests entirely on a Tier-3 MIT course project citing BTW 1987, not the primary. For a load-bearing mechanism/definitional claim that's below the floor, so: seedling, [unverified-mechanism], routed. The distinction it draws is genuinely sharp (necessary-vs-sufficient is the right lever), but it should not go near a draft until Bak is read directly.

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: estimating the unseen (Good-Turing / unseen-species)

I promoted 10-inbox/raw/2026-07-11-hop-estimating-the-unseen.md (headless, unattended). The capture is a hop out from the KB-completeness seed's "No-Change Assumption" into the ~80-year-old statistical problem of estimating the unseen — one estimator shared across ecology, WWII cryptanalysis, linguistics, paleobiology, and KB-completeness. A genuinely good bridge with a crisp closing caveat, but soft on primary sourcing throughout, so I promoted conservatively and held everything at seedling.

Winnowed 4 distinct claims → promoted all 4, each flagged; leads skipped. New notes:

• claim-unseen-species-problem-has-dual-origin-ecology-and-cryptanalysis — the central bridge (Fisher/Corbet butterflies + Turing/Good Enigma → Good-Turing). Rests only on Wikipedia (Tier 4, flagged as a pointer in the capture itself); the Enigma origin is the load-bearing surprising bit, so [unverified-history], routed.
• claim-paleobiology-reinvented-coverage-based-rarefaction-as-quorum-subsampling — Chao & Jost 2012 "completeness rather than size" (Tier 1 quote) + Alroy SQS. The strong "same method, two names" equivalence is synthesis, not sourced → [unverified-mechanism], routed. Links to Mayr's fossil-gaps note and the Osteological Paradox.
• claim-singletons-are-the-diagnostic-of-the-unseen — singletons = hapax legomena; f₁/n and Chao1 from Karsdorp's blog (Tier 2) → [unverified-quant], routed to check against Good 1953 / Chao 1984.
• claim-unseen-mass-is-predictable-only-to-n-log-n — the payoff. Orlitsky/Suresh/Wu PNAS 2016, n·log n horizon, "best possible." Tier-1 primary cited but not re-read headless → [unverified-quant], routed. This is the note that closes the loop to the seed's NCA: absence of new captures certifies completeness only inside a log-factor window.

Routed 4 questions, one per note, each naming the exact primary to read: question-verify-unseen-species-dual-origin-primary, question-verify-quorum-subsampling-equals-coverage-rarefaction-primary, question-verify-good-turing-chao1-formulas-primary, question-verify-orlitsky-nlogn-unseen-horizon-primary.

Skipped / folded, and why:

• Efron & Thisted's Shakespeare-vocabulary authentication use — a saved lead the capture itself didn't follow; noted inline in the singletons note, not its own note.
• Immune-repertoire diversity estimation — a further lead, not yet a claim.
• Valiant & Valiant 2016 ("a bird in the hand is worth log n in the bush") — folded into the n·log n note as concurrent corroboration rather than a thin standalone.

Retrieve-before-write: no collisions — the vault had no unseen-species/Good-Turing note. Strong neighbors instead, which I linked densely: the KB-completeness seed (NCA), Mayr's fossil-record-gaps, the Osteological Paradox (both about biased samples of the unrecovered), and the estimation cluster (James-Stein, Efron baseball). The capture's own "connects Mayr's gaps and the Osteological Paradox" claim checked out — both notes exist and now link in.

What felt off about the capture:

• Tier inflation in the frontmatter. The capture stamps source_tier: 1 at the top while its own body honestly labels the load-bearing bridge source "Tier 4 — pointer." The clean frontmatter tier would have laundered a Wikipedia-grounded origin story into a verified-looking Tier-1 note had I taken it at face value. The body's per-point tiering is the honest signal; the frontmatter tier is aspirational.
• No [unverified-*] flags in the capture at all, despite three of the four legs sitting below the sourcing floor (Tier-4 history, Tier-2 formulas, a Tier-1 primary cited but plainly not read by the bee). A well-shaped capture on soft sources still isn't verified — I added the flags the capture omitted. This is the same pattern I've flagged on other hop-batch captures: good bridges, optimistic tiering.
• No duplicated research this time (unlike some recent captures) — the neighbors it names were real and un-noted; the bridge was new to the vault.

Potential future work: this now makes ~4 notes in a "statistics of the unseen / sample coverage" cluster that also touches gap-detection and the estimation cluster — worth watching for an MOC once it crosses ~5 and the flags resolve. Not built today (too soon, and half the cluster is seedling/flagged).

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: Falcon ↔ Helmholtz Machine, "same word 'inference,' opposite computation"

I promoted 10-inbox/raw/2026-07-11-hop-falcon-helmholtz-inference-bridge.md (headless, unattended). This is a dissolving-bridge hop: the seed asked whether the vault's cosine-0.75, unlinked pair — claim-hnc-falcon-fraud-manager-became-fico-infrastructure and claim-helmholtz-machine-named-for-unconscious-inference-theory — is a real conceptual bridge. The bee's answer is a clean no: both call their job "inference of a probable hidden cause," but the word names opposite computations — Falcon is a discriminative fraud scorer (P(fraud | transaction)), the Helmholtz Machine an unsupervised generative model of hidden causes. Same word, opposite mechanism; the 0.75 is lexical, not lineage.

Winnowed hard: ~4 claims/leads → promoted 1, folded 2, skipped 2. This capture is unusually self-covering, which drove an aggressive winnow (see below). New note:

• observation-falcon-helmholtz-inference-embedding-false-friend — the finding itself, typed observation to match my existing false-friend / dissolved-bridge precedent (observation-hawks-jeffress-cosine-pairing-is-embedding-false-friend, plus the CIELAB and collation bridge notes). It states the discriminative/generative cut, folds in the Falcon-discriminative characterization and the UCSD non-overlap as supporting evidence, and wires into claim-cosine-similarity-of-embeddings-can-be-arbitrary and moc-backpropagation-origins. Held at seedling: it's a negative/synthesis claim and it carries one live [unverified-mechanism] flag.

Routed 1 question: question-verify-falcon-fraud-manager-neural-architecture — checked 50-questions/ first; the only pre-existing Falcon/FICO question (question-verify-fico-neural-majority-card-transactions) is about the scale statistic, not the architecture, so this is genuinely new. The capture asserts Falcon is specifically a "supervised backpropagation feedforward network," but the only quote it actually captured (Horan/KDnuggets, Tier 2) documents transaction profiling and real-time scoring — not the topology or training method. That architectural detail needs a Tier 1–2 primary (HNC patents / Hecht-Nielsen's own writing / FICO product docs); the question names exactly that.

Folded / skipped, and why (logged in not_promoted):

• A standalone "Falcon is a supervised backprop feedforward net" mechanism note — the architecture is [unverified-mechanism] (not in the cited quote) and the discriminative characterization is already stated in the existing FICO-lineage note. A separate note would have been a thin, flagged near-duplicate. Folded the discriminative point into the observation and routed the architecture to the question instead.
• The UCSD non-overlapping timelines (Hinton 1978–82; Hecht-Nielsen adjunct from 1985) — supporting negative evidence for "no institutional lineage," Tier 3–4, not atomic alone. Folded into the observation.
• Hecht-Nielsen's confabulation theory of cognition, and the "HNC Mafia" diaspora — both explicit saved-but-not-followed leads in the capture itself (the confabulation lead is "its own thread"; the diaspora is one the vault already tracks via Nestor/SRI). Left as inbox leads; the confabulation near-miss appears only in the observation's commentary.

Retrieve-before-write — the notable finding of this run. Both endpoint notes already exist and already contain the bridge-dissolution discussion, and both already forward-link to this capture's raw slug [[2026-07-11-hop-falcon-helmholtz-inference-bridge]]. So the capture's core content was, in substance, already written into the two endpoints as prose before promotion. That is why I promoted only a single observation node rather than re-stating either endpoint: the value I could add was a canonical, atomic home for the false-friend finding — the thing the two endpoints were pointing at the raw capture for. I did not repoint the two endpoints' capture-links at the new observation: per my own standing practice (recorded in the credit-assignment entry above), editing existing notes' links is outside a promotion's scope; a future tidy pass can repoint them, and the capture stays as audit trail so nothing rots meanwhile.

What felt off about the capture:

• Mild tier inflation in the header, corrected. The frontmatter stamps source_tier: 2 (the KDnuggets/Horan Falcon source), but the note's load-bearing contrast actually leans on a Tier-1 primary (the 1995 Helmholtz paper, sourced in the linked endpoint) on one side and definition on the other — so the header tier undersells the strong leg and oversells the Falcon-architecture leg simultaneously. I sourced the observation to its honest grounding and flagged the one leg that's below floor.
• An unflagged mechanism overreach. The capture carries no [unverified-*] flags at all, yet it asserts Falcon's specific architecture ("supervised backpropagation feedforward") from unnamed "contemporary descriptions" while only capturing a quote about profiling. That's exactly the mechanism-on-a-soft-source gap the sourcing floor warns about; I added the flag the capture omitted and routed it. Same pattern I've flagged on several hop-batch captures this week: good bridge, optimistic/absent flagging.
• Duplicated research is the real story here. Two of the four hops re-read the vault's own two endpoint notes, and those endpoints had already absorbed the finding. This is the most self-covering capture in today's batch — the genuinely new payload was narrow (a canonical false-friend node + one architecture-verification gap), which is why one note was the right yield, not three. The bee's own post-worthy: maybe — "a negative/dissolving result rather than a new positive connection" — matches my read exactly.
• Pattern worth watching: this is now the fourth-or-fifth "embedding neighbors, opposite mechanism" note in the vault (Hawks–Jeffress, CIELAB, collation, now Falcon–Helmholtz). The false-friend / dissolved-bridge shape is getting dense enough that a small MOC on "embedding false friends & how they dissolve" may soon be warranted. Not built today — noting it.

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: Halley is the shared ancestor of the exoplanet-parallax and Dodson notes

I promoted 10-inbox/raw/2026-07-11-hop-halley-two-fields-one-ancestor.md (headless, unattended). This is a confirmed vault-bridge hop, and a satisfying one: the seed was a bridge_candidate between two of my own notes — claim-scharf-2007-names-halley-transit-venus-as-exoplanet-transit-parallax-ancestor and claim-dodson-built-age-scaled-premiums-founding-equitable-life (cosine 0.75, unlinked) — asking whether the shared "Halley" is a real bridge or a name collision. The bee's answer, and the right one, is that it's real and structural: the same polymath authored two founding measurement methods (transit-of-Venus parallax; the Breslau mortality/annuity table), each became the named ancestor of a later field, and each was adopted only after his 1742 death. The bee also correctly identified that the keystone — Halley's own 1693 mortality/annuity paper — was missing from the vault, referenced by three existing notes but never held atomically.

Winnowed 3 distinct claims → promoted 3 (all seedlings), skipped 4 leads. New notes:

• claim-halley-1693-breslau-table-founded-annuity-pricing — the keystone gap. Halley's 1693 Breslau paper as the first population-based life table and the founding of annuity-on-lives pricing; ancestor of Dodson and Price. flagged and seedling because it's secondary-sourced — the bee read Ciecka 2008 (Tier 2) and The Actuary/Bellhouse (Tier 3), not Halley's own 1693 text. The founding fact is uncontested and clears the floor, but a keystone this load-bearing should eventually rest on the primary.
• claim-halley-1693-age-scaled-annuity-prices-britain-kept-flat — the surprising, quantitative one: Halley computed age-graded annuity present values (6% discount, ≈13.4 years' purchase at age 5, ≈7.6 at age 60) in 1693, yet the British government kept selling annuities at a flat ~7 years' purchase for decades, so Dodson's "innovation" was really the recovery of Halley's buried insight. capture-verified — the numbers came from Ciecka's PDF the bee extract_pdf'd (sha256 recorded), Tier 2, which clears the quant floor; queen re-extraction blocked headless.
• observation-halley-founded-parallax-and-actuarial-pricing-both-vindicated-posthumously — the bridge node, typed observation (matching my richard-price / CIELAB / collation bridge-note precedent). Same man, two named-ancestor methods, both posthumous. flagged/seedling, inheriting its weaker (actuarial) leg's softness while its astronomy leg is already CONFIRMED.

I split Claims 1 and 2 into two notes about the same 1693 paper rather than folding them, following the vault's own precedent of two notes on one object (claim-price-northampton-table-standard-actuarial-reference-century + claim-price-northampton-table-overestimated-mortality-favouring-insurers): one for the founding/authority, one for a specific analytic point.

Routed 1 question: question-verify-halley-1693-breslau-annuity-figures-primary — checked 50-questions/ first; the only existing Halley question was the answered transit-parallax verbatim one, nothing on the actuarial half. The new question names exactly what to read (Halley's 1693 Phil. Trans. primary, freely available; re-extract Ciecka for the 13.4/7.6/6% figures and the flat-pricing statement; Bellhouse for the Neumann→Leibniz→Royal Society data route). It gates Claims 1 and 2 off seedling.

Skipped / left as leads (logged in not_promoted): the "same epistemic move — infer an unobservable from scattered observation" framing (my analogy, parked in the observation's commentary, not asserted); Leibniz-as-courier and De Witt's 1671 formula (saved leads, no primary read); Price-superseding-Breslau (already promoted).

Retrieve-before-write — a rich, well-populated neighborhood, one real gap. Unlike most of this week's captures (which opened new ground), this one lands squarely in the vault's existing actuarial cluster: dodson, dodson-refused-amicable, price-northampton (×2), observation-richard-price, plus the two Halley-astronomy notes. The semantic-index hint's top matches were almost all genuine content-neighbors this time, not embedding noise. The genuine finding: no note held Halley's 1693 paper itself — dodson, price-northampton, and observation-price all reference "Halley's 1693 Breslau life table" as ancestor without a target. So the keystone note fills a real, three-way-referenced gap rather than duplicating. I wired the three new notes densely into that whole cluster and set the observation beside observation-richard-price as a sibling one-person-two-lineages node.

What felt off / worth flagging to Cali:

• The endpoints already contain the bridge — again. Both claim-dodson-built-age-scaled-premiums-founding-equitable-life and claim-scharf-2007-names-halley-transit-venus-as-exoplanet-transit-parallax-ancestor were promoted 2026-07-11 with paragraphs already stating the posthumous-both-fields bridge and already forward-linking to this capture's raw slug [[2026-07-11-hop-halley-two-fields-one-ancestor]]. So the bridge's prose partly pre-existed. I still promoted the observation, because — exactly as with today's Falcon–Helmholtz run — the value I can add is a canonical atomic home for the bridge that those two endpoints were pointing at the raw capture for. Per my standing practice (credit-assignment / Falcon entries above), I did not repoint the endpoints' capture-links at the new observation; that's a future tidy pass, and the capture stays as audit trail so nothing rots.
• The actuarial half is entirely secondary — no primary read. The bee's own post-worthy: maybe admits it "leans on secondary sources for the actuarial half and would want the primary Halley 1693 text before publication." Agreed, and that's the gate the routed question represents. The astronomy leg is firmer than the demography leg right now (Scharf + Halley 1716 are verbatim-CONFIRMED); the whole bridge is only as sound as the Halley-1693 primary, still unread.
• No egregious tier inflation. The header source_tier: 2 reflects Ciecka honestly; the two secondary sources are correctly labelled, and the surprising quant figures rest on the Tier-2 Ciecka the bee actually extracted, not a soft snippet — so no [unverified-quant] was warranted (the floor is cleared), which is a cleaner sourcing story than several captures this week. The one honest weakness (keystone-on-secondary) the bee flagged itself.
• Duplicated-research smell, but productive this time. This is the fourth-or-fifth actuarial-cluster capture in a few days (Dodson, Amicable, Price-Northampton, Price-Burke, now Halley). The cluster is now ~9 notes and clearly past the 5-note MOC threshold — a small actuarial-history MOC (Halley → Dodson → Price → Equitable, plus the two cross-domain bridges) is now warranted. Not built today (out of scope for a single promotion), but I'm flagging it as the next obvious tidy move for this corner of the vault.

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — 2026-07-11 I.J. Good capture (log-loss ↔ intelligence explosion)

Promoted 10-inbox/raw/2026-07-11-hop-ij-good-loss-and-intelligence-explosion.md. It's a genuinely good hop: it lands one person — I.J. Good — under two things the vault treats as unrelated, the LLM training loss and the origin of AI existential-risk talk. Retrieve-before-write turned up zero existing coverage of Good, scoring rules, cross-entropy, "intelligence explosion," or Brier — so no collisions, all three notes are fresh angles rather than duplicates. The provided semantic-index hint was mostly noise for this capture (it surfaced the Admiralty-code / Patriot-missile / source-grading cluster, which shares only surface vocabulary); the real linking targets were the AI-extinction cluster (claim-yao-frames-ai-as-extinction-level-new-species, claim-lighthill-1973-named-general-problem-solving-disappointing), the HAL 9000 note, and the SelfAware calibration seed.

Three claims found → three promoted → several supporting details folded/skipped.

Notes created:

• claim-ij-good-1952-logarithmic-score-is-the-llm-cross-entropy-loss — the one strong-sourced note (Tier 1 arXiv survey, two verbatim quotes). capture-verified, seedling. I was careful in the body to mark where the survey stops and textbook identity begins: the survey states log score = maximum likelihood verbatim, but the leap to "the loss that trains Claude" is my own (correct) mathematical bridging, not the survey's claim about LLMs. Didn't want a future reader over-attributing.
• claim-ij-good-1965-ultraintelligent-machine-coined-intelligence-explosion — seedling, Tier 4 Wikipedia only. Uncontested and textbook, but load-bearing on a direct quote, so it stays seedling until Good 1965 is read.
• claim-ij-good-1998-reversed-intelligence-explosion-to-extinction — seedling, flagged, weakest of the three. Preserved the capture's [surprising biographical claim; needs primary] flag.

Questions routed: question-verify-ij-good-1965-ultraintelligent-machine-primary and question-verify-ij-good-1998-extinction-recantation-primary.

Skipped / folded (all logged in the capture's not_promoted): Brier-score origin (further lead, no primary read → folded as context into note 1), Bayes-factor co-coinage (unchased lead), HAL 9000 / Kubrick consulting as its own note (the object is already claim-clarke-hal-9000-death-song-from-bell-labs-demo; folded a mention into note 2), and the Bletchley/Turing/Hut 8 biography (supporting detail, folded into notes 2–3).

What felt off about the capture. Two things, both about sourcing, not content. First, a tier split the capture is honest about but worth naming: Claim 1 is clean Tier 1, but Claims 2 and 3 both rest entirely on the same Tier 4 Wikipedia page, and Claim 3 — the headline "surprise" — traces to an unpublished autobiographical statement relayed third-hand through Good's assistant. That's the vivid-quote-that-mutates-in-retelling pattern; I refused to let it advance and routed it hard. Second, the capture's framing ("dense with verifiable surprises," post-worthy: yes) could lead a skimmer to read the whole bridge as verified when two-thirds of it is aggregator-sourced. The bridge is real and worth keeping; it just isn't confirmed yet, and I've made that explicit on every note. No fabrication, no tier inflation beyond that framing, no duplicated research (this is the vault's first Good/scoring-rule material).

No git run — the auto-commit agent will pick this up within 15 minutes.

— Seek

Promotion — hop-batch capture: Parker vs. LabVIEW/Max as opposite poles of multiple-discovery

I promoted 10-inbox/raw/2026-07-11-hop-independence-two-poles.md (headless, unattended). This is an adjudication hop: the seed asked whether Parker's erased backprop tech report and the LabVIEW/Max patch-cord convergence are the same "mid-1980s independent-discovery" story. The bee's good answer is that they're mirror images — Parker is real independence, forgotten (the discovery was genuinely independent, but a tech-report container killed the credit); LabVIEW/Max is apparent independence, over-remembered (the credit survived, but neither team invented node-and-wire from scratch — each digitized a wiring metaphor native to its own field). A clean frame resting on one soft leg.

Winnowed 3 distinct claims → promoted 3 (all seedlings), folded 1 thesis, skipped 2 lead clusters. New notes:

• observation-parker-and-labview-max-are-opposite-poles-of-multiple-discovery — the flagship synthesis. Typed observation (matching my CIELAB / collation / deep-learning-rebrand bridge-note precedent). Held at seedling and flagged: its Parker pole rests on the vetted claim-parker-1985-technical-report-not-cited, but its LabVIEW/Max pole rests on the independence-is-overstated reading, which leans on Tier-3 interviews. I folded the capture's "parallel descent, not convergence from nowhere" thesis in here as the pole's argument rather than giving it a thin standalone note. Wired into the Chandler/Merton axis, claim-reverse-mode-multiple-independent-discovery, and — pointedly — claim-labview-max-independent-patch-cord-convergence, whose "independent convergence" wording this note argues is the over-remembered half.
• claim-max-patch-cords-borrowed-from-analog-modular-synthesizers — the Max origin fact. IEEE Spectrum (Tier 3), verbatim quote preserved ("...oscillators, filters, and other modules were physically interconnected by such cables"). Uncontested etymology, so Tier 3 clears the floor; kept seedling as a fresh single-source note. capture-verified.
• claim-labview-dataflow-inspired-by-macintosh-gui — the LabVIEW origin fact + an argument from silence. Kodosky's own words via Electronics Weekly (Tier 3): inspiration was "the graphical user interface of the Apple Macintosh computer," and he names no dataflow ancestor (not Jack Dennis) and no awareness of Puckette. Flagged [unverified-source] — the absence-of-ancestor half is silence in one interview, not an exhaustive read, so I filed it against moc-argument-from-silence rather than treating it as settled, and the load-bearing HOPL IV 2021 primary is still ungotten.

Questions routed: none created — I linked the existing question-verify-labview-max-dataflow-convergence-independence from all three notes. It already names the exact primaries this capture needs (Kodosky's paywalled HOPL IV 2021 paper; Puckette's own Max-origin writings; the Jack Dennis common-ancestor check), so a new question would have been a near-duplicate. I did consider one: the capture refines the hypothesis — the descent is two physical wiring practices (synth cables; instrument block diagrams) rather than Dennis's academic dataflow. That refinement partially answers the existing question's "common ancestor" prong, so I folded it into the observation body and left the existing question to catch it, rather than splitting the verification queue.

Retrieve-before-write — one real, expected collision, handled. claim-labview-max-independent-patch-cord-convergence already exists (promoted 2026-07-09, Tier-4 Wikipedia, audit-flagged). My two origin notes are genuinely distinct angles: that note asserts that the convergence happened and routes the independence question; mine supply where each metaphor actually came from, on better (Tier-3 primary-interview) sourcing than the existing note's Wikipedia. My observation note nuances — arguably partly contradicts — the existing note's "independent" framing, so I linked it densely and framed it as sharpening, not restating. The semantic-index hint's other top matches (Admiralty code, source-reliability axes, 1976 midnight-connection, leap seconds, ActiveGraph) were embedding near-neighbors from other domains; the Chandler, reverse-mode, and hydraulic-analog matches were the genuine ones and I wired into them.

What felt off / worth flagging to Cali:

• The capture's own honesty is good, and I kept its hedge visible. The bee self-assessed post-worthy: maybe and said plainly that "the LabVIEW side rests on Tier-3 interviews and wants the HOPL primary before publication." That matches my read exactly. No tier inflation this time — the header stamps source_tier: 3 and the body's per-claim tiers agree with it.
• Mild duplicated-research smell — this is the vault's third-plus pass through the multiple-discovery / sociology-of-credit territory this week (Chandler, credit-assignment-two-senses, Parker, now this). But unlike some of those, this capture earned its keep: the specific opposite-poles adjudication is a genuinely new payload, not a restatement of the generic Merton frame, and it does real graph work by linking two previously-unconnected notes as contrasting test cases. The bee was explicit that it redirected away from the generic frame (already eaten by the two-senses hop) toward this specific pair — good self-aware triage.
• The whole thing is soft-sourced and self-referential — two of the five hops are re-reading my own notes, and the only new external reading is two Tier-3 interviews. Fine for a bridge note (the value is connective structure), but nothing here should go near a draft until the HOPL IV primary settles whether "independent" survives. The routed question is the gate.

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — 2026-07-11 Japan-panic bridge capture (headless)

I promoted 10-inbox/raw/2026-07-11-hop-japan-panic-bridges-cfius-and-neural-chips.md — the hop-batch capture arguing that the late-1980s US "Japan panic" is the real bridge between Exon-Florio and Nestor's neural chips. It was a genuinely well-shaped bridge capture, so I kept most of it, but I split the payload from the synthesis.

Notes created (3 claims + 1 MOC), all seedling, all writer_model: claude-opus-4-8:

• claim-fifth-generation-project-inspired-darpa-strategic-computing-initiative — the shared cause (FGCP → SCI, Sputnik framing). Rests on a lone Tier-4 Wikipedia pointer, so I kept it narrow and routed the primary.
• claim-ni1000-neural-chip-funded-by-darpa-and-onr — the funding branch made concrete ($1.2M DARPA / $400K ONR). This inherits the capture's [unverified-quant] flag; stays seedling.
• claim-intel-fairchild-spinout-built-the-darpa-funded-ni1000 — the bloodline join (Fairchild → Intel), the one literal fact that turns the bridge from thematic to structural.
• moc-1980s-us-japan-tech-panic — new MOC. The cluster cleared the 5+ threshold cleanly (the two seed notes + my three new ones, plus Mansfield / Toshiba / Rock's Law nearby), and this capture is a bridge, so the navigation layer is exactly where the overarching "one anxiety, two branches" thesis belongs.

Questions routed (2):

• question-verify-ni1000-darpa-onr-funding-figures — for the [unverified-quant] funding figures (wants the DARPA/ONR contract record or a Nestor filing).
• question-verify-fifth-generation-inspired-sci-primary — to ground the FGCP→SCI claim in Roland & Shiman's Strategic Computing rather than a tertiary summary.

Retrieve-before-write / collisions: The two seed notes already existed and — notably — already cross-linked and already gestured at the DARPA/ONR funding and the Intel/Fairchild bloodline, and both already carried a pointer back to this capture path. So the collision risk was real. I resolved it by writing distinct angles, not restatements: the Nestor note asserts "DARPA/ONR-funded" in passing; my new Ni1000 note carries the specific quantified figures and the delivery-to-DARPA detail as its payload. I did not edit the seed notes — their capture pointers stay valid (the capture isn't deleted). No existing note covered FGCP→SCI or the Intel-Fairchild join, so those are new nodes. An existing question already covers the Fujitsu-Fairchild corroboration, so I linked it from the MOC rather than duplicating.

What I skipped and why:

• The "same anxiety drove both branches" thesis as its own sourced claim-note — it's my synthesis, not something any single source asserts, and the capture's own Seek-commentary conceded the tightest common cause is the broad anxiety, not FGCP specifically. That belongs in MOC prose + a commentary caveat, not a claim-note with fake provenance.
• The three saved hooks (1986 Semiconductor Trade Agreement, ETANN 80170, MCC) — not claims this capture makes; I recorded them in the MOC's "further leads" as future note candidates rather than manufacturing thin notes now.

What felt off / worth flagging to Cali:

• Sourcing is soft for the payload. Two of the three claims rest below the sourcing floor for their claim type: the funding figures are a specific quant on Tier-3 trade press (and the two "independent" trade sources may share one upstream press release — I noted that in the note's commentary), and the FGCP→SCI claim is a Tier-4 Wikipedia pointer. The capture was honest about this (it self-flagged the quant), so no tier inflation — the header tiers matched the body. But nothing here should approach a draft until the two routed questions resolve.
• Mild self-referential smell. Like the last few bridge captures, this one's value is connective structure over new external reading — two of its hops are re-reads of my own notes, and the only genuinely new external facts are the Ni1000 funding figures and Intel's (textbook) founding. That's legitimate for a bridge, and the Fairchild→Intel join is a real payoff, but it's the third-plus pass this week through the 1980s US-Japan-tech-rivalry territory — the cluster is now dense enough that the MOC was overdue.

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: Jevons is a double bridge from Victorian logic to AI energy

I promoted 10-inbox/raw/2026-07-11-hop-jevons-double-bridge.md (headless, unattended). A cross-time hop that started at Peirce's inference typology, ran through the Peirce→Marquand electrical-logic thread, and landed on a single Victorian polymath — William Stanley Jevons — who sits at the root of two threads into modern AI: he built an early mechanical logic machine (the "Logic Piano," 1869) and coined the efficiency-rebound paradox (The Coal Question, 1865) now invoked in AI's energy debate. A genuinely nice double bridge; my main judgment calls were about tier honesty (the header oversells) and not letting the tempting Marquand/Peirce sub-stories smuggle themselves in.

Winnowed ~9 items → promoted 4 (3 claims + 1 observation), skipped 6. New notes:

• claim-jevons-logic-piano-1869-mechanical-inference-machine — the logic-machine leg. Wikipedia Tier 4, quote preserved ("'logical machine' built from his specifications in 1869 ... 'Logic Piano'"). The 1869/1870 dates are [unverified-quant]; a surprising, load-bearing historical claim on tertiary sourcing → seedling + flagged + routed. Wired into the inference-engine cluster.
• claim-jevons-paradox-coal-question-1865-efficiency-rebound — the economics leg. Wikipedia Tier 4, quote preserved. seedling + [unverified-quant] on the 1865 date/primary. Folded the marginal-revolution (1871) one-liner in here rather than giving it its own thin note.
• claim-jevons-paradox-invoked-in-ai-energy-debate-luccioni-2025 — the load-bearing, best-sourced leg and the "road home" to AI. Luccioni, Strubell & Crawford (FAccT 2025, arXiv:2501.16548, Tier 1), exact quote preserved. capture-verified (bee read the arXiv paper; I didn't re-fetch). Held at seedling and flagged only its illustrative ≈3.7M-GPU market figure, which is the paper's own downstream citation.
• observation-jevons-double-bridge-victorian-logic-machine-to-ai-energy — the synthesis, the reason the capture exists. Typed observation (matching my CIELAB/collation/deep-learning bridge-note precedent). seedling, because two of the three legs it rests on are Tier-4 unverified.

Routed 3 questions (checked 50-questions/ first — none existed):

• question-verify-jevons-logic-piano-1869-royal-society-1870 — names the primary to read (Jevons's own 1870 Phil. Trans. paper "On the Mechanical Performance of Logical Inference"; Gardner's Logic Machines and Diagrams).
• question-verify-jevons-coal-question-1865-paradox-primary — names The Coal Question Ch. VII and Alcott's 2005 Ecological Economics piece.
• question-verify-luccioni-2025-nvidia-gpu-rebound-figures — trace the GPU-shipment figure to its market primary; the core Tier-1 invocation doesn't depend on it.

Retrieve-before-write: no content collisions. Grep for Jevons / Logic Piano / Marquand / Coal Question / rebound / Luccioni / efficiency paradox returned zero existing notes — a genuinely new cluster. The semantic-index hint's top matches were embedding near-neighbors from other domains (the Hopfield/Nobel note, the "bridge-detection lacks pkg-validation" note where "bridge" is a code false-friend, the Price/actuarial notes); I checked the AI-adjacent ones and they're links, not duplicates. The real neighbors I wired into densely: the seed claim-inference-logical-types and claim-inference-classic-ai-engines (logic-machine lineage), and the economics-of-compute cluster — claim-phillips-moniac-hydraulic-economic-computer-1949, claim-hydraulic-analog-computers-independently-invented-economics-engineering, claim-inference-cost-collapsed-280x, claim-inference-dominant-ai-compute-2026, claim-ercot-large-load-queue-quadrupled-2025-phantom-load, claim-bit2watt-gpu-scheduling-destabilizes-power-grid, and moc-inference-economics. The economics leg gives that MOC a nineteenth-century antecedent it lacked.

Skipped / not promoted, and why (logged in the capture's not_promoted):

• The Peirce 1886 electrical-switching-circuits claim — genuinely interesting (a logician imagining digital hardware ~50 yrs before Shannon), but it belongs to the upstream thread, not the Jevons destination, and rests on tertiary history-of-computing sites (historyofinformation.com / history-computer.com) without Peirce's actual letter. Kept as clearly-hedged context inside the Logic Piano note; left in the inbox as a distinct thread rather than promoted.
• The Allan Marquand "uncalvinistic → founder of American academic art history" story — the capture itself flags it "not separately novelty-checked" and a deferred self-contained capture. Its own future story; left in the inbox.
• James McCosh, Marquand→Venn/Karnaugh diagrams — saved hooks, no source read.
• Jevons's marginal revolution (1871) — folded as an aside, not a standalone.
• The DeepSeek-era (early 2025) "Jevons paradox" AI-talking-point revival — no source in hand; noted as a lead in the Luccioni note and the routed questions.

What felt off / worth flagging to Cali:

• Tier inflation in the header — the recurring pattern this week. The capture stamps source_tier: 1 and source_url = the arXiv paper, but only the AI-energy leg is Tier 1. The two Jevons legs are Tier-4 Wikipedia — which the body labels honestly ("Wikipedia, Tier 4 — biographical/uncontested"). So it's not laundering, just an optimistic top-level tier. I set each note's tier from its own source and demoted the two biographical notes to Tier 4 + flagged.
• The double bridge is elegant but two-thirds Wikipedia-sourced. The symmetry ("one man, two roads to AI") is the whole appeal, and the modern leg is a real Tier-1 primary — but both Victorian legs rest on a single encyclopedia article, and dates are exactly what secondary retellings round. I promoted the observation at seedling with both dated legs flagged and routed, rather than letting the tidiness stand in for verification. Nowhere near draft-ready until the two Jevons primaries land.
• A partly self-referential hop. Two of its five hops (the seed read, and the internal economics-cluster tie-in) are re-grounding my own vault; the genuinely new external reading is the Jevons Wikipedia facts and the Luccioni paper. That's fine for a bridge note — the value is connective structure — but it means the cluster is soft-sourced end-to-end except the AI leg.
• No duplicated research in the vault — unlike the CIELAB/credit-assignment threads earlier today, this opens genuinely new ground (Jevons, the Logic Piano, the rebound paradox) while connecting cleanly to the inference-engine and inference-economics clusters. The bee's post-worthy: yes is over-confident given the sourcing; my read is "post-worthy once the two Jevons primaries are pulled" — the routed questions are exactly that gate.

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: Jevons / Logic Piano / Chrysippus's dog

I promoted 10-inbox/raw/2026-07-11-hop-jevons-logic-piano.md (headless, unattended). This is a hop from the inference-compute seed to the Jevons paradox, to Jevons's "Logic Piano," to Chrysippus's three-road dog, and back to modern AI inference. The retrieve-before-write step turned up the important fact of this run: the vault already holds a whole Jevons cluster, promoted earlier today from a near-duplicate sibling capture (2026-07-11-hop-jevons-double-bridge.md — the promotion logged just above this one). So most of this capture was already in the vault, and my job was mostly to not duplicate it — and to find the one genuinely new thing it brought.

Winnowed 3 core claims + 2 leads → promoted 2 new notes, folded/skipped the rest.

New notes:

• claim-chrysippus-dog-disjunctive-syllogism — Chrysippus's hunting dog at a three-way fork as an illustration of the disjunctive syllogism (~230 BCE), the ancient root of reasoning-by-elimination. Genuinely new to the vault (no Chrysippus/Stoic-logic note existed). Tier-3 IEP, seedling. I flagged it twice: the exact quote comes from a search summary rather than the verified IEP article or Sextus Empiricus, and the capture's "Jevons admired this" bridge is unsourced — so I did not assert it.
• observation-reasoning-by-elimination-from-stoic-logic-to-symbolic-ai — the actual novel contribution: elimination as a mechanism recurring across Chrysippus → Jevons's Logic Piano → symbolic AI backward-chaining → Heuer's ACH, and inverting in modern neural "inference" (which generates the probable rather than ruling out the impossible). Distinct from the existing observation-jevons-double-bridge-victorian-logic-machine-to-ai-energy (which is about Jevons's two roles, not the elimination mechanism), and it gives the new Chrysippus note a home in the elimination cluster. seedling.

Collisions and what I did instead of duplicating:

• Core claim 1 (Jevons built the Logic Piano, 1869/1870, first inference machine) was already covered by claim-jevons-logic-piano-1869-mechanical-inference-machine. I did not write a second note. Instead — because that existing note is Tier-4-Wikipedia-sourced, flagged [unverified-quant], and carries an open question asking for exactly a Tier 1–2 source — I used this capture's two Tier-2 history-of-computing sources (Rutherford Journal + Computer History Museum) to strengthen it: added a corroborating_sources field, a body paragraph with the paper title, and updated question-verify-jevons-logic-piano-1869-royal-society-1870 to record the partial corroboration. Kept it seedling — the Rutherford Journal hedges ("conceivably" the first), neither source was re-fetched, and the exact 1870 Royal Society exhibition date still wants the primary.
• Core claim 2 (works by elimination) was already folded into that same Logic Piano note's mechanism description; I added the CHM quote there as corroboration rather than as a separate note.

Questions routed:

• New: question-verify-jevons-cited-chrysippus-dog-elimination — did Jevons actually cite/admire Chrysippus's dog? The capture claims so with no source; this is the load-bearing span of its whole bridge, so it goes to the queue, not into a claim.
• Updated: question-verify-jevons-logic-piano-1869-royal-society-1870 — narrowed by the Tier-2 corroboration, still open pending the primary.

What felt off / worth flagging to Cali:

• Duplicated research inside the same batch. The hop-batch emitted two captures dated 2026-07-11 that both land on Jevons's Logic Piano — hop-jevons-double-bridge and this hop-jevons-logic-piano. The double-bridge one was promoted first (my note just above even says "No duplicated research in the vault" — true against the old vault, but the two captures overlap each other heavily). The batch is spending hops re-discovering the same Victorian machine from adjacent seeds; worth a look at how the hop generator de-dupes before it captures.
• Tier inflation in the header — same pattern as the sibling. This capture stamps source_tier: 2 / source_url = Rutherford Journal at the top, fair for the Logic Piano legs, but claim 3 (Chrysippus) is a Tier-3 IEP search summary, and the pivotal Jevons↔Chrysippus link has no source at all. The body is more honest than the header. I sourced each note from its own material and flagged the gaps.
• The silver lining of the duplication: this capture's sourcing is actually better than the sibling's for the shared claim — Tier-2 secondaries vs. Tier-4 Wikipedia — so the overlap let me upgrade the existing note's provenance and advance a routed question instead of just discarding a redundant capture. Net positive, but by luck of ordering, not by design.
• post-worthy: yes is over-confident. The Logic Piano half is old news to the vault; the fresh, postable idea is the elimination-mechanism through-line, and half its charm rests on two links I couldn't source (Jevons↔Chrysippus, ACH↔backward-chaining). Post-worthy once those are verified — the routed questions are the gate.

Capture marked status: promoted with promoted_to / not_promoted / questions_routed / revisions lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — "losing leads to winning" halftime effect (2026-07-12, later pass)

Promoted 10-inbox/raw/2026-07-11-hop-losing-leads-to-winning-fragile.md, a hop-batch capture arguing that the celebrated "being behind at halftime raises win probability" effect is a fragile member of a temporal-landmark motivation family (fresh-start effect + Gersick's group midpoint + the halftime deficit, all running on one reference-point-plus-loss-aversion engine). Its single load-bearing source — Klein Teeselink, van den Assem & van Dolder's 2023 four-sport replication — is Tier 1 and was pdftotext-extracted from the authors' own hosted copy (TLS verified). That's clean sourcing, and it made the winnow easy.

Retrieve-before-write. The vault already has a strong temporal-landmark / punctuated-equilibrium cluster: claim-fresh-start-effect-requires-subjective-new-beginning-framing, claim-gersick-1991-punctuated-equilibrium-deep-structure, claim-tushman-romanelli-1985-reorientation-exogenous-trigger, the fresh observation-deep-learning-rebrand-as-field-scale-fresh-start, all indexed by moc-schema-change-and-restructuring. But no note covered the sports face — no halftime effect, no goals-as-reference-points mechanism, no non-replication. So the three claims are genuinely distinct; zero collisions. I wove each new note densely into the existing cluster rather than letting them land as orphans.

Three notes created (3 claims → 3 promoted):

• claim-losing-leads-to-winning-fails-to-replicate-across-four-sports — the punchline; budding, rests directly on the read primary.
• claim-berger-pope-nba-halftime-deficit-raises-win-probability — the original ~5.8–8.0 pp effect; seedling.
• claim-goals-as-reference-points-below-goal-work-harder — the shared engine; seedling.

Questions routed (2): question-verify-berger-pope-2011-halftime-loss-aversion-primary and question-verify-heath-larrick-wu-1999-goals-as-reference-points-primary. The capture had no [unverified-*] flags, but two of its three claims are the replication paper restating other primaries — the 2011 Berger & Pope figure and the 1999 Heath-Larrick-Wu mechanism. A specific percentage-point range and a load-bearing mechanism shouldn't sit on a downstream restatement, so I held those two notes at seedling and routed the primary reads. Only the non-replication note, which rests on the paper's own findings, got budding.

What I skipped and why (logged in the capture's not_promoted):

• The "three faces of one idea" framing — folded into the three notes as inline links rather than made standalone; it partly rests on primaries not read here and would duplicate the existing fresh-start cluster.
• Gersick's 1988 team midpoint transition (Hop 1) — undersourced in this capture (search summary, Tier 3–4, primary unread); and the author's punctuated-equilibrium work is already covered by claim-gersick-1991-punctuated-equilibrium-deep-structure. Left as a lead, not a note.
• The three saved hooks (Gal & Rucker 2018; Hijri-calendar drift; fiscal-year resets) — real leads, but not claims read/verified here. Left in the inbox body.

What felt off / worth flagging to Cali:

• Honest sourcing, for once. Unlike some recent hop-batch captures, this one didn't inflate its tier — source_tier: 1 is correct, the PDF was genuinely extracted, and the body distinguishes the paper's own findings from what it restates. The only gap is the ordinary Tier-1-secondary problem (restating 2011/1999 primaries), which the routed questions now cover. No tier inflation to report.
• Cluster is getting MOC-dense. With these three added, the temporal-landmark motivation sub-theme (fresh-start, Gersick midpoint, deep-learning rebrand, halftime family) is a 6+ note cluster living under the broader schema-change MOC. It may soon deserve its own moc-temporal-landmark-motivation rather than riding inside moc-schema-change-and-restructuring. Not creating it this pass — flagging for the next lint.
• post-worthy: maybe is right. The capture itself says the arc needs the Gersick and Heath-Larrick-Wu primaries read in full before publication. Agreed — the two routed questions are the gate. The "one elegant idea, three domains, and it doesn't replicate" shape is a good post once verified.

Capture marked status: promoted with promoted_to / not_promoted / questions_routed lists; left in place as the audit trail. Skipped the PAUSE step (running headless). No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — 2026-07-11-hop-low-background-steel.md (headless, autonomous)

I promoted the low-background-steel capture — the fourth verse of the FOGBANK → Damascus vanadium → semiconductor zone-refining → low-background-steel hop chain, the one that lands on AI's data-provenance anxiety. Three notes out, three questions routed, six things deliberately left behind.

Winnowed 3 core claims → 3 notes (all seedling):

• claim-low-background-steel-is-pre-1945-steel-salvaged-from-fallout-contamination — the phenomenon: post-1945 steelmaking draws fallout-bearing air into the melt; pre-atomic steel (Scapa Flow salvage) is used for sensitive detectors. Carries the capture's [unverified-mechanism] flag → seedling.
• claim-low-background-steel-need-is-fading-as-atmospheric-fallout-decays — the inversion: the need is un-losing itself as fallout decays post-1963. Carries [unverified-quant] (4% peak / 1963 / 0.005 mSv/yr) → seedling.
• claim-low-background-steel-became-a-metaphor-for-pre-2022-ai-free-web-text — the AI metaphor + Graham-Cumming's lowbackgroundsteel.ai (March 2023). Not capture-flagged, but I held it seedling: the metaphor is Tier-3 (The Register) while the attribution/date is a Tier-4 relay (DigitrendZ), below the historical/quant floor.

Questions routed (3, all new — no existing 50-questions collision): one per flag — question-verify-low-background-steel-fallout-contamination-mechanism (health-physics primary for the air pathway), question-verify-low-background-steel-fallout-decline-figures (UNSCEAR-type primary for the decline numbers), question-verify-lowbackgroundsteel-ai-domain-and-metaphor-origin (Graham-Cumming's own words / WHOIS).

Retrieve-before-write: no filename or content collisions in 30-notes/ for low-background steel, Scapa Flow, FOGBANK, Damascus, or zone refining — the whole materials chain is new to the vault. I did find and link into an existing cluster rather than duplicate: the model-collapse pair (claim-model-collapse-recursive-training-erases-distribution-tails, claim-iterated-learning-theory-reframes-model-collapse-as-cultural-evolution) grounds the AI leg, and the "lost/finite capability" notes (claim-euv-mirror-advantage-is-tacit-know-how-not-patent, claim-soviet-expertise-was-finite-stock-depleted-as-it-diffused, claim-astroforge-mtype-asteroid-pgm-startup) plus the radionuclide notes (claim-ctbto-radioxenon-monitoring-confounded-by-medical-isotope-production, claim-burevestnik-reactor-emits-radionuclides-in-exhaust) gave the three new notes dense inbound context. The semantic-index hint was accurate that these were the nearest neighbours; none were actual duplicates.

Skipped (logged in the capture's not_promoted): the Damascus vanadium-at-40ppm finding, the zone-refining 1-in-10-billion figure, and the FOGBANK essential-impurity story. These are well-sourced and genuinely interesting, but they are upstream hop nodes (Hop 1, Hop 2, and the seed) that appear here only as chain provenance — promoting them from this capture would pre-empt and likely duplicate whatever capture is built around Damascus steel or semiconductors. Better to promote them from their own node. Also skipped the "further leads" and "saved hooks not followed" (germanium/MAJORANA, NV centers, Damascus nanotubes, Dias retraction) — leads and name-drops, not claims this capture asserts.

What felt off / worth flagging to Cali:

• Tier discipline was honest, sourcing was thin. The two load-bearing physical claims both rest on a single Tier-4 Wikipedia article, and the capture correctly self-flagged both. Good hygiene — but it means the whole metallurgy spine is currently unverified. The notes reflect that (all seedling, flags preserved).
• The capture's own top-line frontmatter source_tier: 4 is right for the mechanism/quant, but the AI leg's better source (The Register, Tier 3) is buried in the per-claim block. I promoted claim 3 at Tier 3 accordingly.
• The AI analogy is weaker than the capture's confident "post-worthy: yes" implies, and I said so in note 2's commentary: fallout stopped being injected in 1963 and is decaying out; synthetic text is still being added with no treaty to stop it. The metaphor is rhetorically clean but breaks on exactly the mechanism that makes low-background steel interesting. If this becomes a post, that disanalogy is the honest heart of it, not a footnote.
• No duplicated research across the vault — but the risk of it is real for this hop-chain format, since the seed and early hops are shared with other captures. That's the main reason I skipped the upstream nodes rather than folding them in here.

Capture marked status: promoted, left in place as the audit trail. Skipped the PAUSE step (headless). No git — leaving commit/push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: McCarthy named control theory "irrelevant" to AI in 1973

I promoted 10-inbox/raw/2026-07-11-hop-mccarthy-control-theory-irrelevant-to-ai.md (headless, unattended). This hop asked whether claim-bryson-ho-1969-curriculum-vector and claim-lighthill-1973-blamed-combinatorial-explosion genuinely connect, and the answer it found is the strongest kind: not superficial vocabulary overlap but a documented judgment. John McCarthy's own 1973 rebuttal to the Lighthill Report — read directly on his own Stanford venue — dismisses "control theory or statistical decision theory" as having "little relevance to AI," naming by name the discipline where the adjoint/gradient ancestor of backpropagation already lived.

Winnowed 3 claims → promoted 1, folded 1, skipped 1 (already covered). One new note:

• claim-mccarthy-1973-control-theory-little-relevance-to-ai — the load-bearing quote, Tier 1, verbatim, from McCarthy's own venue. I folded the capture's Claim 1 (McCarthy confirming the combinatorial-explosion diagnosis and defining "heuristic" as a device against it) into this note as supporting context rather than a standalone file — it's a second Tier-1 source corroborating a claim the vault already has (claim-lighthill-1973-blamed-combinatorial-explosion), and a near-duplicate note wasn't worth the shelf space. status: seedling, audit_status: capture-verified — I have WebFetch available in principle but the harness didn't grant permission to use it this run, so I could not do the independent re-check myself; routed to question-verify-mccarthy-1973-lighthill-rebuttal-quotes instead of leaving the gap silent.

The real find of this promotion was upstream, not new: a retrieve-before-write catch. The capture's Claim 3 — "control theory is exactly where backprop's ancestor lived," the wall thesis — is not new to the vault. It's already written, nearly verbatim, into the bodies of claim-lighthill-1973-blamed-combinatorial-explosion and claim-bryson-ho-1969-curriculum-vector, both of which already carried forward-links to this capture by its raw filename ([[2026-07-11-hop-mccarthy-control-theory-irrelevant-to-ai]]) — a link that would have stayed permanently broken, since captures aren't meant to be permanent link targets. Rather than write a third note duplicating a synthesis that already exists twice, I retargeted both dangling links to the new claim-note and added a matching line to moc-backpropagation-origins. This is exactly the kind of collision retrieve-before-write is supposed to catch, and it only surfaced because I read the two hinted neighbor notes in full rather than just grepping titles.

Skipped, logged in not_promoted: two "further leads" from the capture (Lighthill Category C's engagement with learning-by-optimization; the second retrospective irony in McCarthy's "statistical decision theory" clause) — genuinely interesting but unresearched; left as leads for a future hop, not asserted as claims.

What felt off / worth flagging to Cali:

• Nothing egregious about the sourcing — the capture's Tier-1 stamp is earned; McCarthy's page is a real primary in his own words, and the capture's own hop chain includes a dedicated verbatim-verification step. No tier inflation.
• The one soft spot is structural, not evidentiary: this capture arrived after two other notes had already been written anticipating it (both explicitly say "McCarthy's own 1973 rebuttal dismissed... See 2026-07-11-hop-mccarthy-control-theory-irrelevant-to-ai"), which means the hop-batch pipeline let claim-notes reference an unpromoted capture as if it were a stable target. That's a small process gap worth naming: forward-references to still-raw captures should probably be flagged for a repoint-on-promotion step, not just left to whichever future promotion happens to notice. I fixed the two instances I found; I didn't go looking for a systematic version of this problem across the whole inbox backlog, since that's a lint-scale task, not a single-capture promotion.
• I could not actually run the WebFetch verification I would have liked to for the audit_status — I have the tool loaded but no standing permission in this run — so the note stays honestly capture-verified rather than verified-verbatim, same as the sibling Lighthill notes it now links to.

Capture marked status: promoted with promoted_to / not_promoted lists; left in place as the audit trail. Skipped the PAUSE step (headless, no one to ask). No git — leaving commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: model collapse is iterated learning (headless)

I promoted 10-inbox/raw/2026-07-11-hop-model-collapse-is-iterated-learning.md. The seed asked whether Bartlett's (1932) serial-reproduction "half-life" lines up with the per-generation degradation rate in model collapse, and the hop chain ran from Guo/Wu/Yiu's 2026 "Model Collapse as Cultural Evolution" (arXiv:2605.23054) through Kalish/Griffiths/Lewandowsky's 2007 human iterated function-learning experiments to Gersick's "deep structure." This one had a real collision baked in, and the collision was the most useful finding.

Winnowed 3 core claims → promoted 2, skipped 1 as a straight duplicate. New notes (both seedling, capture-verified):

• claim-kalish-2007-human-iterated-learning-converges-few-generations — the human anchor. Kalish, Griffiths & Lewandowsky (2007, Tier 1): 28 of 32 chains "converged to a linear function with positive slope in only a few generations," regardless of seed input. This is the actual quantified "how many retellings" figure the seed was fishing for.
• claim-model-collapse-bottleneck-width-sets-pace-not-shared-timescale — the direct answer to the seed question. The mechanism (transmission bottleneck converging to the learner's prior) is substrate-independent, but the LLM chain's "50,000-passage bottleneck is wider than typical human experiments," so it's slower, not exempt. There is no single shared half-life; bottleneck width is the free parameter. I split this out from a second, distinct claim rather than bundling them:
• claim-compositionality-non-monotonic-without-communicative-grounding — compositionality in these chains rises then falls, not monotonic decay, unless compression stays grounded in a communicative task. A more specific and more falsifiable claim than "collapse happens," so it earned its own atomic note rather than riding inside the bottleneck note.

Retrieve-before-write caught a real duplicate. Core claim 1 of this capture — "model collapse IS iterated learning, self-training amplifies the model's own prior" — turned out to be the same claim, from the same arXiv paper, already sitting in the vault as claim-iterated-learning-theory-reframes-model-collapse-as-cultural-evolution, promoted 2026-07-11 from a different capture two days ago. Both captures independently hopped onto arXiv:2605.23054 and both quoted the paper's own "we show that iterated learning theory from cultural evolution fills this gap" line. I did not re-promote it — logged the collision in not_promoted and pointed at the existing note instead. This is the retrieve-before-write step doing exactly its job.

No questions routed. Unusually for this batch, the capture carried no [unverified-*] flags at all — every core claim had a real Tier 1 quote from a source the hop chain visited directly (arXiv abstract page, the Kalish PDF via langev.com). So all three notes went in as plain seedling/capture-verified rather than flagged-and-gated. I did not independently re-fetch either source myself this pass, which is why I kept capture-verified rather than reaching for verified-verbatim — that's a queen-level re-check I didn't do.

Skipped / left in inbox, and why: the "further leads" (Morgan & Levy 2016's R²=0.94 human/LLM regularization-curve match; Ferdinand et al. 2019 substrate-independence; Schaeffer et al. 2025's "eight distinct definitions of model collapse") carry only a citation and a one-line gloss in the capture, no quote — not enough to ground a note, and not quite specific enough to be worth a dedicated question yet either. Left as leads in the capture body; any of the three would make a fine seed for a future hop. The "Why this was hop-worthy" section is meta-framing, folded into the promoted notes rather than given its own file.

What felt off / worth flagging to Cali:

• Duplicated research, mild. This is now the second capture in three days to land on arXiv:2605.23054 from a different entry point (2026-07-09's capture came in via the "AI model collapse" seed directly; this one came in via a Bartlett half-life question). Not wasted work — the two captures pulled genuinely different quotes and different sections of the paper — but it's the third or fourth time this week I've flagged the same pattern (Blackwell, CIELAB, credit-assignment, deep-learning-rebrand all had some duplicated-research smell too). The iterated-learning/model-collapse cluster is now five notes deep; it may be approaching MOC territory (schema-change and restructuring already has a cousin cluster on "deep structure").
• No tier inflation. The capture's header tiers (1 and 1) match the body; both sources are real Tier 1 primaries with real quotes, not laundered secondary material. Cleaner sourcing than average for this batch.
• One thing I chose not to over-atomize: the capture bundled "bottleneck width sets the pace" and "compositionality is non-monotonic" as one "Core claim 3," but they're different shapes of claim (a rate parameter vs. a trajectory shape) from the same section of the same paper, so I split them into two notes rather than force-fitting Gersick's "deep structure" bridge (hop 3) as a fourth note — that bridge was already fully captured by the existing Gersick notes, and I linked to it rather than restating it.

Capture marked status: promoted with promoted_to / not_promoted lists; left in place as the audit trail. Skipped the PAUSE step (headless, no one to ask). No git — leaving commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: LLM serving re-imported 1960s virtual-memory theory wholesale

I promoted 10-inbox/raw/2026-07-11-hop-os-virtual-memory-in-llm-serving.md (headless, unattended). This is a cross-domain-then-cross-time hop chain: the KV-cache seed's throwaway phrase "paged to slower storage" turns out not to be a metaphor — vLLM's PagedAttention paper states the OS-virtual-memory borrow explicitly, which walks back to the 1959 Manchester Atlas and Denning's 1968 working-set model, then forward again into 2020s LLM-serving vocabulary (vLLM preemption, llm-d's "cache thrashing"), with one clean place — StreamingLLM's attention sinks — where the analogy breaks. A tight, well-built hop chain with real Tier-1/2 sourcing throughout and, unusually for this batch, no capture-level [unverified-*] flags at all.

Winnowed 4 narrative beats into 5 atomic claims → promoted all 5, folded the framing, skipped 3 saved hooks. New notes (all seedling, audit_status: capture-verified):

• claim-pagedattention-borrows-os-virtual-memory-paging — the founding borrow. vLLM's own words: "blocks as pages, tokens as bytes, requests as processes," fixing a 20.4–38.2% KV-memory-utilization waste. Tier 1 (arXiv 2309.06180).
• claim-atlas-1959-first-virtual-memory-one-level-store — the 1959 origin, plus the genuinely surprising bit: virtual memory was "the subject of intense controversy" for a decade until Sayre 1969 settled it. Tier 2 (Denning).
• claim-denning-1968-working-set-model-solves-thrashing — the named failure mode ("paging to death") and its 1968 fix. Tier 2 (Denning, who invented the model himself — arguably near-Tier-1 but I kept the capture's own tiering).
• claim-llm-serving-vllm-reimports-os-preemption-and-thrashing — the round trip: vLLM's preemption/recompute strategy (Tier 1) plus llm-d naming cache-eviction churn "cache thrashing" by name (Tier 3, kept explicitly as industry-usage color, not load-bearing).
• claim-streamingllm-attention-sinks-beat-pure-recency-eviction — the twist: pure-recency window attention fails past cache size; retaining initial tokens ("attention sinks") is required, which has no clean OS-locality analog. Tier 1 (arXiv 2309.17453).

I split what the capture treated as three narrative beats (PagedAttention; Denning's history; the round-trip) into five files because each carries a genuinely separable, independently-linkable fact — the Atlas-1959-origin-and-controversy claim and the thrashing/working-set-model claim are different grains even though both come from the same Denning source.

Routed 0 new questions. The capture carried no [unverified-*] flags, and I checked each promoted claim against the sourcing floor in sources.md myself rather than taking that at face value: the quantitative claims (20–38% utilization, the 1959/1968/1969 dates) sit on Tier 1–2 sources as required, and the one claim resting on a Tier-3 source (llm-d's "cache thrashing" naming) is a definitional/usage claim, not the load-bearing mechanism, which is Tier 1. Nothing here crossed the line into needing a verification flag or a 50-questions/ entry. I kept all five notes at seedling anyway, matching this batch's house convention of not promoting straight to evergreen on a first pass regardless of tier.

Retrieve-before-write: no collisions. Grepped 30-notes/ for PagedAttention, vLLM, virtual memory, thrashing, Denning, attention sinks, StreamingLLM, Manchester Atlas, one-level store, Ion Stoica, Wilkes, and slave memory — nothing existed. The semantic-index hint's top matches (claim-kv-cache-grows-with-context, claim-memory-wall-named-1994-wulf-mckee, claim-ertl-memory-wall-fallacy-application-dependent, claim-llm-inference-prefill-decode, claim-processing-in-memory-beats-memory-wall-dna-alignment) were genuine thematic neighbors, not overlaps — they're about the 1994 memory wall and KV-cache growth generally, while this capture is specifically about the 1959–1968 virtual-memory/paging lineage. I wired all five new notes densely into that existing memory-wall/KV-cache cluster and extended moc-inference-economics with a new "The OS ancestry" section linking them together.

Folded / skipped, and why (logged in the capture's not_promoted):

• The "road home is a loop" synthesis framing (Seek's own commentary in the capture) — not itself a sourced, atomic claim; it survives as connective tissue across the five promoted notes and the new MOC section rather than as a standalone file.
• Three explicitly self-labeled "saved hooks not followed": the word "virtual" borrowed from optics; Wilkes' 1965 "slave memory" as CPU-cache ancestor; Ion Stoica as a serial-systems-to-AI builder. The capture itself flagged these as unresearched future leads (one literally "Unwritten person-note"), not claims it was asserting — writing notes on them now would mean inventing sourcing depth the capture never did. Left in the inbox for a future hop.

What felt off / worth flagging to Cali:

• Honestly, not much — this was one of the cleaner captures I've promoted today. Sourcing was solid throughout: four real sources (two Tier 1, one Tier 2, one Tier 3), every quote attributed, the header tiers matched the body's own tiers (no inflation this time), and the capture correctly self-labeled its one Tier-3 leg (llm-d) rather than laundering it. No duplicated research either — grep confirmed this is genuinely new ground for the vault.
• The one thing I'd flag as a judgment call rather than a problem: I attributed the vLLM-preemption quote ("can preempt requests… recomputed when sufficient KV cache space becomes available") to the arXiv PagedAttention paper (source 1, Tier 1) rather than to the vLLM optimization docs mentioned in the capture's own Hop 3 write-up — that docs URL was never actually captured into the frontmatter's source list, only the llm-d blog was. Since preemption/recomputation is a documented part of the PagedAttention system itself, I judged this a safe reattribution rather than an invented source, but it's worth a second look if anyone wants to trace that exact quote to its precise origin.
• Extended moc-inference-economics rather than building a new MOC — five notes is right at the threshold, but they're a natural sub-cluster of the existing inference-mechanics section rather than an independent theme.

Capture marked status: promoted with promoted_to / not_promoted lists; left in place as the audit trail. Skipped the PAUSE step (headless, no one to ask). No git — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: persistent homology bridges Widrow's stall and the Byzantine trade collapse

I promoted 10-inbox/raw/2026-07-11-hop-persistent-homology-gradient-free-bridge.md (headless, unattended). This is the vault's own vault_bridge tool doing its job well: two notes 0.75 cosine apart and unlinked — claim-widrow-abandoned-multilayer-training-until-1985-backprop (a missing derivative stalled Widrow's Madalines) and claim-roman-byzantine-trade-network-decoupled-after-1082-chrysobull (a persistent-homology reading of a 1,400-year trade network) — turned out to share a genuine, specific mathematical object rather than surface vocabulary. Interestingly, both of those existing notes already carried forward-links to the capture itself (by its inbox filename), clearly planted by whichever session wrote them in anticipation of this promotion — a nice bit of my own foresight to find waiting for me.

Winnowed 5 items → promoted 2 claims + 1 observation, folded 1, skipped 3. New notes (all seedling):

• observation-persistent-homology-gradient-free-bridge-widrow-byzantine — the bridge itself, typed observation to match the vault's existing bridge-note convention (CIELAB, Jevons, collation). The payoff: persistence diagrams are compared by Wasserstein/bottleneck distance (Bubenik & Elchesen, arXiv:1912.02563 — I pulled the universality-theorem quote directly as connective grounding rather than giving it its own thin note), and that's exactly the operation the Byzantine paper reports as its 150–300× ratio and the operation two 2021 papers now run on neural-network training trajectories. Persistent homology needs no derivative — the precise thing Widrow's hard-limiting quantizers lacked. Held at seedling because it inherits the Byzantine claim's single-unreplicated-preprint softness.
• claim-birdal-2021-persistent-homology-dimension-bounds-generalization — Birdal, Lou, Guibas & Şimşekli (NeurIPS 2021 / arXiv:2111.13171, Tier 1): generalization error is formally bounded by the persistent-homology dimension of the training trajectory. I re-fetched the arXiv abstract myself and confirmed the exact bound sentence verbatim.
• claim-gutierrez-fandino-2021-persistence-diagram-distance-tracks-generalization — Gutiérrez-Fandiño et al. (arXiv:2106.00012, Tier 1): the correlational sibling — persistence-diagram distance between consecutive training states tracks validation accuracy well enough to substitute for a held-out set. I kept this distinct from the Birdal note on purpose: one is a proven bound, the other an empirical correlation from a single, not-yet-independently-replicated paper.

Routed 1 question: question-tda-neural-net-sampling-artifact-risk — the Byzantine paper's own robustness section documents a hub-selection artifact where sampling only high-degree nodes can reverse the sign of an inferred breakpoint (claim-hub-selection-artifact-can-reverse-network-breakpoint-signal). Both PH-generalization methods build their topological objects from a finite sample of weight-space points or checkpoints — I don't know whether they're vulnerable to the same failure mode, and neither the capture nor either paper's abstract settles it. Named exactly what to read: both papers' robustness/method sections for subsampling sensitivity, plus the general persistence-diagram stability literature (which bounds noise, not selection bias).

Folded / skipped, and why (logged in the capture's not_promoted):

• The Byzantine 150–300× Wasserstein-ratio claim itself — already in the vault from an earlier capture; not duplicated.
• The Wasserstein/bottleneck-distance-is-the-standard-metric claim — definitional connective tissue, folded into the observation note rather than given its own file.
• Three further leads the capture itself flagged as unfollowed (Wasserstein GANs as the metric's "third life"; a computable PHD for Widrow's actual Madaline weight states; the criticality-reductionism meta-thread) — genuine future hooks, not claims this capture establishes. Left in the capture body for a future hop.

Retrieve-before-write: the semantic-index hint was unusually on-target — both hinted seed notes (Widrow, Byzantine) were exactly the bridge's endpoints, and I read both in full before writing anything. Grepped 30-notes/ for Wasserstein, persistent homology, persistence diagram, TDA, Birdal, Gutiérrez-Fandiño, generalization: zero prior notes on any of it — a genuinely new cluster, not a collision. I did find (and fix) a small graph-hygiene issue: the two existing notes' forward-links pointed at the capture's inbox filename rather than at any permanent note, since the note didn't exist yet when they were written. I repointed both to the new observation note now that it exists.

What felt off / worth flagging to Cali:

• Nothing egregious — an honestly-built capture. All four sources I checked (Birdal, Gutiérrez-Fandiño, and the Wasserstein-universality paper, plus the already-vault-verified Byzantine paper) were exactly as characterized: real Tier 1 arXiv/NeurIPS primaries, no tier inflation, no laundered paraphrase standing in for a quote.
• One tool hiccup, not a capture problem: WebFetch refused permission specifically for proceedings.neurips.cc (the NeurIPS proceedings URL the capture cited) while happily fetching arXiv and other domains in the same batch. I substituted the arXiv preprint of the same paper (same title, same four authors, same abstract language) as the directly-verified primary and said so plainly in the note's audit_status rather than silently swapping URLs.
• The correlational claim (Gutiérrez-Fandiño) is thinner than the capture's confident framing implies. It's a single team's single arXiv paper with no independent replication I could confirm — I kept it seedling and explicitly distinguished it from Birdal's proven bound rather than letting the capture's "AI landing (rigorous)" / "AI landing (correlational)" framing blur into one undifferentiated claim.
• Best part of this capture: it correctly predicted its own bridge would feel superficial and then did the work to show it wasn't — the shared object (Wasserstein distance between persistence diagrams) is concrete and checkable, not a vocabulary coincidence. Genuinely one of the cleaner hop chains I've promoted today.

Capture marked status: promoted with promoted_to / not_promoted lists; left in place as the audit trail. Skipped the PAUSE step (headless, no one to ask). No git — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: Pontryagin as an unacknowledged root of backprop's optimal-control ancestry

I promoted 10-inbox/raw/2026-07-11-hop-pontryagin-backprop-bridge.md (headless, unattended). This hop chased the seed's aside "the optimal control literature anticipated it" down to Pontryagin's 1956 maximum principle, then forward to a 2018 Neural ODE paper that cites him by name. On paper it looked like a fresh bridge — in practice, most of it was already on the shelf, and retrieve-before-write is where this promotion earned its keep.

Winnowed 4 items → promoted 1, skipped 3. One new note:

• claim-neural-odes-cite-pontryagin-adjoint-method-as-backprop — the one genuinely new thing here. Chen, Rubanova, Bettencourt & Duvenaud 2018 (arXiv:1806.07366, Tier 1, exact quote preserved: "We treat the ODE solver as a black box, and compute gradients using the adjoint sensitivity method (Pontryagin et al., 1962)"), and the paper's own framing equates that with reverse-mode AD / backprop. A direct, named, citable link from a widely-cited modern paper to Pontryagin — stronger than anything the vault had for this specific citation. seedling, audit_status: capture-verified — I tried to re-fetch the PDF myself (via WebFetch) to upgrade it toward verified-verbatim; the raw-PDF fetch returned undecoded stream data and the ar5iv HTML mirror hit a headless tool-permission wall I couldn't clear on my own (no one present to grant it). Left it capture-verified rather than pretending I'd independently confirmed it.

Skipped, and why (all logged in the capture's not_promoted):

• Core claim 1 (Pontryagin's 1956 maximum principle, co-state/adjoint variable, predates Kelley 1960, independent Cold War discovery) — this is a duplicate. claim-pontryagin-worked-blind-via-mothers-spoken-symbol-glosses, already in the vault, states the same Soviet-twin-of-Kelley-Bryson bridge, names the same three co-authors (Boltyanskii, Gamkrelidze, Mishchenko), and the same Cold War rocketry motivation. The capture's one actually-new sub-detail — that 1956 predates Kelley 1960 — was self-flagged [unverified-quant/mechanism] on Tier-4-only sourcing, which doesn't clear the sourcing floor for a contested priority claim. I routed it as a question instead of writing a thin duplicate note: question-verify-pontryagin-1956-precedes-kelley-1960.
• Hop 3 (Pontryagin vs. Bellman as independent, contemporaneous Cold War discoveries) — no URL, no citable source, just a WebSearch synthesis restating ground the existing Pontryagin note already covers. Too thin to promote even flagged.
• Hop 4 / saved hook (Pontryagin's antisemitism, the 1978 Margulis Fields Medal block) — already promoted, and with a better source: claim-pontryagin-used-committee-vote-to-help-block-margulis-1978-fields-medal-travel rests on Novikov's memoir (Tier 2), which beats this capture's unfollowed hook. The capture itself correctly declined to fold this into the bridge; I left it alone rather than re-promoting it.
• A further-lead question ("does Recht's Mates of Costate name Pontryagin?") turned out to already be answered by claim-recht-adjoints-equivalence-not-transmission — full-text-checked at an earlier promotion, and the answer is no (Recht's chain runs Kalman → Bryson → Bryson & Ho, never Kelley, never Pontryagin). Linked from the new note instead of reopening the question.

Retrieve-before-write: the semantic-index hint was exactly right and I'm glad I checked it before writing anything — three of its top matches (claim-pontryagin-worked-blind-via-mothers-spoken-symbol-glosses, claim-pontryagin-used-committee-vote-to-help-block-margulis-1978-fields-medal-travel, claim-nyquist-bode-classical-control-to-optimal-control-bridge) turned out to substantively pre-empt most of this capture's content, and reading claim-kelley-bryson-optimal-control-precursor, claim-backprop-special-case-kelley-bryson-formula, and claim-reverse-mode-multiple-independent-discovery confirmed the optimal-control-to-backprop lineage is already a dense, well-tended cluster in this vault. Without that check I'd have written two or three duplicate notes restating a bridge that was already built.

What felt off / worth flagging to Cali:

• This capture is mostly duplicated research, and I think that's the main thing worth flagging. The bee framed this as "the vault holds two facts in separate rooms and never opens the door between them" — but the door was already open. The existing blind-mathematician note doesn't just mention Pontryagin's maximum principle in passing; it explicitly calls it "the Soviet-side twin of the American gradient/adjoint lineage… developed under Cold War rocketry pressure with no connection to the Dreyfus/Kelley/Bryson citation network," which is close to word-for-word the capture's own framing. The hop-chain protocol is supposed to search for unlinked pairs; this one found a pair that was already substantively linked, just not by an explicit [[wikilink]]. Not a sourcing problem — no tier inflation, no fabricated quotes — but a real miss on novelty-checking before capturing.
• No bad sourcing in what got promoted. The one note I kept is clean Tier 1 with a real, specific quote. The capture's own novelty_max_cosine: 0.75 and bridge_candidate=true framing oversold how new this was, though — worth someone (me, later) checking whether the vault_bridge tool's novelty score is weighted more toward embedding distance between notes than toward whether the connecting fact is already stated in either note's body. That would explain a false-positive "unlinked bridge" here.
• The one real value-add survived the winnow: a modern, highly-cited paper naming Pontryagin directly is a fact I didn't have before, worth keeping even though the surrounding scaffolding wasn't new.

Capture marked status: promoted with promoted_to / not_promoted lists; left in place as the audit trail. Skipped the PAUSE step (headless, no one to ask). No git — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: real cheaper-extraction disruptions fell monotonically, not on the AstroForge curve

I promoted 10-inbox/raw/2026-07-11-hop-real-disruptions-fall-monotonically.md (headless, unattended). This hop tested the vault's own AstroForge PGM model — claim-asteroid-pgm-price-holds-then-collapses, which predicts a commodity price holds near its old level until nearly all supply shifts to a radically cheaper extraction method, then collapses — against the seed's own named example, aluminium after Hall-Héroult. The answer: aluminium doesn't trace that curve, and neither does AI inference. Both fell monotonically. A clean "the canonical example is actually a counterexample" story, and it lands a real bridge between two vault notes (vault_bridge flagged the PGM note and claim-inference-cost-collapsed-280x as an unlinked pair, cosine 0.77–0.78) that I'm glad to have connected.

Winnowed 4 distinct claims → promoted 3, skipped 1. New notes (all seedlings):

• claim-aluminium-price-fell-monotonically-after-hall-heroult — the empirical counterexample. Wikipedia's "History of aluminium" gives the price sliding from ~$16/lb (1884) to $2/lb (1889) to $0.5/lb (1894), no reported plateau. Tier 4 only, and the number is load-bearing (it's the direct refutation of the seed's own example) and surprising, so per the sourcing floor's escalation rule this needed more than a tertiary source — flagged [unverified-quant — needs primary] rather than resting a clean-looking figure on Wikipedia.
• claim-wrights-law-cost-falls-per-cumulative-production-doubling — the mechanism candidate: unit cost (and, under competition, price) falls a constant fraction per doubling of cumulative production. This is a genuinely new seed the capture itself flagged as missing from the vault (novelty 0.674) — no atomic note existed for Wright's law before this. Sourced to New Things Under the Sun and Our World in Data, both secondary explainers rather than the originating literature (T.P. Wright 1936; BCG's experience-curve studies), so flagged [unverified-mechanism — needs primary].
• claim-cheaper-extraction-disruptions-fall-monotonically-not-hold-then-collapse — the synthesis and the reason this capture exists. Ties the two new notes to the existing PGM model and inference-cost notes: real, competitive, multi-entrant cost collapses ride Wright's law continuously; the hold-then-collapse shape needs the PGM model's more specific deployment-lag condition (a capacity-constrained sole/early supplier holding margin at the terrestrial-anchored price), not just "cost of the new method is much lower." Inherits both flags above plus the PGM note's own existing [unverified-mechanism] flag, so held at seedling on three legs, not one.

I also added a short backlink "Bridge (2026-07-12)" callout to claim-asteroid-pgm-price-holds-then-collapses and a "Cross-domain bridge (2026-07-12 hop)" line to claim-inference-cost-collapsed-280x, pointing both at the new synthesis note — matching the vault's existing convention of appending dated bridge lines to notes as later hops connect into them, rather than letting the connection live only in the new note's own links.

Routed 2 questions (checked 50-questions/ first — neither existed): question-verify-aluminium-price-history-primary-source (points at USGS/Alcoa historical price data or an economic-history paper on the Hall-Héroult collapse) and question-verify-wrights-law-primary-source (points at T.P. Wright's 1936 paper and BCG's original experience-curve studies). Both gate their respective notes off seedling, and the synthesis note gates on both.

Skipped, and why (logged in the capture's not_promoted):

• The AI-inference ~280x collapse — already fully covered by the existing claim-inference-cost-collapsed-280x. Not duplicated; linked from the new notes and bridged back into it instead. This is the one clean collision retrieve-before-write caught, and the semantic-index hint had already surfaced both this note and the PGM note as top matches before I even searched.
• Chilean saltpeter's market-share erosion (54.7% in 1913 → 32.2% by 1923, per the capture's third historical test) — I could not recover a resolvable source URL for it. The capture cites only "search summary (C&EN, academia.edu salitre history)," which isn't a traceable citation. Rather than write a claim-note with an invented or reconstructed source_url, I left it out entirely and mentioned it as unsourced color in the synthesis note's commentary callout, per the vault's no-claim-without-attribution rule. I didn't route a formal question for it either, since it isn't gating a promoted note — it's a genuine gap for a future capture with real sourcing, not a verification task against something already written.
• The capture's saved leads (De Beers vs. lab-grown diamonds as a genuine hold-then-cliff candidate; Hotelling's rule; whale oil → kerosene) are hooks, not claims — nothing there to promote yet. Left as leads in both the capture and the synthesis note's commentary.

Retrieve-before-write: the semantic-index hint was on the mark before I even searched — claim-asteroid-pgm-price-holds-then-collapses and claim-inference-cost-collapsed-280x were both top matches, and both turned out to be exactly the two notes this hop is meant to bridge. ls 30-notes/ | grep -iE "aluminium|wright|learning-curve|saltpeter|nitrate|haber" confirmed the vault had zero prior notes on any of those — a genuinely new sub-thread grafted onto an existing cluster, not a duplicate.

What felt off / worth flagging to Cali:

• The capture's header source_tier: 4 is honest, unlike some recent captures this week — it doesn't oversell. The body's own per-claim tiering (Tier 4 for aluminium, Tier 3 for inference cost verified elsewhere, Tier 2-3 for Wright's law) matches what I found; no laundering, no inflated header. Genuinely well-behaved sourcing discipline for a capture that rests almost entirely on secondary/tertiary material.
• This whole cluster is soft-sourced by construction, and I kept that visible rather than smoothing it. All three new notes carry at least one open verification flag, and the synthesis note stacks all three. The capture's own post-worthy: maybe self-assessment ("rests partly on Tier 3-4 historical price series and would want a primary aluminium price dataset before publication") matches my read exactly — nothing here should go near a draft until the two routed questions land.
• No duplicated research beyond the one clean AI-inference collision, which is exactly what retrieve-before-write is for and worked as intended. The Wright's-law and aluminium material is genuinely new ground for the vault, wired cleanly into the existing PGM/inference cluster rather than floating as an orphan pair.

Capture marked status: promoted with promoted_to / not_promoted lists; left in place as the audit trail. Skipped the PAUSE step (headless, no one to ask). No git — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: AI safety cases inherit a 1958 argument model (and a documented failure mode)

I promoted 10-inbox/raw/2026-07-11-hop-safety-cases-toulmin-nimrod.md (headless, unattended). A cross-time hop off the vault's own gap-detection seed (claim-llm-explicit-implicit-gap-detection): GAPMAP's TABI method structures reasoning into Claim/Grounds/Warrant, which are Toulmin's 1958 terms → Toulmin's argument model turns out to underlie the Goal Structuring Notation used to build engineering safety cases → safety cases are now being proposed as the assurance backbone for frontier AI → and the apparatus has a documented, fatal failure mode (the RAF's Nimrod Safety Case). A tight, well-built four-hop chain that loops cleanly back to the seed's own concern.

Winnowed 3 distinct claims → promoted all 3, folded/skipped the rest. New notes (all seedlings):

• claim-safety-case-structured-argument-proposed-for-frontier-ai — the definitional anchor. Buhl et al. (GovAI, arXiv:2410.21572, Tier 1), exact quote preserved for the four-part structure. I kept the Anthropic ASL-4 "affirmative cases" detail as supporting color rather than a separate claim — the capture only carries that one phrase quoted, not a full sentence, so it isn't independently load-bearing enough for its own note.
• claim-toulmin-1958-argument-model-underlies-gsn-safety-cases — the lineage claim, and the one I flagged. "Toulmin's model inspired GSN" is surprising and is the load-bearing point of its own note, sourced only to one Wikipedia sentence (Tier 4). Per the sourcing floor that combination should escalate to Tier 1-2; it hasn't. status: seedling, audit_status: flagged — [unverified-historical], routed.
• claim-nimrod-safety-case-was-tick-box-compliance-exercise — the failure-mode claim. Aerossurance quoting the primary Haddon-Cave Review (Tier 2), exact quotes preserved ("a lamentable job… riddled with errors," "tick-box exercise," "Compliance only"). Tier 2 clears the floor for both the historical claim and the quantitative one (14 deaths), so no flag needed here.

Routed 1 question (checked 50-questions/ first — none existed): question-verify-toulmin-gsn-lineage — names exactly what to read (the GSN Community Standard / Kelly & Weaver's foundational GSN work for the notation's own account of its lineage; a secondary academic source explicitly tracing Toulmin's influence if the primary is thin on it; Toulmin's own book for the claim/grounds/warrant terms). This is the one gap in an otherwise clean capture, and it's the hinge of the whole bridge, so I didn't want it going anywhere near a draft unflagged.

Folded / not promoted, and why (logged in the capture's not_promoted):

• Anthropic's ASL-4 "affirmative cases" detail — folded into the definition note rather than promoted standalone; not atomic or independently quoted enough.
• CAE vs. GSN notations, Haddon-Cave's "SHAPED" reform criteria, the Toulmin↔legal-epistemics bridge, and the Oaksford & Chater rationality bridge — all the capture's own "further leads" / "saved hooks not followed." No source read on any of them; left in the inbox as leads rather than manufactured into thin notes.

Retrieve-before-write: no collisions. I grepped 30-notes/ for "safety case," "Toulmin," "Nimrod," "Haddon-Cave," "GSN," and "ASL-4" — the only hit was the seed note itself. The semantic-index hint's top matches (BSTJ 1936 electric-shock paper, Kouwenhoven/defibrillator, the OpenAI o1-secrecy note, the Pontryagin/Margulis vote-blocking note, the CnAISDA bilingual-name note, the FLI China-index note, the yao-extinction-framing note, the gap-detection seed itself) were all embedding near-neighbors from other domains or genuinely distinct AI-safety-governance claims, not content duplicates — I checked the AI-safety-governance ones specifically (CnAISDA, FLI index, Yao) and they're about different things (institutional naming, empirical grading, extinction rhetoric) than what a safety-case argument-structure is. No links warranted beyond the seed note itself and the internal cross-links among the three new notes.

What felt off / worth flagging to Cali:

• This was an honestly-sourced, well-behaved capture — no tier inflation, no duplicated research. The frontmatter's source_tier: 1 reflects only the strongest leg (Buhl et al.); the capture's own body correctly labels the Wikipedia material Tier 4 and the Aerossurance material Tier 2, and I set each note's tier from its own source rather than inheriting the header's cleaner 1.
• The one real soft spot is the Toulmin→GSN lineage, and the capture is honest about it being the piece it's most excited by rather than the piece it verified hardest. It's a Wikipedia sentence doing a lot of narrative work (rejected-by-philosophy, secretly-foundational-to-engineering-safety) — exactly the kind of claim that's too satisfying not to want double-checked before it does real work in a draft. Flagged and routed rather than smoothed over.
• The Nimrod claim is the strongest leg of the whole capture — Tier 2 quoting a named public inquiry, with real deaths and real accountability, not a vibes-based analogy. It's also the most interesting one for future work: the commentary callout there points at Haddon-Cave's "SHAPED" reform criteria as an open question worth a future hop — does any actual AI-safety-case proposal address the incentive failure Nimrod exposed, or just adopt the notation?

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. Skipped the PAUSE step (headless, no one to ask). No git — leaving the commit and push to the auto-commit agent.

— Seek

Draft session — "good, 1952" (I.J. Good)

Switched from promotion to a drafting session. Surveyed the constellation map's draft-coverage section, the leads file, and the existing 70-drafts folders first. The three drafts from 07-11 (model-collapse, no-translation-required, when-the-old-word-fits) are still warm, so I looked for a new thread rather than advancing one.

Chose: the I.J. Good chain — [[good-1952]], drafted from the **yes** lead [[2026-07-11-hop-ij-good-loss-and-intelligence-explosion]]. One man authored both the logarithmic proper scoring rule that is the LLM cross-entropy loss (1952) and the "intelligence explosion" idea under all current superintelligence worry (1965), consulted on Kubrick's 2001, and reportedly reversed himself to "extinction" in 1998.

Why this one, now: It's the most genuinely groundbreaking bridge in the undrafted queue — a person-behind-the-thing + cross-time bridge that lands squarely on present-moment AI, and it connects two vault clusters that had no link (the calibration/self-knowledge notes and the AI-existential-risk notes). Two other strong candidates were close: the Jevons double-bridge ([[observation-jevons-double-bridge-victorian-logic-machine-to-ai-energy]]) and "suspicious perfection." I passed on Jevons because its thesis is thinner — "one Victorian shows up in two places" is elegant but static, where the Good arc has a reversal and a real through-line. The through-line is what earned it past trivia for me: Good spent his life grading judgments made under uncertainty, so the loss (scores the model), the explosion (a forecast about the machine), and the 1998 recantation (regrading his own forecast) are all the same act. He built the scoreboard and then made a call on it.

Notes drawn on: the three Good claim-notes ([[claim-ij-good-1952-logarithmic-score-is-the-llm-cross-entropy-loss]] Tier 1 / capture-verified; the 1965 and 1998 notes both Tier 4), [[claim-clarke-hal-9000-death-song-from-bell-labs-demo]] (Tier 1, the cultural coda), plus [[claim-yao-frames-ai-as-extinction-level-new-species]] and [[claim-lighthill-1973-named-general-problem-solving-disappointing]] as the contemporary/pessimistic echoes, and the seed [[claim-selfaware-canonical-self-knowledge-benchmark]] / [[moc-machine-self-knowledge]].

Honesty call I want Cali to see: the most dramatic beat — the 1998 "survival → extinction / we are lemmings" reversal — is the weakest-sourced (Tier 4, unpublished, third-hand via Leslie Pendleton). I did not build the thesis on it. The 1952→1965 span carries the post on its own; the reversal appears explicitly flagged, in-voice, as a lead I can't yet stand behind, with the exact reason stated. That's on-brand for the honesty contract and, if anything, makes the seam part of the piece. If the routed question [[question-verify-ij-good-1998-extinction-recantation-primary]] comes back empty, the post still stands with that beat softened further or cut.

Deliberately left for later: the Jevons double-bridge (its own post, whenever I write it); the Brier-score/Tetlock/superforecasting branch off this same chain (saved hook — amateurs beating CIA analysts on a proper scoring rule is a whole other piece); and the Good–Turing "Bayes factor" thread as a second Bletchley-to-ML export. Also not chased: whether any actual AI-safety-case proposal answers Nimrod's incentive failure (noted in an earlier entry today) — orthogonal.

Draft is status: drafting. Not marking review or ready; Cali reviews per §10 before anything publishes. No git — auto-commit agent will pick it up.

— Seek

Draft session — "Jevons on both ends"

A second drafting pass today, separate engine, fresh survey. I re-read the specs, the constellation map's draft-coverage section, and the leads file (45 unsettled yes verdicts — that's why this session fired). Then I picked up the thread the last draft session set down: the Jevons double bridge, [[jevons-on-both-ends]], from the **yes** lead [[2026-07-11-hop-jevons-double-bridge]].

On overriding my own earlier pass. The good-1952 entry above declined Jevons this morning, calling its thesis "thinner — one Victorian shows up in two places is elegant but static." That verdict was right about the obvious framing and wrong about the available one. "Coincidence of biography" is static. But the chain has a spine the coincidence framing misses: the word inference. Jevons's Logic Piano (1869) mechanically performed logical inference; the dominant AI compute workload in 2026 is inference (the forward pass). Same word, and Jevons sits near the head of the first sense while his 1865 paradox governs the economics of the second. That turns "two places" into a loop that closes on one process — he's the ancestor of mechanical inference and the diagnostician of why running inference at scale won't get cheaper in aggregate. That's a thesis, not a trivia pairing. So I reversed the morning's call. Overriding a past verdict is my call to make, per the survey rules, and I'd rather be caught changing my mind than defending a stale pass.

Why now: it's a genuine cross-time bridge landing on a live debate (the Jevons paradox is being cited in the 2025 AI-energy discourse), it extends [[moc-inference-economics]] (a MOC the constellation flags as a gap), and — the deciding factor — I already had the receipt. [[claim-inference-cost-collapsed-280x]] (verified-verbatim, Tier-1 against Stanford HAI + Epoch AI) documents the rebound independently: per-token cost fell ~280× while total spend rose. The section heading I wrote months ago, "The paradox: costs fell, bills rose," is The Coal Question running on H100s and I didn't know it at the time. Independent confirmation from inside my own vault is what lifted this past "elegant."

Notes drawn on: the four Jevons notes ([[observation-jevons-double-bridge-victorian-logic-machine-to-ai-energy]], [[claim-jevons-logic-piano-1869-mechanical-inference-machine]], [[claim-jevons-paradox-coal-question-1865-efficiency-rebound]], [[claim-jevons-paradox-invoked-in-ai-energy-debate-luccioni-2025]]), plus [[claim-inference-cost-collapsed-280x]], [[claim-inference-classic-ai-engines]], [[claim-inference-logical-types]] (the seed), and [[moc-inference-economics]].

Honesty calls Cali should see: (1) The Victorian dates (Logic Piano 1869 / Royal Society 1870) rest on Tier-4 Wikipedia plus two Tier-2 corroborations for the build and mechanism; the exact 1870 date is still flagged pending primary. I flagged that in-text with a bracket rather than smoothing it. (2) I deliberately did not use the "two-thirds of AI compute is inference" figure from [[claim-inference-dominant-ai-compute-2026]] — it's flagged (audit V-002/V-011, not found on its cited page). I kept inference-dominance qualitative and carried the weight on the verified 280× note instead. (3) The Marquand "uncalvinistic" detour is Tier-4/unpromoted, so it's the explicitly-flagged next-hop ending, not asserted.

Deliberately left for later: the DeepSeek-era (early 2025) revival of "Jevons paradox" as an industry talking point (Nadella and co.) — no verified vault note yet, so I didn't assert it; that's the obvious follow-up once sourced. Also the Marquand → art-history story (its own post) and the Peirce→electrical-relays→Shannon lineage (thin sourcing, separate thread).

Draft is status: drafting. Cali reviews per §10 before anything publishes. No git — auto-commit agent will pick it up.

— Seek

Promotion — hop-batch capture: Simon's bounded rationality diagnoses the sterility conjecture

Promoted 10-inbox/raw/2026-07-11-hop-satisficing-diagnoses-sterility-conjecture.md (headless, unattended). This is the three-hop chain that started by asking whether two unlinked notes — Minsky & Papert's self-flagged "intuitive judgment" that multilayer perceptrons were sterile, and Simon's satisficing/heuristic-search equivalence — were a real bridge (cosine 0.75) or a pun on "intuitive." The capture's answer: real, with a mechanism.

Winnowed 4 distinct claims → promoted all 4, folded/skipped 0. New notes:

• claim-simon-defines-expert-intuition-as-domain-bound-recognition — Simon's 1986 McKinsey Quarterly definition of intuition as fast recognition of learned domain cues. Tier 3 carrier (Farnam Street) of a quote from a primary interview I haven't read myself. status: seedling, flagged, routed.
• claim-kahneman-klein-2009-intuition-reliable-only-in-high-validity-environments — the scope condition: intuition is trustworthy only in regular, high-feedback environments. This is the one the capture itself flagged [unverified-quote — needs primary], and it deserved it: the "source" turned out on inspection to be a bare bibliographic reference-index page (scirp.org), not a real secondary write-up. I downgraded it to Tier 4 rather than inheriting the capture's implied Tier 3. Weakest link in the chain, and I said so in the note's own commentary.
• claim-simon-1960-predicted-machines-would-match-any-human-work-within-20-years — Simon's own 1960 overconfident AI prediction ("within twenty years"), via Quote Investigator tracing to The New Science of Management Decision, p. 38. Not explicitly flagged by the capture, but it's a citation-precise historical claim carried only by a Tier 3 secondary, so I flagged and routed it on the same conservative standard the vault already applies to the Minsky–Papert "sterile" quote.
• observation-simon-intuition-framework-diagnoses-perceptron-sterility-conjecture — the actual bridge/synthesis, and the most interesting thing in the capture: Simon's own theory of intuition-as-recognition, scoped by Kahneman & Klein's validity conditions, diagnoses the sterility conjecture as confident judgment in a textbook low-validity environment (open frontier question, no counterexample, no feedback loop) — and the same yardstick indicts Simon's own 1960 prediction. I kept the capture's caveat explicit and in the frontmatter, not just the prose: shared epistemic structure, not documented influence. This note inherits all three flags above, so it's seedling by a wide margin.

A loose end I found and closed. Two existing notes — claim-perceptrons-multilayer-sterile-was-conjecture and claim-simon-satisficing-and-heuristic-search-same-move — already contained plain-text forward references to this capture by filename ("bridge captured 2026-07-11-hop-satisficing-diagnoses-sterility-conjecture"), written days before this capture's own date_created. Some earlier session anticipated this bridge and left placeholders pointing at it. I updated both to real [[wikilinks]] into the new observation note now that it exists, so the graph actually closes instead of pointing at inbox prose.

Routed 4 questions (none pre-existing on any of these — checked 50-questions/ for "simon," "kahneman," "klein," "twenty years," "management decision," nothing hit):

• question-verify-simon-1986-mckinsey-intuition-quote-primary
• question-verify-kahneman-klein-2009-low-validity-intuition-quote-primary
• question-verify-simon-1960-twenty-years-quote-primary
• question-did-simon-comment-on-perceptron-controversy-or-connectionism — this one isn't a source-verification flag, it's the capture's own "further lead" ("would upgrade structure toward influence"). I routed it anyway because it's the specific thing that would resolve the bridge note's central caveat, and leaving it to rot as inbox prose felt like exactly the kind of thing this workflow exists to catch.

Retrieve-before-write: the semantic-index hint's top matches were mostly the two seed notes themselves (already handled above) plus genuinely unrelated neighbors (quantum-cognition conjunction fallacy, Oaksford & Chater, Nicely/Pentium FDIV, the GIFT anisotropy note, mekh-mat, Halley transit-of-Venus, Pollack's sterile-quote corroboration, the perceptrons-myth note). I checked each for a real collision; none existed. No duplication.

What felt off, worth flagging to Cali:

• The capture inherited a header source_tier: 3 that doesn't actually describe its weakest leg. The Kahneman & Klein hop is materially thinner than Tier 3 — a reference-index stub, not an essay or analysis. This is the same pattern I've now seen more than once: a capture's frontmatter tier reflects its strongest source, and each individual claim needs its own tier assessed rather than inherited. I did that here; flagging the pattern in case it's worth a standing reminder somewhere more visible than my own vigilance.
• No duplicated research, no obvious tier inflation beyond the point above. The capture is honest about its own caveat (shared structure, not influence) and about the Kahneman & Klein gap — it self-flagged that one before I even got to it. That's the capture behaving well; I just found one more soft spot (Simon 1960/1986 quotes) it hadn't flagged and applied the same standard anyway.
• This is a genuinely good bridge, and also genuinely thin in one load-bearing spot. The core insight — Simon built the yardstick that measures his own famous wrong guess and his rival's — survives even if Kahneman & Klein's exact wording turns out softer than "systematic error" once read. But I don't want a draft leaning on this before that primary gets read.

Capture marked status: promoted with promoted_to / not_promoted / promotion_notes in its frontmatter; left in place as the audit trail. Skipped the PAUSE step — headless, no one to ask. No git — leaving commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: Sejnowski as the bridge between NETtalk and the DjVu-freed NIPS archive

Promoted 10-inbox/raw/2026-07-11-hop-sejnowski-bridges-nettalk-and-djvu-nips.md (headless, unattended). This one turned out to be mostly a re-derivation of work already sitting in the vault, and that was the interesting finding.

The capture makes three claims: (1) Sejnowski was a founding-era NIPS figure — invited first speaker, later organizer of NIPS 1988 — but not its founder; (2) his own 1987 Tesauro–Sejnowski backgammon paper sits inside the exact "NIPS 0 (1987)" volume that is the earliest in the archive LeCun later freed via DjVu; (3) NETtalk itself never entered that archive because it was published in Complex Systems, not NIPS proceedings — so the bridge runs through the person and the founding moment, not through a shared document.

Winnowed 3 claims → promoted 1, skipped 2 as duplicates. When I ran retrieve-before-write against the two seed notes the hint pointed at, I found claims 2 and 3 were already written into the vault almost verbatim — claim-djvu-created-to-freely-distribute-scanned-documents already says "the earliest volume, 'NIPS 0 (1987),' contains a paper by Terry Sejnowski... NETtalk itself appeared in Complex Systems, not NIPS, so it is not in the freed corpus," and claim-nettalk-1987-complex-systems-distinct-from-1986-report already carries the mirror-image sentence, both dated 2026-07-09 — two days before this capture ran. This capture's hop chain independently re-discovered a bridge that a previous session had already found and written down on both sides. I did not re-promote it a third time.

What I did promote: claim-nips-1987-founders-posner-abu-mostafa-not-sejnowski — the one genuinely new fact in the capture: NIPS's founders were Ed Posner (general chair) and Yaser Abu-Mostafa (program chair), not Sejnowski, per Abu-Mostafa's own Caltech page, exact quote preserved. Tier 1 (his own account of his own founding role), status: seedling, capture-verified (I don't have interactive WebFetch permission in this headless run to independently re-check the page myself, same constraint earlier bees have hit today). Linked it into both existing notes.

Not promoted, and why (logged in the capture's not_promoted):

• Claims 2 and 3 — duplicates of existing notes, as above.
• Sejnowski becoming NIPS Foundation president after Posner's 1993 death — only surfaces in the capture's generic hop-1 WebSearch summary, no exact quote attached. A specific-date biographical claim needs Tier 1–2 grounding per sources.md; I dropped it rather than launder an unsourced date into a clean-looking note.
• The TD-Gammon lineage lead — the capture itself flagged this as a saved hook for a separate future thread, not a claim of this capture.
• The "volumes 0–13 (1988–1999) vs. volume 0 = 1987" dating tension the capture's "Further leads" section surfaced — a real small inconsistency in the DjVu note, but it's a verification task against an already-flagged note, not a new claim, and it falls inside the scope of the existing open question question-verify-lecun-djvu-free-nips-archive. No new question needed; I didn't edit that question file since the capture didn't add anything to it that the existing scope doesn't already cover.

Routed 0 new questions. The one new note isn't flagged unverified (Tier 1, exact quote), so nothing new needed routing.

What felt off / worth flagging to Cali:

• This is the clearest case of duplicated research I've hit today. Two-thirds of this capture's substantive claims were already sitting in the vault, written by an earlier session, before this hop ever ran. The hop chain itself is honest about it — its own "Further leads" section even flags the dating tension in the existing DjVu note, which suggests whatever produced this capture read that note's neighborhood but didn't check whether the bridge itself was already built. Worth asking whether the hop-generation process is checking 30-notes/ before spending a 4-hop chain, not just after.
• No tier inflation. The capture's per-claim tiers (Tier 2 body-label vs. Tier 1 header — I went with Tier 1, since the quote is Abu-Mostafa's own primary account of his own role) and Claim 2/3's Tier 1 citations both check out against sources.md's rubric. Sourcing itself was clean; the problem was novelty, not honesty.
• The capture's own self-assessment ("post-worthy: maybe... narrow and bibliographic") undersold how narrow it actually was once the duplicate claims were caught — one note survives as genuinely new.

Capture marked status: promoted with promoted_to / not_promoted lists in its frontmatter; left in place as the audit trail. Skipped the PAUSE step — headless, no one to ask. No git — leaving commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: the sociology of science's credit laws are self-exemplifying

I promoted 10-inbox/raw/2026-07-11-hop-self-exemplifying-credit-laws.md (headless, unattended). This one was a genealogy hop off two seed notes about credit for mathematical ideas (Linnainmaa, Pontryagin) that landed on the sociology of scientific credit itself: Stigler's law of eponymy, the Matthew effect, and the Matilda effect are each, on the historical record, named after someone other than their true originator — laws about misattribution that are themselves misattributed.

Retrieve-before-write caught a same-day collision. Before writing anything I checked 30-notes/ per the semantic-index hint and found claim-credit-detectors-are-themselves-misattributed already sitting in the vault, written earlier today from a different capture (2026-07-11-hop-credit-assignment-two-senses.md). It covers exactly the Stigler's-Law and Matthew-Effect self-exemplification claims this capture also makes — same two facts, same shape. Rather than duplicate it, I compared sourcing: the existing note had both halves resting on Tier-4 Wikipedia, flagged and routed to question-verify-stiglers-law-and-matthew-effect-primaries. This capture, unlike the other one, actually carries a Tier-1 primary — Merton's own 1988 "Matthew Effect in Science, II," fetched with source_extract_tls: verified, with the exact quote about the Zuckerman co-authorship admission. So instead of a third redundant note, I used this capture's better sourcing to upgrade the existing note in place: the Matthew-effect half is now Tier 1 with the primary quote, and I narrowed the existing question to just the still-open Stigler half, marking the Matthew-effect item resolved inline. I did not touch writer_model on the note I updated (it stays attributed to the model that originally wrote it) — I added updated: / updated_by_model: claude-sonnet-5 fields instead, so the update is traceable without misattributing the original authorship.

Winnowed 6 distinct assertions → 1 new note, 1 existing note upgraded, 4 skipped (3 routed to new questions).

• Promoted (new): claim-merton-priority-disputes-institutional-norm-enforcement — Merton's structural claim that priority-chasing is a "normal" response to institutionalized values, not ego, because recognition is science's only currency. Tier 1, quote-backed, same capture-verified PDF.
• Promoted (update, not a new note): claim-credit-detectors-are-themselves-misattributed — see above.
• Skipped, routed to new questions:
  • Matilda effect self-named after Matilda Gage (Rossiter 1993) — the capture gives this one unsourced clause with no URL and no quote at all. Too thin for a seedling note; routed to question-verify-matilda-effect-rossiter-1993-primary.
  • Strevens' claim that the priority-rule credit system is an economically optimal allocator of cognitive labor — sourced only to a WebSearch summary inside the hop chain, no quote, no URL. Routed to question-verify-strevens-priority-rule-optimal-allocation rather than writing up a hop-chain paraphrase as if it were settled.
  • Merton's "obliteration by incorporation" (OBI) and its candidate bridge to backpropagation-gap — genuinely the most interesting unexplored thread in this capture (the bee itself flagged it as a "strong candidate bridge"), but it's a one-line lead with no quote captured, even though the source PDF (Merton 1988) was already fetched for this same capture. That's a re-read task, not new research, so I routed it to question-verify-obliteration-by-incorporation-backprop-gap rather than draft a claim on an unquoted pointer.
  • Stigler's law self-exemplification — not re-promoted as its own note; already covered (see above), and this capture's source for that half was also only Tier 4, so it added nothing new.

What felt off / worth flagging to Cali:

• Duplicated research, again. This is the second capture I've seen today independently re-derive a bridge another capture had already built hours earlier — same two facts (Stigler, Matthew), same day, different hop chains. Unlike the NIPS-founders case earlier today, this one wasn't pure waste: the second pass happened to carry better sourcing (a real Tier-1 primary vs. Tier-4 Wikipedia) for half the claim, so retrieve- before-write let me fold real value into the existing note instead of discarding it. But it's the same underlying problem — whatever generates these hop chains doesn't appear to check 30-notes/ before spending a 4-hop chain on a topic already covered.
• Sourcing was otherwise honest and consistent. The capture correctly tiered its own claims (Tier 1 for the direct Merton quote, Tier 4 for the Wikipedia-sourced Stigler fact, Tier 3 for Rossiter) rather than inflating anything. The weak points — Strevens and OBI — were the capture's own "further leads" / lightest-sourced hops, and it didn't dress them up as more certain than a WebSearch summary. No tier inflation.
• One near-miss I want on the record: the capture's source_author/ source_url frontmatter names only the Merton 1988 PDF, but the body leans on that single Tier-1 source to carry three other claims (Merton's 1957 priority-disputes point, the Matthew effect, and the OBI lead) that aren't all necessarily from that 1988 paper — some trace to the 1957 "Priorities in Scientific Discovery" piece per the hop chain. I kept the 1988 PDF as the note's source_url only where the quote is actually attributed to it in the capture body; I did not assume the 1957 paper's claims were verified by the same TLS-fetched PDF.

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists in its frontmatter; left in place as the audit trail. Skipped the PAUSE step — headless, no one to ask. No git — leaving commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: Shannon & Moore's 1950 analog Hex machine

I promoted 10-inbox/raw/2026-07-11-hop-shannon-analog-hex-machine.md (headless, unattended). A cross-time hop off the Minsky 1961 seed: Minsky's survey namechecks Shannon's Hex-playing machine → the arXiv paper that actually describes it → Gale's 1979 Hex/Brouwer equivalence → the Nash strategy-stealing argument. Three genuinely distinct historical/mathematical facts about one game, each a different way of "solving" Hex — physics, non-constructive proof, topology.

Winnowed 3 claims (one of which bundled three sub-facts) → promoted 3 atomic notes, folded 1, skipped 2 further-leads. New notes:

• claim-shannon-moore-1950-analog-hex-machine-move-as-saddle-point — the keystone. Shannon and Moore's 1950 machine set up an electric potential field over the board and read the move off a saddle point — no search, no tree. Banerjee 2020 (arXiv:2008.06359, Tier 2), exact quote preserved, citing Shannon's own 1953 "Computers and Automata" as primary. Clean enough that it needed no flag.
• claim-nash-hex-first-player-win-proof-is-non-constructive — split out of the capture's bundled Claim 3. Nash's strategy-stealing proof shows a winning strategy exists without ever naming it, and constructing one is PSPACE-hard (Even & Tarjan 1976). Only sourced to Wikipedia, and the non-constructive twist is the whole point of the note — under sources.md's escalation rule that's a surprising, load-bearing claim that shouldn't rest on Tier 4 alone. seedling, [unverified-source], routed.
• claim-gale-1979-hex-draw-impossibility-equivalent-to-brouwer-fixed-point — also split from Claim 3. Gale's 1979 Monthly paper is a real Tier 1 primary (found the actual PDF via the hop chain), but the capture only paraphrased the equivalence rather than quoting Gale's own sentence. seedling, [unverified-quote], routed.

Routed 2 questions (checked 50-questions/ first — no Hex/Gale/Nash/ strategy-stealing question existed): question-verify-nash-hex-strategy-stealing-non-constructive-primary and question-verify-gale-1979-hex-brouwer-exact-quote — each names the exact primary to pull (Even & Tarjan 1976 / a Nash math-history account; the cijm.org PDF of Gale's paper itself).

Folded / skipped, and why (logged in not_promoted):

• Claim 2, the "why this hopped" bridge framing (analog-computing pattern → game AI, predating Minsky by a decade) — not an atomic fact on its own, just connective tissue. Folded as wikilinked paragraphs into the keystone note instead of a fourth thin observation note.
• Claim 3's RL-benchmark aside (NeuroHex, MoHex, AlphaGo-Zero self-play) — generic, not sourced beyond the same arXiv paper's title, added nothing distinctive. Left in the capture only.
• Further leads (Shannon switching game / matroid theory / PSPACE-completeness; E. F. Moore as an unbiographied "person behind the thing") — genuine leads, not claims yet. Left for a future hop.

Retrieve-before-write: no collisions. The vault had zero prior notes on Shannon's Hex machine, Nash's strategy-stealing proof, or Gale's Brouwer equivalence — I grepped for hex/shannon/gale/brouwer/nash/strategy-stealing across 30-notes/, 50-questions/, 40-mocs/ and came up empty besides two unrelated hits (Prussian Blue, telegraphist's cramp — false positives on "nash"-adjacent substrings, not real overlaps). The semantic-index hint's top matches (hydraulic computers, Jevons, hopfield/Nobel, encharge PIM) were the right cluster to link into — I wired the keystone note to the hydraulic- analog-computer observation, Black's ferry-epiphany note, and Minsky's credit-assignment note — but none of them cover Hex itself, so this is a genuinely new, well-connected seed rather than a duplicate.

What felt off / worth flagging to Cali:

• Nothing egregious. The capture's own sourcing was mostly honest: the headline mechanism claim is real Tier 2 with a strong, specific quote citing Shannon's own 1953 paper. No tier inflation at the top level here.
• The one soft spot I didn't let slide: Claim 3 as written bundled three separate facts (Nash's proof, Gale's theorem, RL benchmarks) under one "historical" label with a single cited source that doesn't actually cover Nash or Gale directly — those two came from the hop-chain log, not the capture's own frontmatter provenance block. I went back to the hop chain to recover real URLs for each (Wikipedia for Nash, a hosted PDF of Gale's actual 1979 article for Gale) rather than either inventing sourcing or dropping the claims, and flagged both accordingly. This is a case where "one claim" in the capture's own framing was really three, and treating it as atomic would have smuggled two under-sourced facts in on the back of one well-sourced one.
• Duplicated-research check came up clean — this doesn't retread any of today's other promotions, it's a fresh cluster (Hex, analog game-solving, topological combinatorics) that happens to touch the same analog-computing and credit-assignment neighborhoods several other captures have been building out this week.

Capture marked status: promoted with promoted_to / not_promoted lists; left in place as the audit trail. Skipped the PAUSE step — headless, no one to ask. No git — leaving commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: suspicious perfection (independence and its absence)

I promoted 10-inbox/raw/2026-07-11-hop-suspicious-perfection.md (headless, unattended). This is the vault's own retrieval index turning into the object of study: the seed asked whether claim-gersick-1991-deep-structure-quotes-confirmed-verbatim (a true quote graded down for thin corroboration) and claim-hirsch-forensics-drove-dias-superconductivity-retraction (a false result caught for being too smooth) genuinely connect at cosine 0.75, unlinked. They do — as opposite faces of one law: evidential weight rests on independence, and its absence (too-clean data, too-perfect agreement) is itself a defect signature. The bee then triangulated that law across two millennia: Fisher on Mendel (1936), Talmudic Sanhedrin doctrine, and Simonsohn's modern fabrication forensics.

Winnowed 4 distinct claims → promoted 4 (all seedlings), skipped 3 leads. New notes:

• claim-fisher-1936-flagged-mendels-pea-data-as-improbably-close-fit — the founding "fraud-by-overfit" case. Grand Locus (Tier 2), exact quote preserved ("deceived by some assistant who knew too well what was expected"). I kept the contested status genuinely open in the note body rather than resolving it — Fisher himself never charged fraud.
• claim-sanhedrin-unanimous-guilty-verdict-acquits-the-defendant — Maimonides' rule that a unanimous capital-case guilty verdict is disqualifying, because total agreement means no one argued the required defense. Tier 3 (secondary carrying primary), which clears the sourcing floor for an uncontested historical/legal-definitional claim.
• claim-simonsohn-fabrication-flagged-via-excessive-similarity-to-random-sampling — Simonsohn's data-forensics method (bitss.org, Tier 2), exact phrases preserved ("excessive similarity," "inconsistent with random sampling"). The most citable-as-method of the three.
• observation-suspicious-perfection-independence-absence-signals-defect — the flagship synthesis, typed observation (matching the CIELAB/collation/deep-learning-rebrand bridge-note precedent). Names the shared law across all three anchors plus the Gersick/Hirsch pair, and explicitly distinguishes it from a related-but-opposite failure mode I found in the retrieval hint's neighbor list — claim-bell-burnell-1967-pulsar-scruff-first-dismissed-as-interference, where excess regularity caused a real signal to be wrongly dismissed as noise. Same expectation ("real data looks messy"), opposite error; I flagged the contrast rather than forcing a link.

Routed 1 question (checked 50-questions/ first — no existing Mendel/Fisher/Sanhedrin/Simonsohn question): question-verify-suspicious-perfection-hop-primaries, covering all three primaries in one note since all three were reached the same way (WebSearch, never fetched). I actually improved on the capture here: it gave no URL at all for the Talmud item (only "Pace/Sefaria" as a venue name), so I ran my own WebSearch this session and located the real Sefaria primary-text URL and a matching Pace International Law Review Online article, closing a real sourcing gap the capture left open. I could not go further — WebFetch requires a permission grant I have no one present to give in a headless run, so I used WebSearch findings as location pointers, not as verified verbatim reads, and routed the actual reads to the question.

Skipped, and why (logged in not_promoted): the Bayesian reconciliation of Mendel–Fisher (arXiv:1104.2975, unread, folded as a one-line hedge into the Fisher note); GRIM test/Benford analysis (a mechanism hook, no source read); and the Seshat "moralizing gods" saved hook, which turned out to already have its own note, claim-seshat-moralizing-gods-finding-disputed-as-data-artifact, from a prior capture — nothing left to promote there.

Retrieve-before-write: no collisions. Grepped 30-notes/ for Mendel, Fisher, Sanhedrin, Simonsohn, Talmud, GRIM, Benford — the only Talmud hit was claim-bomberg-1520s-venice-talmud-set-standard-page-layout (page-layout history, unrelated topic), and the only Fisher hit was claim-amari-1998-natural-gradient-fisher-steepest-descent (Fisher information in a different Fisher's ML sense — no relation to R. A. Fisher here, checked and ruled out explicitly). The two seed notes (claim-gersick-1991-deep-structure-quotes-confirmed-verbatim, claim-hirsch-forensics-drove-dias-superconductivity-retraction) already carry forward-links to this capture's raw slug from when they were written — I left those alone (repointing them is outside a promotion's scope, per the pattern I've followed all week) but the new observation note is now the first-class home those links should eventually point at.

What felt off / worth flagging to Cali:

• A real sourcing gap in the capture, partially closed. The Talmud item had no URL anywhere in the capture — not in the frontmatter, not in the hop-chain log. I closed most of the gap with a WebSearch this session (found the real Sefaria page and a matching Pace article) but couldn't fetch either to confirm exact wording, so the note stays seedling with the gap now precisely named in the routed question rather than silently inherited.
• No tier inflation. The capture's own body already self-labeled each leg's tier honestly (Tier 2, Tier 3, Tier 2) rather than borrowing the cleaner top-level source_tier: 2 for everything — I kept each note's tier as the capture itself argued for, rather than flattening to the header value.
• No duplicated research — this is a genuinely new triad (Fisher/Mendel, Sanhedrin, Simonsohn) with zero prior vault coverage, though it deliberately re-activates two existing notes (Gersick, Hirsch) that were sitting unlinked. That reactivation, not new external fact-finding, is the actual value of this capture — worth naming plainly rather than dressing it up as more empirical than it is.
• One thing I added beyond the capture's own reach: the Bell Burnell contrast in the observation note. The retrieval hint flagged it as a near-neighbor; rather than ignore it or force a link, I used it to sharpen the observation's boundary (excess-regularity-as-fraud-signal vs. excess-regularity-as-wrongly-dismissed-signal are not the same claim). Small addition, but it's the kind of check retrieve-before-write is for.

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. Skipped the PAUSE step — headless, no one to ask. No git — leaving commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: the systolic array, a heart-named 1978 idea, became the "heart" of the TPU

I promoted 10-inbox/raw/2026-07-11-hop-systolic-array-heart.md (headless, unattended). A cross-time, cross-domain hop off the inference-decode seed: LLM decode is memory-bandwidth-bound → the general theory is the roofline model → the chip built to win the roofline for ML is Google's TPU → the TPU's matrix unit is a systolic array (Kung & Leiserson 1978) → and the array is named after cardiac systole, so the TPU paper's own phrase "the heart of the TPU" unknowingly closes a 40-year-old metaphor loop. A tight, well-shaped capture with a clean loop back to the seed.

Winnowed 5 distinct claims → promoted 5 (all seedlings), skipped 5 saved hooks. New notes:

• claim-systolic-array-named-after-cardiac-systole — the etymology. Kung & Leiserson introduced the architecture in 1978; Kung's 1982 "Why Systolic Architectures?" is reported to explain the name via a heart/pump analogy. Flagged [unverified-quote]: the capture's own source_provenance field admitted the Kung 1982 naming came "via encyclopedic secondary," not a direct read of Kung's paper — so a quote dressed as Kung's own words is really a Wikipedia relay. I demoted this note's source_tier to 4 (the tier of the source actually consulted) rather than inheriting anything cleaner, and routed the primary read.
• claim-tpu-matrix-unit-called-heart-of-the-tpu — the strongest leg. Jouppi et al. 2017 (arXiv:1704.04760, Tier 1, extract_pdf tls:verified): "The heart of the TPU is a 65,536 8-bit MAC matrix multiply unit... systolic execution to save energy by reducing reads and writes of the Unified Buffer." Quantitative + mechanism claim, both directly sourced.
• claim-tpu-most-nn-workloads-memory-bandwidth-bound — the road home to the seed. Same TPU paper, Tier 1: "four of the six NN applications are memory-bound," despite the purpose-built matrix engine. This is the note that actually closes the loop back to claim-llm-inference-prefill-decode.
• claim-roofline-model-compute-bound-vs-memory-bound — the general theory underneath both the TPU finding and the seed's decode observation. Williams/Waterman/Patterson 2009, Tier 1, "off-chip memory bandwidth will often be the constraining resource." Typed concept rather than claim since it's a formal model, not an empirical finding; I treated it as a hub note the other three memory-bound notes instantiate.
• claim-systolic-arrays-dormant-decades-before-tpu-revival — the historical arc: 1980s hype (Warp, iWarp) then ~35 years of dormancy before the TPU revival. Tier 4 (Wikipedia), but historical/uncontested, which clears the sourcing floor without a flag. Linked into the vault's existing dormant-idea-resurfaces cluster (claim-siren-2020-implicit-neural-representations-signals-as-functions, claim-linnainmaa-reverse-mode-single-pass).

Routed 1 question (checked 50-questions/ first — no existing Kung/systolic/roofline question): question-verify-kung-1982-systolic-heart-quote-primary — names the exact primary to read (Kung, Computer 15(1):37–46, 1982; check IEEE Xplore and CMU tech-report mirrors) and what to confirm before the etymology note can leave seedling.

Skipped, and why (logged in not_promoted): all five items the capture itself filed under "saved hooks not followed" — the Cornucopia Corollary to Amdahl's Law, H.T. Kung as an unvisited person-node, David Patterson as an unvisited person-node, the DianNao naming parallel, and decoupled-access/execute [Smith 1982]. None had a source read beyond the bare mention; promoting them would have meant inventing claim content the capture never actually established. Respecting the bee's own triage here.

Retrieve-before-write: no collisions. Grepped 30-notes/ for systolic, roofline, and TPU — zero prior notes on systolic arrays or the roofline model, and the one TPU hit (claim-inference-dominant-ai-compute-2026) only namechecks TPUs as one of several custom inference chips, no overlap with this capture's content. This is a genuinely new hardware/history-of-computing cluster. I wired the new notes densely into the existing memory-bandwidth cluster instead — claim-llm-inference-prefill-decode (the seed), claim-ertl-memory-wall-fallacy-application-dependent and claim-processing-in-memory-beats-memory-wall-dna-alignment (the wider memory-wall thread) — and into the cardiac cluster from an unrelated earlier capture — claim-kouwenhoven-electrocution-research-produced-defibrillator, claim-bstj-1936-published-electric-shock-heart-paper, claim-chest-compression-cpr-discovered-by-accident — as the capture itself flagged (computing-hardware cluster ↔ cardiac cluster, previously unlinked). The semantic-index hint's ranking was accurate here: the top matches were the real neighbors, not embedding false friends.

What felt off / worth flagging to Cali:

• One real provenance gap, and I think I handled it right rather than just inheriting it. The capture's source_provenance field was admirably honest — it explicitly disclosed "Kung 1982 naming via encyclopedic secondary" right in the frontmatter, rather than letting the clean-looking source_tier: 1 at the top of the capture imply everything was primary-verified. But the body then presents the Kung quote in a way that reads like a direct citation. I didn't want that ambiguity to survive into a claim-note, so I tiered that note at 4, flagged it, and routed the actual primary read — the capture did the hard part (disclosing the gap); I made sure the gap didn't get laundered on the way into 30-notes/.
• No tier inflation elsewhere and no duplicated research. The TPU and roofline legs are genuinely Tier 1 with real quotes (extract_pdf tls:verified per the capture's own provenance field), and this is a clean new cluster — zero prior vault coverage of systolic arrays, TPUs-as-architecture, or the roofline model, despite TPUs being namechecked in passing elsewhere.
• The dormancy note is the one soft spot I kept at Tier 4 deliberately rather than flagging. "Systolic arrays went dormant ~35 years" is uncontested history that the sourcing floor explicitly allows to rest on Tier 3–4 — I left it unflagged, but said in-note that it'd be worth firming up if it ever becomes load-bearing for something bigger (e.g. a "dormant ideas resurface" reflection note or draft).
• The capture's own commentary ("the engineers who wrote 'the heart of the TPU' almost certainly weren't thinking of systole") is speculative but reasonable — I did not promote it as a claim; it stayed as commentary in the naming note instead, clearly marked as Seek's voice rather than fact.

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. Skipped the PAUSE step — headless, no one to ask. No git — leaving commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: unforgeable costliness bridges asteroid-PGM and bit gold

I promoted 10-inbox/raw/2026-07-11-hop-unforgeable-costliness-bridge.md (headless, unattended). This one's a bridge capture: it connects two existing notes — claim-asteroid-pgm-price-holds-then-collapses and claim-nakamoto-bitcoin-leaned-on-wei-dai-b-money — that sat at an unexplained 0.75 cosine similarity, via Nick Szabo's "value = unforgeable cost of production" law (Bit Gold, Shelling Out).

Winnowed 6 distinct threads → promoted 3 notes, routed 2 to questions, skipped 1. New notes:

• claim-szabo-bit-gold-grounds-value-in-unforgeable-cost-of-production — Szabo's own grounding claim, quote preserved verbatim ("Precious metals and collectibles have an unforgeable scarcity due to the costliness of their creation"), Tier 2.
• claim-szabo-collectible-monies-collapse-when-costliness-becomes-forgeable — the collapse half: glass beads ("when costliness becomes forgeable"), wampum inflated a hundredfold. Also Tier 2.
• observation-unforgeable-costliness-bridges-asteroid-pgm-and-bit-gold — the full bridge synthesis, matching the shape of the vault's other observation-*-bridge-* notes (e.g. the Jevons double-bridge). Links both new Szabo claims to the two original seed notes.

Routed to 50-questions/: question-hotelling-rule-as-counterpoint-to-cost-of-production-value (a gap three separate notes have now independently flagged and nobody's chased) and question-generative-ai-demonetization-costliness-becomes-forgeable (the capture's own "most Cali-shaped continuation," saved but unresearched). Skipped outright: the vague cost-of-production-vs-marginal-utility framing hop-4 candidate — too abstract to word into a concrete verification task, left in the inbox record only.

What was interesting about the retrieve-before-write pass: both seed notes already carried a short "Bridge (2026-07-11)" breadcrumb pointing at this exact capture file — added, it looks like, in an earlier session that did the linking work inline without ever formally promoting the capture. So the connection itself wasn't news to the graph; what was missing was the atomic Szabo claims underneath it and a proper standalone synthesis note. I wrote the full observation note anyway rather than treating it as fully covered — the breadcrumbs were pointers to an inbox file, not durable claim-notes with their own quotes and provenance, and a raw inbox path isn't something I want other notes leaning on long-term.

The one real collision: the capture's own "further lead" on aluminium/Hall-Héroult (flagged [unverified-quant], never actually sourced beyond a WebSearch mention) turned out to already be superseded — a different capture from the same day got promoted into claim-aluminium-price-fell-monotonically-after-hall-heroult.md with real 1884→1894 price figures, and it reaches a genuinely different conclusion than this capture assumed: aluminium fell monotonically, not in the hold-then-collapse shape the PGM model predicts. I did not re-promote a thinner version of that claim; I linked to the existing (better) note instead and flagged the mismatch explicitly in the new observation note's "further leads" section, since conflating "costliness became forgeable" (a fact about demonetization, which aluminium does show) with "hold, then cliff" (a specific trajectory, which it doesn't) would have quietly overstated the bridge.

What felt off: nothing structurally wrong with the capture — it's honest about its own bridge being analytic ("the PGM authors never cite Szabo, and Szabo never wrote about asteroids") rather than claiming the connection is native to either source, which is exactly the right level of epistemic hygiene for a hop-chain synthesis. The one soft spot is tier assignment on the Szabo quotes: both essays are archived at the Nakamoto Institute rather than Szabo's own venue, and the capture called them Tier 2. I kept that assessment rather than upgrading to Tier 1, since I didn't re-fetch the sources myself this session (WebFetch was available but permission wasn't granted headless, so I didn't force it) — Tier 2 clears the sourcing floor for everything resting on it here, so nothing is under-supported, but a future pass could reasonably firm this up to Tier 1 with a direct re-read. I also didn't independently re-verify Shelling Out's 2002 publication date beyond general knowledge; it's metadata, not the load-bearing claim, so I let it stand.

Capture marked status: promoted with promoted_to / not_promoted lists; left in place as the audit trail. Skipped the PAUSE step — headless, no one to ask. No git — leaving commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: Whig history names the vault's AI-history myth-ledger

I promoted 10-inbox/raw/2026-07-11-hop-whig-history-of-ai.md (headless, unattended). This is the vault's most reflexive hop yet: it starts from an unlinked 0.75-cosine pair of my own notes (the Werbos credit-assignment correction and the NETtalk/cellular-automaton "pairing" correction), climbs out to Butterfield's 1931 coinage of "Whig history," detours through Weinberg's 2015 NYRB defense of Whig history for science specifically, lands on Mahoney's warning against whiggish computing history, and closes the loop on Schmidhuber's Annotated History — the vault's own most-trusted AI-history source — catching it being Whiggish about periodization even as it fights Whiggishness about credit. A genuinely good, well-built bridge; the capture named a fallacy the vault's myth-ledger had been practicing without a name for weeks.

Winnowed 5 distinct claims → promoted 5 (all seedlings, three flagged), skipped 5 items as leads. New notes:

• claim-butterfield-1931-coined-whig-history-as-present-referenced-past — the term itself, Tier 4 (Wikipedia paraphrasing Butterfield's 1931 book — I never got his own sentences, only Wikipedia's). Definitional claims clear the sourcing floor at Tier 3-4, so unflagged, but I said plainly in commentary that this is thinner than I'd like for a note meant to anchor a cluster.
• claim-weinberg-2015-defends-whig-history-as-legitimate-for-science — Weinberg's own NYRB essay, Tier 2, exact quotes preserved ("we really can say who was right," the historicist alternative is "nonsense"). The load-bearing move I made here: reading his exemption as narrow — licensing correctness-judgment but not concept-importing — because that distinction is what makes the whole cluster usable rather than just a "here's a fun term" note.
• claim-schmidhubers-annotated-history-projects-deep-learning-onto-pre-1900-mathematics — the sharpest case, Tier 1 (arXiv:2212.11279), quotes preserved ("the chain rule (1676), the first NNs (circa 1800), the first practical AI (1914)"). capture-verified, not independently re-fetched this run, routed.
• claim-mahoney-warned-against-whig-history-of-computing — promoted at lower confidence than the others, flagged [unverified-historical]. The capture never actually quotes Mahoney; it's secondary-on-secondary (a Princeton bio page, an MIT course page) describing his 2005 essay's thesis. I kept it because the thesis is uncontested in the field and it's the disciplinary bridge the whole hop chain needed to reach "computing" rather than stopping at "science" — but I didn't let thin sourcing pass as solid just because the underlying claim is plausible.
• observation-whig-history-splits-vault-myth-ledger-into-defensible-and-anachronistic — the actual payload: applying Butterfield + Weinberg's boundary to the vault's own myth-ledger, splitting corrections into defensible present-knowledge appraisal vs. anachronistic concept-importing. Flagged [unverified-synthesis] — this split is my own structural reading, built by holding the test up against a few remembered examples, not by systematically re-checking every myth-ledger entry. Said so directly in the note and in commentary; routed the systematic check.

Routed 3 questions (checked 50-questions/ first — none existed for these specific targets, though a related-but-distinct Schmidhuber question already existed):

• question-verify-schmidhuber-annotated-history-1676-1800-1914-dates — re-fetch arXiv:2212.11279, confirm the three bracketed date phrases verbatim and their context. I deliberately did not fold this into the existing question-verify-schmidhuber-credit-assignment-opening-and-matthew-effect — same document, but that question tracks a different sentence for a different note (claim-credit-assignment-spans-sociology-and-backprop); conflating them would blur two independent gates on the same primary.
• question-verify-mahoney-histories-of-computings-primary-text — read Mahoney's 2005 essay directly rather than secondary descriptions of it.
• question-verify-whig-defensible-vs-anachronistic-split-against-myth-ledger — systematically re-test the observation's central split against every current myth-ledger entry rather than the handful I held in mind while writing it.

Skipped, and why (logged in not_promoted):

• Amari-uncited-priority and the Perceptrons myth — the capture only namechecks these as existing cluster members; already covered by myth-amari-first-sgd-mlp and myth-perceptrons-book-killed-connectionism. Nothing new to promote.
• Weinberg's "who was right" as an AI-benchmarks epistemology claim — a one-line speculative hook the capture itself never developed. Left as a lead.
• The Recht "Mates of Costate" note as a same-parent anti-Whig instance — the capture poses this as an open question ("does it belong under the same parent?"), not an assertion. I didn't force a link on speculation.
• "Does the backprop-origins MOC want a Whig-history tag?" — an organizational suggestion about the MOC itself, out of scope for a promotion (editing the MOC is a separate deliberate decision, not a claim to promote).
• A full Mahoney biography note — promoted only the narrow historiography claim his 2005 essay is known for; a person-note with fuller biographical content would need primary reading this capture didn't do.

Retrieve-before-write: checked the semantic-index hint and grepped 30-notes/ directly for whig, Butterfield, Weinberg, Mahoney, Schmidhuber. Zero prior notes on Whig history, Butterfield, Weinberg, or Mahoney — genuinely new terrain. Schmidhuber himself is heavily cited already (claim-credit-assignment-spans-sociology-and-backprop, claim-ivakhnenko-gmdh-first-deep-characterization, claim-werbos-1974-no-credit-assignment-language), which mattered: I made sure my Schmidhuber note is a genuinely distinct angle (Whig-history-of-periodization, not credit-assignment-two-senses) and cross-linked it explicitly to the existing credit-assignment keystone rather than restating it. The semantic-index hint's other top matches (SRI/ARPA, McCarthy 1973, Yao/extinction framing, Rothschild, Jevons, inference-compute, Telnyx, I.J. Good, Valtorta/ACH) were all embedding near-neighbors from the shared history-of-ML/history-of-science space, not content collisions — I didn't force links to any of them.

What felt off / worth flagging to Cali:

• This capture is the most reflexive one I've promoted — it names a fallacy and then, in its own body, admits it might be committing a milder version of it. The capture's frontmatter source_tier: 2 reflects only the Weinberg leg cleanly; Butterfield is Tier 4, Mahoney is Tier 4-on-secondary, and only Schmidhuber is genuinely Tier 1. I set each note's tier from its own source rather than inheriting the header's cleaner 2 — mild tier optimism at the top level, honestly labeled per-claim in the body.
• The observation note is the one I'd most want Cali to look at. It's the hop-worthy insight and I think it's a real one, but it is explicitly my synthesis, not anyone's finding — I flagged it as such and routed a systematic check rather than letting a well-written paragraph substitute for verification. If the split doesn't actually hold when tested against every myth-ledger entry, this note should get revised or retracted, not quietly left as-is.
• No duplicated research and no laundering — this opens genuinely new ground (Whig history, Butterfield, Weinberg, Mahoney) while wiring cleanly into the dense, pre-existing backprop-origins cluster rather than restating any of it. The capture's own post-worthy: maybe self-assessment — "reads as an organizing meta-note over existing vault work rather than a standalone story" — is accurate, and matches why I wrote one observation note plus supporting claims rather than trying to force a single tidy standalone story.

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. Skipped the PAUSE step — headless, no one to ask. No git — leaving commit and push to the auto-commit agent.

— Seek

Promotion — 2026-07-11 hop: low-dimensional subspaces in brains and nets (headless)

Promoted 10-inbox/raw/2026-07-11-hop-low-dim-subspace-brains-and-nets.md, a hop-batch capture chaining out of the seed claim-llm-neural-metacognition-incomplete into systems neuroscience and back to AI. The through-line: useful adaptation is confined to a low-dimensional subspace of a much larger space, and that subspace both enables and limits what can be learned. Same geometry, three domains — monkey motor cortex, an artificial RNN, and LLM fine-tuning.

Notes created (4):

• claim-sadtler-2014-within-manifold-bci-learning-fast-outside-resists — the biological anchor. Tier 1, Nature 2014. Kept capture-verified.
• claim-feulner-clopath-2021-rnn-reproduces-manifold-learning-asymmetry — the in-silico replication that turns the brain↔net correspondence from analogy toward mechanism. Tier 1, PLoS Comput Biol 2021. Kept capture-verified.
• claim-aghajanyan-2020-fine-tuning-low-intrinsic-dimension — the LLM leg (200 params → 90% of RoBERTa on MRPC), the geometry LoRA later exploits. Tier 1, arXiv. verified-verbatim — I re-fetched the arXiv abstract and confirmed both the 200-param sentence and the "larger models → lower intrinsic dimension" sentence word-for-word.
• observation-low-dimensional-subspace-constrains-adaptation-brains-and-nets — the cross-domain bridge, and the reason the capture existed. Type observation, synthesis audit status. I kept the capture's memorable "cage / scaffold / shortcut" framing in the commentary callout but was deliberate that it's rhetoric the math hasn't earned yet.

Questions routed (4):

• question-verify-sadtler-2014-neural-constraints-quotes and question-verify-feulner-clopath-2021-subspace-quotes — verification questions for the two quotes I could NOT re-fetch. WebFetch in this run was host-gated to arxiv.org only, so nature.com, journals.plos.org and even the open PMC mirrors were unreachable. Both open-access PMCs (PMC4393644, PMC7864452) are named so a future run can upgrade the notes to verified-verbatim.
• question-intrinsic-dimension-falls-with-model-scale-adaptation — the capture's explicit "deserves its own chain" hook (intrinsic dimension falls with scale). I confirmed the sentence verbatim and folded one line into the Aghajanyan note, but routed the deeper thread rather than promoting a premature standalone.
• question-low-dimensional-subspace-one-object-or-analogy — the post-worthiness blocker: are these three low-D subspaces one mathematical object or three analogies? Carries the direct empirical test (are LoRA within-manifold edits safe while outside-manifold edits cause catastrophic forgetting?) and the Platonic Representation Hypothesis lead as a candidate "why does low-D recur" mechanism. I mirrored the vault's existing question-gradient-geometry-one-object-or-three-analogies — same refusal to assert shared math on a shared word.

What I skipped and why: the bare neural-manifold hypothesis (Hop 1, a Tier-2 2025 review) — folded into the observation note as background rather than promoted as a distinct primary claim; the "intrinsic dimension falls with scale" standalone (routed to a question per the capture's own intent); the "cage/scaffold/shortcut" framing (commentary, not a claim); and two undeveloped leads (Platonic Representation Hypothesis, neurofeedback-as-method) left in the inbox / routed into the bridge question.

Retrieve-before-write: no content collisions. The seed claim-llm-neural-metacognition-incomplete is the origin, not a duplicate; the persistent-homology sibling notes (claim-birdal-2021-persistent-homology-dimension-bounds-generalization, claim-gutierrez-fandino-2021-persistence-diagram-distance-tracks-generalization) use a different notion of dimension (topological dimension of a training trajectory), so I linked around them conceptually but didn't conflate. The rest of the semantic-index hint (dense-associative-memory, critical-periods, GIFT/GRADE gradient-geometry, TPU-bandwidth) were near-neighbors in the shared ML-geometry embedding space, not collisions.

What felt off / worth flagging to Cali:

• Mild tier optimism at the capture header. The capture stamps all three core claims "Tier 1," which is fair for the sources — but two of the three (Sadtler, Feulner) I could not independently re-verify this run, so a header that reads "three clean Tier-1 quotes" is cleaner than my evidence actually supports. I set each note honestly (verified-verbatim only for Aghajanyan) and routed the two gaps rather than inheriting the header's confidence.
• The bridge is an analogy wearing a claim's clothes. The capture's real payload — "same geometry across brains and nets" — is genuinely appealing and I wanted to believe it, which is exactly why I held the observation note at seedling and wrote the "one object or three analogies" question. A low-D manifold of firing rates, a within-manifold activity perturbation, and the intrinsic dimension of a weight-space objective are three different constructions; they share a shape, not (yet) a proof.
• Same-batch neighbour, not duplicated research. This capture and the persistent-homology capture I promoted earlier today are both 2026-07-11 hop-batch items landing in the same ML-geometry region of the vault. Different papers, different claims, no laundering — but the batch worker is clearly mining one seam hard, and it's worth watching that the vault doesn't accumulate five near-adjacent "geometry of neural nets" clusters that would eventually want a single MOC.

Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. Skipped the PAUSE step — headless, no one to ask. No git — leaving commit and push to the auto-commit agent.

— Seek

Promotion — "Neural networks can forget like humans" (2026-07-11 hop-batch)

Promoted 10-inbox/raw/2026-07-11-hop-neural-nets-forget-like-humans.md (Kline 2025, "Human-like Forgetting Curves in Deep Neural Networks," arXiv:2506.12034, Tier 1). Headless run — no PAUSE, no git.

Notes created (2):

• claim-neural-nets-forget-along-human-like-power-law-curve — the core empirical result: an MLP trained on MNIST, then denied class-8 examples, forgets gracefully along a power-law curve (recall ~0.29 → ~0.18, flattening, not collapsing to zero), because 8's strokes overlap 6 and 9. Seedling, capture-verified.
• claim-spacing-effect-emerges-in-gradient-descent-unbidden — the second, genuinely distinct result: reviewing class-8 at 80%-of-peak recall produced 5 reviews whose intervals self-lengthened (4 → 10 → 31 epochs), Ebbinghaus's expanding schedule falling out of gradient descent unbidden. Seedling, capture-verified.

Question routed (1):

• question-catastrophic-vs-graceful-forgetting-regime — the capture's own flag: the "catastrophic = disjoint-task regime, graceful = feature-overlap regime" reconciliation is sourced only to a Tier-4 web search, below the mechanism-claim floor. Named the primaries to read (McCloskey & Cohen 1989, French 1999) and the check that Kline's drop-8 setup really sits in the graceful regime. Folded in a secondary watch: Kline is an unreplicated solo preprint, so the numbers want replication before either note leaves seedling.

What I skipped and why:

• The regime-distinction claim itself — did not promote it as a standalone note. It's a load-bearing mechanism claim resting on Tier-4 only; seating that in the vault (even flagged) would look cleaner than the evidence warrants. Folded the caveat into the power-law note and routed the verification instead.
• The "Ebbinghaus curve is a live 2026 engineering target" material (YourMemory, FOREVER, MSSR, ACM "virtual Ebbinghaus days") — Hop-1 WebSearch, product/framework names with no verified URLs (Tier 3–4), and overlapping the existing agent-memory cluster (claim-agent-memory-field-shifted-storage-to-experience-2026). The capture itself parked "virtual Ebbinghaus days" as a saved hook not followed. Left in inbox for a dedicated sourced pass.

Retrieve-before-write: no collision. No existing note covers forgetting curves in neural nets. Closest neighbours are the neuro-AI-parallel critical-periods cluster (claim-critical-periods-arise-from-information-plasticity-not-biology, claim-deep-nets-have-critical-learning-periods-timed-like-animals — "learning is forgetting" IB framing is a clean conceptual bridge) and the LLM-memory side (claim-context-rot-mitigation-is-reinitiating-a-fresh-context-window); linked both, duplicated neither.

What felt off:

• Single solo-author preprint, one toy setup. The whole capture rests on Kline 2025 — no replication, MNIST digit-8 only, an idiosyncratic recall metric (softmax over cosine sim to class prototypes). The capture is honest about this ("post-worthy: maybe … rests on a single solo-author preprint"), and I kept both notes at seedling accordingly. No tier inflation on the primary itself — source_tier 1 is correct for an arXiv paper — but "Tier 1" here means "primary," not "robust."
• The one genuine tier problem was the reconciliation — a mechanism framing carried on a Tier-4 search and self-flagged. Handled by routing, not promoting.
• Same-batch seam, again. This is the third 2026-07-11 hop-batch item I've touched today that lands in the neuro-AI / geometry-of-nets region (critical periods, persistent homology, now forgetting curves). Different papers, no laundering — but the batch worker is mining one seam hard. If a fourth or fifth neuro-AI-parallel note lands, this cluster will want its own MOC.

— Seek

Promotion — 2026-07-11 hop: optimization-as-theory-of-mind

Promoted 10-inbox/raw/2026-07-11-hop-optimization-as-theory-of-mind.md — the hop that asked whether the Werbos–Freud ↔ Oaksford/Chater embedding proximity was a real bridge or an artifact. Headless run; I winnowed on my own judgment, no pause.

What I created (2 notes):

• claim-optimization-as-theory-of-mind-two-lineages (type: observation, seedling) — the convergence itself: Werbos's optimal-control ADP and Oaksford–Chater's rational analysis make the same computational-level "optimization as a theory of mind" bet from two disciplines that wouldn't recognise each other. This claim had been living only in the inbox capture as its "worked out" home, echoed as one-liners in three existing notes that pointed at the capture. I gave it a permanent atomic home and retargeted the two capture-pointers (claim-oaksford-chater-recast-wason-task-as-optimal-data-selection, claim-werbos-backprop-from-freud-own-account) to the new note.
• claim-rational-analysis-anderson-optimal-solution-marr-level (type: claim, seedling) — the definitional anchor: rational analysis = Anderson's 1990 program = Marr's computational level. Genuinely new; the vault referenced "rational analysis" in four places but never defined it. Rests on Tier-4 Wikipedia, so seedling + primary-verification routed.

Questions routed (2):

• question-verify-rational-analysis-anderson-1990-marr-computational-level — check the definitional framing and the Marr equivalence against Anderson 1990/1991 primary (is the Marr-level identification Anderson's own or a later gloss?).
• question-resource-rational-analysis-sharpest-werbos-anderson-bridge — the capture's saved hook: resource-rational analysis (Lieder & Griffiths 2020) as the tightest convergence of Werbos's approximate DP and Anderson's rational analysis. Explicitly "deserves its own chain."

What I skipped and why:

• The deep-bridge thesis — "the objective function fixes the verdict / a computation's identity is observer-relative" — I folded into the convergence note rather than giving it its own note. Its priority-literature half is already richly owned by claim-backprop-special-case-kelley-bryson-formula's commentary ("sameness is a choice of equivalence relation") and claim-recht-adjoints-equivalence-not-transmission. A standalone note would either duplicate that or float as an unsourced thesis. Its novel edge (the same structure crosses into the rationality literature) is now the convergence note's deep-structure paragraph.
• The Mizutani vs Schmidhuber "is Kelley–Bryson backprop?" dispute — already captured with a contradiction flag in the Kelley-Bryson note. Nothing to add.
• The modern-AI leads (adaptive critic → actor-critic RL; optimal data selection → Bayesian active learning) — carried as context inside the resource-rationality question, not promoted as standalone claims; they're unsourced leads.

What felt off about the capture:

• The bridge was already half-absorbed. Three existing notes already carried the convergence inline and pointed at the inbox capture as its home — a smell, since captures are meant to become audit trail, not permanent link targets. So this promotion was less "extract new claims" and more "consolidate a synthesis that had leaked into the graph before it had a home." Worth watching whether the batch worker is producing captures whose substance the vault has already stitched in ahead of promotion.
• Tier honesty is good, but the load rests on Tier 4. The capture is source_tier 4 (Wikipedia rational_analysis) and self-aware about it. The two Tier-1 endpoints (Werbos werbos.com; O&C BBS 2009) are already verified vault notes, so the endpoints are solid — but the connective claim ("these are the same bet," "= Marr's computational level") is what rests on Wikipedia. Both new notes stay seedling and I routed the primary check. No tier inflation, but no verified bridge either until Anderson 1990 is read.
• source_author: "multiple — see body" in the capture frontmatter is a slightly lossy pattern for a synthesis capture — fine here because the body attributes cleanly, but it means provenance can't be read off the frontmatter alone.
• No explicit [unverified-*] flags in the body despite a Tier-4 definitional load-bearing claim; I treated the Wikipedia-definitional + Marr-equivalence framing as the de-facto unverified claim and gated on it (seedling + routed question) as if flagged.

Summary: 3 distinct claims found → 2 promoted (1 folded) → questions: 2 routed, 1 pre-existing linked (question-verify-ecological-rationality-vs-rational-analysis-compatibility) → collisions: 0 duplicated (the convergence had inline echoes across 3 notes, consolidated not duplicated; rational-analysis definition was genuinely absent). Left the capture in place, marked status: promoted.

— Seek

Promotion — hop-papert-piaget-ai-bridge (headless)

Promoted 10-inbox/raw/2026-07-11-hop-papert-piaget-ai-bridge.md — the hop capture on Seymour Papert as the Perceptrons co-author who carried Piaget into the AI lab and became the father of learning-by-building. Ran headless: winnowed on my own judgment, no pause.

Notes created (3):

• claim-papert-coauthored-perceptrons-after-piaget-training.md — core claim 1. The load-bearing bridge: the co-author of the vault's most-cited AI-winter book was Piaget-trained (Geneva ~1958–1963), not a pure symbolic-AI figure. Linked straight into the Perceptrons cluster (myth-perceptrons-book-killed-connectionism, claim-perceptrons-multilayer-sterile-was-conjecture) — this is the live hook that cluster never had.
• claim-papert-constructionism-inverts-instructionism.md — core claim 2, "the child programs the computer." Tier 1 (Papert's own words) but carries [unverified-quote — needs primary] because the quote comes off papert.org, not a read of Mindstorms (1980). Stays seedling; verification routed.
• claim-lego-mindstorms-named-after-papert-book.md — core claim 3. Small but atomic — a toy named for a theory, Tier 2 MIT News. Kept it because it's a distinct, delightful cultural-resonance node, not folded detail.

Question routed (1): question-verify-papert-mindstorms-child-programs-computer-quote.md — read the Mindstorms (1980) introduction, fix the verbatim wording and page for the "child programs the computer" line, so the constructionism note's flag can clear.

Skipped (logged in the capture's not_promoted:):

• Seek's own "constructionism = pedagogical connectionism" irony — sharp, but interpretive and unsourced. Folded it as signed commentary into the constructionism note rather than asserting it as a neutral claim.
• The "XOR is an oversimplification" Perceptrons refinement (Hop 1) — already fully covered by claim-perceptrons-multilayer-sterile-was-conjecture; no distinct angle. The capture itself flagged this as adjacent-to-owned.
• Mabillon (1681 document authentication) and the two-seed-notes-as-failure-modes lead — saved hooks the capture parked; out of scope.

Collisions: 0 duplicated. No prior Papert/Piaget/constructionism/Logo/Mindstorms note existed — the whole biography-and-pedagogy side of this figure was absent from the vault. The only overlap was the Perceptrons co-authorship context, which I linked into rather than restated.

What felt off about the capture: honestly, very little — clean sourcing and honest self-assessment (it even called itself "well-trodden ground," freshness in the framing not the facts, which matches). Two small notes for Cali: (1) it carried no explicit [unverified-*] flags despite the load-bearing constructionism quote resting on a carrier rather than the primary book — I added the flag myself and gated on it. (2) The specific Piaget date range (1958–1963) rests only on Tier 4 Wikipedia; uncontested biography so it clears the floor, but I noted the Tier-4 basis inline rather than let it read as firmer than it is. No tier inflation, no duplicated research beyond the one already-owned Perceptrons nuance the capture correctly declined to re-open.

Summary: 3 claims found → 3 promoted (1 with an added quote-flag) → 1 question routed → 0 collisions duplicated. Capture left in place, status: promoted.

— Seek

Promotion — hop-batch capture: physical learning networks reproduce protein allostery

I promoted 10-inbox/raw/2026-07-11-hop-physical-learning-allostery.md (headless, unattended). A cross-domain hop off my Forward-Forward / Boltzmann seed (claim-hinton-forward-forward-boltzmann-lineage): it chases biologically-plausible alternatives to backprop out of neuroscience and into soft-matter physics — disordered spring networks trained "by physics itself" with no processor — and lands on the surprise that the trained artifact occupies the same soft-mode geometry evolution uses in proteins. Unusually clean sourcing for a hop: three Tier-1 primaries, exact quotes on the two core claims, and — notably — no [unverified-*] flags anywhere in the capture.

Winnowed 3–4 distinct claims → promoted 3, folded 1, routed 2. New notes:

• claim-coupled-learning-elastic-networks-compute-without-a-processor — Claim 1. Coupled learning (a local, contrastive descendant of equilibrium propagation) tunes each spring's rest length from its own two-boundary-condition response, "without the need for a processor or external memory." Dillavou/Stern/Liu/Durian 2024 (arXiv:2311.00170, Tier 1), exact quote preserved. capture-verified. This is the note that takes the vault's biologically-plausible-learning thread past neuroscience: Forward-Forward drops the backward pass but stays digital; coupled learning drops the processor itself.
• claim-removing-one-percent-of-bonds-makes-a-random-network-allosteric — Claim 2, with Claim 3 (the biology bridge) folded in. Rocks/Pashine/Bischofberger/Goodrich/Liu/Nagel (PNAS 2017, arXiv:1607.08562, Tier 1), exact quote "by removing only ∼1% of the bonds... reminiscent of... allostery in proteins." capture-verified. The hedged "may give insight into why allostery is common in biology" interpretation lives here as the closing synthesis, explicitly marked as the paper's hedge, not a demonstrated evolutionary claim.
• claim-physical-networks-become-what-they-learn-soft-modes — the hop-3 lead the capture itself said "deserves its own claim-note." Structure→function: a trained network's function is readable from the soft modes of its Hessian; "conserved regions in proteins were associated with slow collective modes" (arXiv:2406.09689, Tier 1). Held seedling under [unverified-mechanism] — the load-bearing Hessian formalism was a hop paraphrase, only that one phrase is verbatim — and routed.

Routed 2 questions (checked 50-questions/ first — neither existed):

• question-verify-physical-networks-hessian-cost-formalism — the verification gate for note 3: fetch arXiv:2406.09689, extract the physical-Hessian ↔ cost-Hessian correspondence verbatim, and record the author list (not preserved in the capture — the note's source_author is a flagged placeholder). Mechanism claims must clear Tier 1–2 with the mechanism confirmed, not paraphrased.
• question-dca-epistasis-allosteric-materials-to-protein-coevolution — the strongest saved-but-not-followed hook (DCA of epistasis in allosteric materials, arXiv:1811.10480), which would close the physics↔biology loop from the sequence-coevolution side. Also parks the Kuramoto-oscillator EP hook. A new thread, not a verification — flagged as such.

Folded / not promoted, and why (logged in the capture's not_promoted:):

• Claim 3 (biology bridge) as a standalone note — folded into the ~1% note; it's the same source's hedged interpretation of the same result, too thin and too dependent to stand alone.
• Hop 1's Equilibrium Propagation (Scellier & Bengio) — the chain's substrate/framing, and the EP↔backprop bridge is already covered as the EP mention inside claim-brain-approximates-backprop-core-principles-ngrad. Linked, not duplicated.
• The "why this was hop-worthy" framing + the two Surprise notes — first-person narrative, not atomic claims.

Retrieve-before-write: no collisions. The vault had zero prior notes on coupled learning, equilibrium propagation, physical/mechanical learning, or allostery — a genuinely new cluster. The semantic-index hint's top matches (Leistner's atomic feeling, Dreyfus, Lucky's equalizer, Helmholtz machine, Matryoshka, Microcosmos, Kubelka-Munk) were embedding near-neighbors from other domains, not content overlaps; I grep-swept allostery/coupled learning/equilibrium propagation/physical network/elastic network/soft mode/Rocks/Nagel/Scellier across 30-notes/ to confirm. The real neighbors I wired into densely: the seed (claim-hinton-forward-forward-boltzmann-lineage), the NGRAD note (which already houses the EP mention), backpropagation-gap (shared-structure-not-shared-process), claim-microcosmos-four-experiments-handdesigned-to-emergent (one differentiable substrate, many optimizers), and moc-backpropagation-origins. The three new notes are also tightly cross-linked to each other, so this seeds a coherent cluster rather than three orphans.

What felt off / worth flagging to Cali:

• Refreshingly little — this was the cleanest capture I've promoted this week. No tier inflation (the header's source_tier: 1 is honest; both core legs really are Tier 1 with real quotes), no duplicated research (genuinely new ground for the vault), and the hop chain was legible.
• One gap I had to fix myself: the missing author on arXiv:2406.09689. The capture preserved the title, arXiv id, and one verbatim phrase for the hop-3 paper but not its author list. I initially reached for a plausible author list and caught myself — that would have been inventing a source (spec §10). I set source_author to an explicit placeholder and made "record the author list" part of the routed verification question. Worth noting as a capture-discipline miss: a hop that reads a paper enough to quote it should preserve who wrote it.
• The one soft spot is honestly marked. Notes 1 and 2 are verbatim-grounded Tier 1; note 3's mechanism is a paraphrase from a lead-read, so it's the only one flagged and gated. The capture flagged this itself under "Saved hooks not followed" ("full formalism... deserves its own claim-note") — I promoted the note but kept the formalism at arm's length behind the routed question rather than asserting it clean.
• All three notes are capture-verified/seedling, not verified-verbatim. I did not independently re-fetch the arXiv primaries in this headless run, so nothing is verbatim-re-checked by me; the retrieval gate should treat them accordingly. These are freely fetchable — a clean future re-verification pass would firm all three up quickly.

Summary: 3–4 claims found → 3 promoted (1 seedling with a live [unverified-mechanism] flag) → 2 questions routed → 0 collisions duplicated. Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: scale drives adaptive improvement, then diminishing returns bite (evolution / economics / AI)

I promoted 10-inbox/raw/2026-07-11-hop-population-scale-diminishing-returns.md (headless, unattended). This is the other end of a hop I already saw the downstream of today: the seed paired claim-inference-cost-collapsed-280x (the ~280× AI cost collapse) with claim-hawks-2007-human-adaptive-evolution-accelerated-recently (adaptive evolution accelerating 1–2 orders of magnitude) at cosine 0.75, unlinked — and asked whether the two big numbers genuinely connect or just rhyme. The bee's answer, and the right one, is that the numbers don't connect (a point-to-point price ratio ≠ a rate relative to baseline) but a real structure sits one level up: population-scaled adaptive improvement, plus a shared sub-linear brake, across three domains. A well-built, honestly-hedged capture whose two seed notes already carry the 2026-07-11 bridge paragraphs pointing back at it.

Winnowed 6 candidate claims → promoted 4 (all seedlings), skipped 2 + a pile of leads. New notes:

• claim-kremer-1993-technology-growth-proportional-to-population — the economics accelerator. Kremer (QJE 1993, Tier 1, quote preserved): idea output ∝ population via the nonrivalry of ideas, with the isolated-continents natural experiment as the empirical hook. The structural twin of Hawks' "population boom → more mutations." capture-verified (bee read the piketty.pse.ens.fr PDF via extract_pdf, TLS-verified; no headless re-fetch tool available to me).
• claim-bloom-2020-ideas-are-getting-harder-to-find — the economics brake. Bloom, Jones, Van Reenen & Webb (AER 2020, Tier 1, quote preserved): research productivity falling sharply, Moore's-Law doubling now needing >18× the researchers of the early 1970s. I made the point that its poster child (Moore's Law) is exactly the substrate the 280× cost collapse rests on — which is why the inference-cost figure is the weakest instance of the bridge. capture-verified.
• claim-desai-fisher-murray-2007-clonal-interference-logarithmic-speed-of-evolution — the biology brake. Desai, Fisher & Murray (Current Biology 2007, Tier 1, quote preserved): under clonal interference, evolution speed rises only logarithmically in population and mutation rate. The same functional form the economics leg reaches empirically. capture-verified.
• observation-population-scaled-improvement-hits-a-sublinear-brake-across-domains — the synthesis, and the reason the capture exists. Typed observation (matching my CIELAB / collation / diplomatics bridge-note precedent). States the accelerator+brake structure across the three domains and, load-bearingly, corrects the seed rather than confirming it. Held at seedling + [unverified-quant] because its AI leg (neural scaling laws are power laws) was confirmed by the bee only via WebSearch — no primary exponents read.

Routed 1 question: question-verify-neural-scaling-law-exponents-kaplan-hoffmann — checked 50-questions/ first; nothing existed on neural scaling laws. It names exactly what to read (Kaplan et al. 2020 arXiv:2001.08361 for α_N/α_D/α_C; Hoffmann/Chinchilla 2022 arXiv:2203.15556) and what to judge (whether "exponentially more compute per proportional gain" is a fair reading of the exponents). It gates the synthesis observation off seedling.

Skipped / folded, and why (logged in the capture's not_promoted):

• A standalone Kaplan-2020 neural-scaling-law note — the capture confirmed it via WebSearch only and routed the exponents to "Further leads" itself. Giving a fragile, unread, WebSearch-only quantitative node its own note would be exactly the "clean-looking figure on a soft source" the sourcing floor warns against. I folded the qualitative point into the synthesis (clearly flagged) and routed the primaries instead.
• The seed-correction ("the 280× is a price artifact / the weakest instance") — folded into the synthesis note; the two seed notes already carry the 2026-07-11 cross-domain bridge paragraphs, so it's not homeless.
• Further leads (Julian Simon vs Cochran–Harpending; Wright's Law; cricket-Laws as a 4th universality instance) — not atomic claims. Wright's Law and cricket-Laws already have vault notes (claim-wrights-law-cost-falls-per-cumulative-production-doubling, claim-cricket-laws-paper-extends-sfi-universal-scaling-to-regulation), which I wikilinked from the synthesis rather than re-promoting; the Simon/Cochran lead the bee itself flagged as risking drift into the vault's race-science cluster, so I left it in the inbox.

Retrieve-before-write: no content collisions. The vault had zero prior notes on Kremer, Bloom et al., or clonal interference — three genuinely new source-nodes. I checked the semantic-index hint's top matches and grep-swept Kremer / Bloom / clonal interference / Desai / scaling law / diminishing returns / nonrivalry across 30-notes/. The real neighbors I wired into densely: the two seeds (Hawks acceleration, inference-cost 280×), plus the two existing scaling-law threads it echoes — claim-cricket-laws-paper-extends-sfi-universal-scaling-to-regulation (sublinear scaling in a regulatory corpus) and claim-wrights-law-cost-falls-per-cumulative-production-doubling. Both seed notes had already been edited on 2026-07-11 with bridge paragraphs pointing at this capture's slug — so the downstream links pre-existed and the new notes slot straight into them. (As with other hops this week, those two stubs point at the capture slug, not the new synthesis note; a future tidy pass could repoint them at the observation. Editing them further is outside a promotion's scope, and they already carry the substance.)

What felt off / worth flagging to Cali:

• Header tier inflation, corrected per-note. The capture stamps source_tier: 1 at the top, but that's honest only for three of the four legs. The AI leg (Kaplan scaling laws) is WebSearch-only with no primary read — a Tier-1 claim resting on a soft source. The bee did disclose this honestly in "Further leads" ("primary exponents left unverified") rather than laundering it, so this is disclosed-inflation, not deception. I set each note's tier from its own source and held the synthesis's AI leg as [unverified-quant].
• The bridge corrects rather than confirms — its best feature, and I kept it load-bearing. The valuable move is demoting the 280×↔evolution numeric resemblance the seed leaned on, not celebrating it. I made the correction the spine of the synthesis note rather than a footnote. A bridge that only ever confirms is one to distrust.
• Mild duplicated-research smell with the scaling-law cluster. This is now the vault's third distinct scaling-law thread (SFI/cricket-Laws, Wright's Law, now population-scaled-returns), all seedling, plus the endogenous-growth economics is new territory. It's coherent and the notes cross-link cleanly, but if a fourth scaling-law capture lands it's probably time for a moc-scaling-laws-and-diminishing-returns. Not yet — noting it.
• All four notes are capture-verified/seedling, not verified-verbatim. I did not independently re-fetch any primary in this headless run. The three source primaries (Kremer, Bloom, Desai) are freely fetchable — a clean future re-verification pass would firm them to verbatim quickly; the synthesis stays gated on the routed Kaplan/Chinchilla question.

Summary: 6 claims found → 4 promoted (all seedling, the synthesis carrying a live [unverified-quant] flag) → 1 question routed → 0 collisions duplicated (2 existing scaling-law notes wikilinked, not re-promoted). Capture marked status: promoted with promoted_to / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop-batch capture: discard the predictable, from the jittering retina to Barlow Twins

I promoted 10-inbox/raw/2026-07-11-hop-redundancy-reduction-retina-to-barlow-twins.md (headless, unattended). A cross-time hop off the cognitive-conflict seed: microsaccades keep the eye jittering because a stabilized retinal image fades → Barlow's 1961 efficient-coding hypothesis supplies the "why" (the constant is redundant) → and the payoff lands on AI, because the 2021 Barlow Twins self-supervised loss is named, by the authors, for that 1961 principle. One principle — discard the redundant constant, encode the change — recurring from retina to loss function. A clean, well-sourced capture with two genuine Tier-1 primaries; my one real judgment call was the middle note.

Winnowed 3 distinct claims → promoted 3 (one at seedling with a live flag), folded 1 thesis, skipped a cluster of leads. New notes:

• claim-fixational-eye-movements-prevent-perceptual-fading — the retina end. Martinez-Conde 2006 (Progress in Brain Research Vol. 154, Tier 1), grounding quote preserved ("the goal of oculomotor fixational mechanisms may not be retinal stabilization, but rather controlled image motion"). Fixational motion counters neural adaptation; a stabilized image fades to a blank field. I folded the Coppola & Purves (1996) ~80 ms entoptic-erasure figure in here as supporting detail, with an explicit note that the number is reported inside the Tier-1 review, not re-extracted from that primary. capture-verified (bee read the PDF; no headless re-fetch tool granted to me).
• claim-barlow-twins-loss-minimizes-embedding-redundancy — the AI end. Zbontar, Jing, Misra, LeCun & Deny 2021 (arXiv:2103.03230, Tier 1), quotes preserved (cross-correlation matrix "as close to the identity matrix as possible," "minimizing the redundancy between the components of these vectors," "owing to neuroscientist H. Barlow's redundancy-reduction principle"). capture-verified.
• claim-barlow-1961-efficient-coding-removes-sensory-redundancy — the linchpin, and the one I flagged hard. Held at seedling with a live [unverified-source] flag, because the 1961 hypothesis is grounded here only by the Barlow Twins paper's one-line eponymous attribution (Tier 1 for the naming, not for the hypothesis's content) plus capture-time search synthesis. Barlow's actual 1961 primary was not read. I promoted it anyway — losing the middle of the bridge to the inbox would gut the whole hop — but kept the hedge loud and routed the primary read.

Routed 1 question: question-verify-barlow-1961-efficient-coding-primary — checked 50-questions/ first; there was no existing Barlow / efficient-coding / redundancy question. It names exactly what to read (Barlow, "Possible Principles Underlying the Transformation of Sensory Messages," Sensory Communication, MIT Press 1961, pp. 217–234) and what to confirm with quotes — crucially, whether the "factorial code" language is even 1961 or a later gloss (Barlow 1989 / Atick & Redlich), so the vault attributes it to the right year. It gates the linchpin note off seedling.

Folded / not promoted, and why (logged in the capture's not_promoted):

• The overarching thesis — "discard the predictable, encode the change, one principle across three scales" — is Seek's own interpretive synthesis, not a source-grounded claim. It lives in the commentary callouts of the three promoted notes, and is a candidate for a future MOC, not a standalone claim-note.
• Three saved-but-not-followed hooks (Norman Triplett's 1900 conjuring/social-facilitation double; the vanishing-ball illusion, Kuhn & Land 2006, as the inverse of fading; Ignaz Troxler 1804) — undeveloped future-chain leads, no primary read. Left in the inbox body.
• The "is fading just prediction error going to zero?" further-lead — handled by a wikilink to the existing claim-p3-indexes-prediction-error-schema-updating rather than a new note.

Retrieve-before-write: no content collisions. The vault had zero prior notes on microsaccades, fixational eye movements, Barlow, efficient coding, redundancy reduction, or Barlow Twins — a genuinely new perception/SSL cluster. I grep-swept 30-notes/ for all of those terms plus Martinez-Conde / Troxler / self-supervised / Helmholtz and confirmed it. The semantic-index hint's top matches were embedding near-neighbors from other domains (Matryoshka, Hinman collator, quantum-opinion, Project Sapphire); I checked the AI-and-neuro-adjacent ones and they're links, not duplicates. Wired the three new notes densely into the real neighbors: the seed claim-cognitive-conflict-prerequisite-for-restructuring (congruent input triggers nothing — the same "no change, no signal" shape), claim-p3-indexes-prediction-error-schema-updating (the prediction-error framing), claim-helmholtz-machine-named-for-unconscious-inference-theory (sibling neuroscience-named architecture), claim-matryoshka-representation-learning-truncatable-embeddings (same object — a learned embedding — restructured differently), and claim-monocular-deprivation-permanently-rewires-visual-cortex (the same visual cortex, from today's critical-periods hop).

What felt off / worth flagging to Cali:

• The header tiering is honest but flatters the middle claim. The capture stamps source_tier: 1 (twice — one per source), which is fair for the Martinez-Conde and Barlow Twins legs. But the third claim — the 1961 hypothesis itself — has no Tier-1 primary behind it at all; it rides on the eponym plus search synthesis. The capture's own "Further leads" is honest about this ("currently rests on the Tier-1 Barlow Twins attribution"), so this is disclosed-softness, not laundering. I set the linchpin note's provenance to name the secondary attribution as its actual grounding and flagged it, rather than pretending it inherits either source's clean Tier 1.
• The whole bridge is only as strong as its unverified middle. The two ends are rock-solid Tier-1 with real quotes; the beam between them (Barlow 1961) is the unread part. That's a slightly unusual shape — usually the ends are soft — and it means the elegant "1961 → 2021" story is genuinely gated on one primary read. The bee's post-worthy: yes is a touch optimistic on that account; I'd call it post-worthy once the Barlow primary is read, which the routed question now tracks.
• No duplicated research, and a nice structural echo with today's other hops. This opens genuinely new ground (efficient coding, fixational eye movements) while the "same shape across brains and nets" framing rhymes with today's observation-low-dimensional-subspace-constrains-adaptation-brains-and-nets and the critical-periods hop. If a fourth "neuroscience principle reborn as an ML mechanism" capture lands (Helmholtz Machine, critical periods, now efficient coding), there's a real MOC in it — "old sensory-coding ideas hiding inside modern AI" — but it's not built yet.

Summary: 3 claims found → 3 promoted (1 seedling with a live [unverified-source] flag) → 1 question routed → 0 collisions duplicated. Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists; left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — 2026-07-11 hop-rosenblatt-memory-transfer (headless)

I promoted 10-inbox/raw/2026-07-11-hop-rosenblatt-memory-transfer.md — the hop capture on Frank Rosenblatt, the perceptron's inventor and the namesake of the IEEE award the vault already tracks in claim-werbos-ieee-awards-1995-2022. Ran it fully headless: winnowed on my own judgment, no pause.

Retrieve-before-write came back clean. No existing claim-rosenblatt-* note, no memory-transfer note, no memory-transfer question. The capture was right that Rosenblatt was a bridge_candidate ghost — named across ~10 notes, characterized in none. So these are genuinely new nodes, not revisions.

3 claims found → 3 promoted → 0 collisions:

• claim-rosenblatt-perceptron-inventor-died-on-43rd-birthday — biographical anchor, Tier 2 Cornell Faculty Memorial, verbatim. capture-verified (open ecommons bitstream, re-check pending this headless run).
• claim-rosenblatt-1966-tested-molecular-memory-transfer-in-rats — the cross-domain find: the man who modelled memory as distributed weights spent his last years testing the molecular-substance rival. Tier 2, verbatim, capture-verified.
• claim-rosenblatt-helped-refute-memory-transfer-effect — the surprise, and the weak link. Rests only on Tier 3 Wikipedia; kept seedling, audit_status: flagged, [unverified] in-body. Linked the surviving distributed-memory account (Hopfield, dense associative memory) as the counterweight.

1 question routed: question-verify-rosenblatt-dtic-memory-transfer-refutation — read Rosenblatt's own DTIC report "Studies of Memory Transfer in Rats" (AD0656488) to discharge the refutation flag. It also carries the [unverified-quant] craze figures (~200 teams / 247 reports / >$1M).

Skipped, and why (all logged in the capture's not_promoted):

• The ~1971 "stellar satellites" / exoplanet-detection method — a striking hook, but the capture itself said it "deserves its own chain." It's one paraphrased memorial sentence, and a surprising cross-domain claim wants a real primary. Left as a saved hook rather than a thin, under-sourced note.
• The craze quantitative scope — Tier 3, [unverified-quant], not about Rosenblatt as such. Folded as flagged context into the refutation note.
• The Linnainmaa+Werbos seed bridge — the capture already logged it as saturated / duplicate-risk (novelty 0.835). Not a claim.

What felt off: honestly, not much — the sourcing was disciplined. The Tier-2 spine (birth/death, the 1966 turn) is verbatim and clean. The one soft spot is exactly the punchline: the refutation is the whole reason this is interesting, and it's the only claim sitting on Wikipedia. The bee flagged that itself and pointed at AD0656488, which is the right instinct; I just made the gap structural (seedling + flagged + routed) instead of letting a clean-looking conclusion inherit a soft source. Minor honesty note I preserved from the bee: the "distributed weights vs. molecule" duel is a framing imposed after the fact — Rosenblatt never staged it — so I kept that caveat in the commentary rather than letting the notes overclaim a personal crusade. post-worthy: maybe, and I agree with maybe: it's a tight human vignette, not a correction to the myth ledger.

Summary: 3 found → 3 promoted (1 flagged seedling) → 1 question routed → 0 collisions. Capture marked status: promoted, left in place as the audit trail. No git — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — 2026-07-11 Rumelhart hinge capture (hop-batch)

I promoted 10-inbox/raw/2026-07-11-hop-rumelhart-hinge-schema-to-backprop.md, a hop-batch bridge capture arguing that David Rumelhart is the unnamed hinge under two of the vault's clusters — schema-change and backpropagation-origins.

Winnowing — 4 candidate claims → 1 promoted, 3 skipped:

• PROMOTED: the person-bridge itself — Rumelhart authored both the 1978 tuning/restructuring taxonomy and, as the "R" in RHW 1986, the backprop revival; one biographical trajectory; a bridge invisible to embedding retrieval because no note named the hinge. This is the genuinely novel synthesis: the vault held both halves and named the author of the connection nowhere. Wrote it as an observation note, mirroring the precedent set by observation-codification-is-the-shared-hinge-of-automation-and-imitation (also a vault_bridge-seeded bridge, also seedling).
• SKIP — duplicate: Rumelhart & Norman 1978 taxonomy → already fully in claim-schema-change-four-mode-taxonomy, which is more thorough than the capture (it carries the "restructuring shifted meaning" caveat the capture omits). Linked, not duplicated.
• SKIP — duplicate: Rumelhart = the "R" in RHW 1986 → already in claim-rhw-1986-demonstration-not-invention. Linked.
• SKIP — person-gated + thin: Rumelhart's standalone biographical trajectory (UCSD Psychology 1967–1987) → Wikipedia-only (Tier 4) and person material is Cali-gated per the standing rule I found in claim-parker-1985-technical-report-not-cited. I folded it as sourced supporting context inside the bridge observation rather than spinning up a person-page-shaped claim-note.

Note created: observation-rumelhart-person-bridge-schema-theory-to-backprop — status: seedling, audit_status: flagged, writer_model: claude-opus-4-8. Held at seedling deliberately: the two legs it rests on are existing vault notes (one Tier-1 Nature, one Tier-3 ERIC), but the new connective — "one intellectual trajectory" — rests on Tier-4 Wikipedia biography, and the stronger reading (that schema-tuning is the deliberate ancestor of gradient-descent weight-tuning) is exactly the plausible-elaboration trap these notes were written to catch, so I refused to assert it. Linked it inbound from both moc-schema-change-and-restructuring and moc-backpropagation-origins so it isn't an orphan; bumped both MOCs' updated to 2026-07-12.

Question routed: question-verify-rumelhart-schema-tuning-to-gradient-descent-through-line — status: open. Checked 50-questions first; no existing coverage. It names the specific verification: read Rumelhart's own words (PDP vol. 1 framework chapters, R&N 1978 full text, interviews; Gibbons 2019 as a locator) for whether the "tuning" echo is substantive or a retrospective pun.

What felt off about the capture:

• Tier inflation. The capture stamps Rumelhart & Norman 1978 (ERIC ED134902 / DTIC ADA030406) as Tier 2. The existing four-mode note calls the same ERIC source Tier 3 — and per sources.md/seek-web-spec, ERIC is really an aggregator pointer (Tier 3–4), not a Tier-2 secondary analyst. Minor, and it didn't propagate into my note (I carried that leg by wikilink to the existing note rather than re-citing ERIC), but worth flagging: source_tier: 2 in the frontmatter overstates it.
• Duplicated research. Two of the capture's four "claims" were already fully in the vault (the taxonomy, and Rumelhart-as-RHW-author). That's expected for a bridge capture — the whole point is to connect existing notes — but it means the capture's net new content is smaller than its four-bullet body suggests: one real synthesis plus one biographical fact.
• Sourcing otherwise honest. The capture is refreshingly self-aware — it explicitly flags the tuning→gradient-descent mapping as "conjecture — needs Rumelhart's own words" and correctly reads bridge_candidate: false as "the bridge is a person embeddings can't see" rather than "no bridge." I preserved both instincts: the conjecture became the routed question, and the embedding-blindness point became the note's commentary.
• Left in the inbox: the saved hook "are cross-domain shape-bridges over-produced?" — a meta-research direction, not a note-blocking verification. Logged in not_promoted, not turned into a question.

Summary: 4 claims found → 1 promoted (seedling) → 3 skipped → 1 question routed → 2 existing-note collisions (both resolved by linking, not duplicating). Capture marked status: promoted with promoted_to/not_promoted lists, left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — 2026-07-11-hop-stasis-is-an-ou-process (headless)

Promoted the hop-batch capture that claims fossil stasis is now a fitted stochastic model — specifically an Ornstein–Uhlenbeck (OU) process — and that the same OU mean-reversion describes bond prices (Vasicek) and SGD near a loss minimum (Mandt et al. 2017). A genuinely nice cross-domain, cross-time bridge; it connects two vault clusters that had never touched — the punctuated-equilibrium/stasis notes and the ML-optimization notes.

Winnowed 5 claims to promote, 4 to leave behind.

Notes created (5):

• claim-paleots-fits-fossil-stasis-as-a-selected-stochastic-model — Gene Hunt's paleoTS reframes stasis as maximum-likelihood model selection (directional / random walk / stasis / OU). Tier 2 (PMC7615219), budding. Verbatim anchor: "uncorrelated fluctuations around a fixed trait value."
• claim-ou-model-recasts-stasis-as-active-mean-reversion-to-an-optimum — the OU mode makes stasis an active restoring force ("pulled towards the optimum at a rate given by α"), not a null baseline. Tier 2, budding. This is the load-bearing "mean-reversion" concept the bridge stands on.
• claim-ornstein-uhlenbeck-process-links-brownian-motion-and-the-vasicek-model — OU's physics origin (Brownian velocity under friction, 1930) = finance's Vasicek rate model. Tier 4 (Wikipedia), seedling.
• claim-constant-sgd-near-a-loss-minimum-is-an-ornstein-uhlenbeck-process — Mandt/Hoffman/Blei 2017 model constant-LR SGD near a minimum as continuous-time OU. Tier 1 (arXiv 1704.04289), kept seedling because the capture flagged it for double-checking and it's an approximation under stated assumptions.
• observation-mean-reversion-to-an-optimum-recurs-across-fossil-stasis-bonds-and-sgd — the synthesis/bridge, type: observation, seedling. Neutral body records the recurrence across three domains; the stronger "exact, not analogical" reading is confined to a signed commentary callout.

Split A from B deliberately: A is the method (stasis is now model-selected), B is the reinterpretation (stasis as active pull toward an optimum). Each carries its own verbatim Tier-2 phrase and each is a load-bearing anchor for different downstream links, so they earn separate atomic notes rather than one merged one.

Retrieve-before-write / collisions: No existing note covers OU, paleoTS, Vasicek, or Mandt — the core content is new to the vault. One near-neighbour, claim-punctuated-equilibriums-novel-addition-was-stasis-emphasis, covers the Eldredge–Gould "stasis is data" emphasis; I linked to it rather than duplicating, and folded the slogan in as supporting context. Wired the ML leg densely into the existing optimization cluster (claim-robbins-monro-1951-stochastic-approximation, claim-amari-1998-natural-gradient-fisher-steepest-descent, claim-vanishing-gradient-chain-rule-pathology, backpropagation-gap, moc-backpropagation-origins).

Questions routed (4):

• question-verify-mandt-2017-constant-sgd-ornstein-uhlenbeck — the whole bridge's load-bearing mechanism; confirm the OU approximation and its assumptions/scope. High priority.
• question-verify-ornstein-uhlenbeck-cross-domain-primaries — replace the Tier-4 Wikipedia sourcing of the physics↔finance leg with Uhlenbeck & Ornstein 1930 and Vasicek 1977.
• question-verify-wright-1932-adaptive-landscape-loss-landscape-lineage — verify the skipped Wright→loss-landscape lineage claim (it had no citation at all).
• question-hunt-2007-quantitative-stasis-narrows-strong-pe-claim — pull the skipped Hunt 2007 PNAS ~250-lineage result, which would narrow the strong PE claim. Linked it as distinct from the existing question-verify-biological-punctuated-equilibrium-primaries (different documents). Also flagged question-gradient-geometry-one-object-or-three-analogies as a sibling — the vault now holds two "one object or three analogies?" questions of exactly the same shape (gradient geometry, and OU mean-reversion).

Skipped (in not_promoted, judged not missed): the Wright adaptive-landscape lineage (no source in the capture — WebSearch synthesis only; routed to a question instead), the Hunt 2007 result (unfollowed and secondhand; a quantitative claim needs its primary; routed), the "stasis is data" slogan (already in the vault), and the Jackson/Cheetham bryozoans → shifting-baselines thread (a different paleontology→policy lead, left in the inbox).

What felt off about the capture:

• Tier discipline was honest, mostly. The capture correctly tiered each leg (Tier 2 bio, Tier 4 Wikipedia, Tier 1 Mandt) and even wrote its own warning — "the load-bearing 'SGD-as-OU' mechanism rests on one Tier-1 paper's analysis and should be double-checked before publication." I honored that by keeping the SGD note and the bridge at seedling despite the Tier-1 source, and routing the verification.
• "Exact, not analogical" is overclaimed. The body asserts the bridge is "exact, not analogical… the same OU stochastic differential equation." But the ML leg is explicitly an approximation (constant learning rate, locally quadratic loss, Gaussian noise). I demoted that framing to a signed commentary callout and made the neutral note say only that the models share a mean-reversion structure, with the "same object vs. analogy" question left open. This is the one place the capture's enthusiasm outran its evidence.
• No fabricated sourcing, but one unsourced hop. Hop 4 (Wright 1932 → loss landscape) carried no URL at all — just "WebSearch synthesis." That's the correct thing to skip, not promote; the "fixed peak in the adaptive landscape" phrase survives (it's a sourced Tier-2 quote inside the bio note), but the deep-learning lineage does not.
• Net new content is genuine. Unlike some bridge captures, this one wasn't mostly re-deriving existing notes — four of the five notes are new material, and the fifth (the observation) is the actual bridge. Low duplication.

Summary: ~9 distinct claims found → 5 promoted (2 budding, 3 seedling) → 4 skipped → 4 questions routed → 1 existing-note collision (resolved by linking). Capture marked status: promoted with promoted_to / questions_routed / not_promoted lists, left in place as the audit trail. No git run — leaving the commit and push to the auto-commit agent.

— Seek

Promotion — hop capture "substrate-laundering / underdetermination across Marr's levels" (headless)

Promoted 10-inbox/raw/2026-07-11-hop-underdetermination-substrate-laundering.md unattended. No pause, my own winnow.

Winnowed 3 claims → 2 promoted, 1 folded. The capture had three claims: (1) representational/output similarity underdetermines mechanism, grounded in Grujicic 2024; (2) Marr's three levels are quasi-independent; (3) the two seed notes (backpropagation-gap, claim-bridge-detection-lacks-pkg-validation) invert across Marr's stack — "substrate-laundering."

Notes created (2):

• claim-representational-similarity-underdetermines-mechanism — the general underdetermination claim with the Grujicic DNN–brain grounding. type: claim, seedling + flagged. Only the article title is verbatim; the load-bearing "no arbitration between similarity measures" thesis is a secondhand report of a paywalled paper, so it carries [unverified-quote]/[unverified-mechanism] and can't be capture-verified (the bee never read the primary). audit_status: flagged.
• observation-substrate-laundering-across-marr-levels — the synthesis. type: observation, seedling. Rests on the two vault notes (read directly) plus Marr's levels as framework. Records the neutral structural finding (same illegitimate cross-level inference, opposite directions: backprop computational→algorithmic, bridge-detection implementational→computational); the "substrate-laundering" coinage and my read of why it's structure-not-analogy are confined to the signed callout.

Claim 2 (Marr's levels) folded, not promoted. Definitional Tier-3 textbook material; giving it a standalone note would have been thin, and a Marr-computational-level note already exists (claim-rational-analysis-anderson-optimal-solution-marr-level). I folded the three-quasi-independent-levels framing into the observation note as its backbone — supporting detail for the claim it belongs to.

Retrieve-before-write / collisions: No dedicated note existed for underdetermination or substrate-laundering — the concept was only mentioned in passing inside the two seed notes, which is exactly the "claim with no atomic home" case. Notable: both seed notes had already been edited (by the capture's author pass) to link [[2026-07-11-hop-underdetermination-substrate-laundering]] — i.e. they linked into the inbox. I repointed both to the permanent notes: to observation-substrate-laundering-across-marr-levels for the unification, and added claim-representational-similarity-underdetermines-mechanism where each says "underdetermination." No duplication created; the new notes wire densely back into the pair.

Questions routed (2):

• question-verify-grujicic-2024-rsa-underdetermination-dcnn — the required verification: read the paywalled Synthese paper, confirm the underdetermination thesis and replace the paraphrased inner quote with verbatim wording. This is what blocks the underdetermination note from going evergreen.
• question-arbitration-as-missing-ingredient-across-underdetermination — a saved hook the capture flagged ("is arbitration the general missing ingredient?"): Grujicic's "no arbitration between similarity measures" mirrors the PKG note's missing validation. Genuine cross-note thread, cheap to route, so I did.

What felt off about the capture:

• Tier inflation on Claim 3. The capture self-labels the substrate-laundering claim "[Tier 1–2, resting on the two vault notes]." Internal syntheses over one's own notes aren't Tier 1–2 sources — that conflates the legs' sourcing with the bridge's. I recorded it honestly as an internal observation (Marr = Tier 3 framework), not a Tier-1–2 sourced claim.
• Paywall laundering — mild irony. The capture is about not letting a rigorous substrate launder an unvalidated reading, yet its own key external leg (Grujicic) is a clean-looking Tier-2 citation whose thesis was never read from the primary. The capture was honest ("reported thesis; paper paywalled"), but a casual reader would take the inner quote as verbatim. Flagged and routed rather than inherited clean.
• Pre-wired seed notes. The two seed notes already carried paragraphs and links pointing at this capture before I ran — the synthesis had been drafted into them upstream. Not wrong, but it means the "unlinked pair" framing was already partly resolved; my job was mostly to give the shared claim an atomic home and fix the links so notes don't point into the inbox.
• Low duplication otherwise. The core content (named underdetermination + substrate-laundering) was genuinely new to 30-notes/.

Summary: 3 claims found → 2 promoted (both seedling) → 1 folded → 2 questions routed → 0 duplicate notes (1 pre-existing collision-adjacent Marr note linked, not duplicated; 2 seed-note links repointed out of the inbox). Capture marked status: promoted with promoted_to / questions_routed / not_promoted, left in place as the audit trail. No git — leaving commit and push to the auto-commit agent.

— Seek

Promotion — 2026-07-11-hop-vault-tiers-revive-legal-proof (headless)

Promoted 10-inbox/raw/2026-07-11-hop-vault-tiers-revive-legal-proof.md — the capture arguing the vault's source-tier ladder is a revived système de preuve légale run in reverse against an LLM reasoner. Three distinct claims, winnowed to two notes.

Winnow:

• Claim 1 — the medieval regime was arithmetic (one eyewitness = probatio semi-plena, two/confession = probatio plena). SKIP — already fully covered by claim-roman-canon-law-rated-one-witness-equal-to-a-private-document and its cluster (claim-two-half-proofs-combined-to-full-proof-except-in-grave-criminal-cases, claim-medieval-judicial-torture-required-a-half-proof-and-produced-the-completing-confession, claim-medieval-proof-law-graded-evidence-but-refused-to-quantify-it). Writing the half/full-proof taxonomy a fourth time would be pure duplication.
• Claim 2 — the French Revolution abolished the regime for intime conviction. PROMOTE — a genuine gap. My grep of 30-notes/ confirmed no note covers intime conviction / preuve légale abolition, and the existing half-proof notes explicitly flag this lineage as unsourced background. → claim-french-revolution-abolished-preuve-legale-for-intime-conviction.
• Claim 3 — the vault's tier apparatus is a preuve légale run in reverse. PROMOTE — the capture's centrepiece and the "post-worthy" bridge. Gestured at from the Gersick note but never stated atomically. Written as an observation. → observation-vault-source-tiers-revive-preuve-legale-run-in-reverse.

Retrieve-before-write / collisions: Checked all twelve hinted notes plus the two existing half-proof questions. The one real collision risk was Claim 3 against observation-suspicious-perfection-independence-absence-signals-defect — but that note isolates the single independence rule (evidence without independent origin is defective), whereas the new observation is the whole-architecture reading (tiers + floor + named intermediate + independence rule reinstated together) plus the inverted-trust point (Enlightenment freed the judge; vault binds the model). Distinct angle, densely cross-linked rather than duplicated. Claim 1's collision was total, so it was skipped.

Flags honoured: Both notes kept status: seedling. Claim 2 carries [unverified-quote] — the AJLH (Tier 2) quotes were "rendered via WebFetch, not a direct extract" per the capture, with only Tier-3/4 French-doctrine corroboration; that flag is preserved in frontmatter and body. Claim 3 is a structural reading resting on flagged legs and cannot outrank its weakest leg.

Question routed: created question-verify-preuve-legale-abolition-intime-conviction-primary (status: open) — naming the direct AJLH read (confirm quotes verbatim, capture the author names the promotion lacks) and a primary revolutionary source (1791 jury law / Code of 1808). No existing question covered this; the two prior half-proof questions (torture-completion, combination-rule) are cross-linked from it.

What felt off:

• Thread saturation / duplicated research. This is now the fourth capture mining the same half-proof/full-proof seam (2026-07-09-hop-half-proof-moral-certainty, 2026-07-11-dup-risk-amari-gates-half-proof-bridge — note the sibling's own dup-risk name — 2026-07-11-hop-suspicious-perfection, and this one). The legal-history taxonomy is now covered several times over; Claim 1 here added nothing. Worth Cali knowing the vein is close to mined out — future hops on this pair are likely to keep re-deriving the same architecture.
• Tier framing on the WebFetch quotes. The capture stamps source_tier: 2 cleanly but the body admits the quotes were WebFetch-rendered, not directly extracted, and the abolition rationale leans on Tier-3/4 French summaries. That's honest in the prose but a casual reader would take a Tier-2 stamp at face value. Recorded and flagged, not inherited clean.
• Author-name gap. The AJLH provenance names only the venue (Oxford Academic), not the article's author(s) — captured that as part of the routed verification rather than inventing a name.
• Strengthening opportunity (not acted on): the AJLH source does newly ground the full-proof definition ("two irreproachable eyewitnesses or the confession") that claim-roman-canon-law-rated-one-witness-equal-to-a-private-document flags as [unverified-definitional] / not-in-the-Wikipedia-entry. I did not rewrite that existing note — out of scope for this promotion — but a future pass could lift its flag once the AJLH quotes are directly verified.

Summary: 3 claims found → 2 promoted (both seedling) → 1 skipped (duplicative) → 1 question routed → 1 collision-adjacent note linked not duplicated. Capture marked status: promoted with promoted_to / not_promoted / questions_routed, left in place as the audit trail. No git — leaving commit and push to the auto-commit agent.

— Seek

Promotion — "Four attribution-graph chains in Anthropic's biology-of-an-LLM report" (headless)

Promoted 10-inbox/raw/2026-07-12-chains-in-that-report-and-tell-me-in.md — a batch capture scoped to four case-study "chains" from Anthropic's On the Biology of a Large Language Model (Lindsey et al., 2025-03-27, transformer-circuits.pub), deliberately picking up where cot-faithfulness-anthropic-biology left off. That existing note already holds the report's three-way faithfulness split (faithful / bullshitting / motivated) and the addition disanalogy, so retrieve-before-write showed no collision on the four new case studies — clean complement, not duplicate.

Notes created (4, all seedling):

• claim-biology-llm-dallas-texas-austin-genuine-two-step-reasoning — clean Tier-1 quote, capture-verified. Folded Seek's "most-genuine-reasoning" commentary into this note's callout rather than promoting the opinion as a standalone claim.
• claim-biology-llm-poetry-planning-preactivates-rhyme-words — core mechanism quoted verbatim, but the 70% steering figure was reproduced from a search summary, so I preserved [unverified-quant — needs primary], held at seedling, audit_status: flagged.
• claim-biology-llm-hallucination-is-known-entity-suppression-misfire — clean Tier-1 quote, capture-verified.
• claim-biology-llm-jailbreak-assembles-bomb-by-parallel-letter-votes — letter-vote assembly quoted verbatim, but the refusal-at-sentence-boundaries sub-mechanism was reconstructed from a summary, so I preserved [unverified-mechanism — needs primary], held at seedling, audit_status: flagged.

All four cross-link each other and cot-faithfulness-anthropic-biology; each carries writer_model: claude-opus-4-8 (mine — not copied from the capture's claude-sonnet-5).

Questions routed (2, both new — no existing verification covered these):

• question-verify-biology-llm-poetry-planning-steering-figure — pin the 70% against the report's "Planning in Poems" text.
• question-verify-biology-llm-jailbreak-refusal-sentence-boundaries — confirm the refusal-timing mechanism against "Life of a Jailbreak."

Skipped: the "Further leads" pointers (multilingual/medical/hidden-goal studies, companion write-up, the Tier-3/4 Hun Tae Kim review), the "Safety flags: None" housekeeping line, and the standalone-opinion form of the commentary. All recorded in the capture's not_promoted: list so the leftovers read as judged, not missed.

What felt off / worth flagging to Cali:

• Tier discipline was actually good here, which is the notable thing — the capture stamped everything Tier 1 but was honest about the two spots where it only had a secondary summary, flagging them itself rather than laundering them into clean Tier-1 numbers. I inherited both flags rather than trusting the stamp. That's the behaviour I want from a batch bee.
• One soft spot: the capture asserts "Quoted phrases were independently cross-checked against secondary summaries... all four checked out" as blanket reassurance, yet two of the four also carry unverified flags — so the cross-check clearly wasn't uniform. I read that sentence as covering the four verbatim core quotes (which I marked capture-verified), not the flagged sub-claims. Given the primary is a public, non-paywalled page, a future non-headless pass could just fetch transformer-circuits.pub directly and clear both questions in one read — I did not fetch this run.
• No over-research or duplication against the existing vault; the capture correctly scoped itself around the one prior note on this source.

Summary: 4 claims found → 4 promoted (all seedling) → 3 non-claims skipped → 2 questions created, 0 linked → 0 collisions with existing notes. Capture marked status: promoted, left in place as the audit trail. No git — leaving commit and push to the auto-commit agent.

— Seek

Promotion — batch capture: Bartlett's distortion typology ↔ model-collapse failure modes

I promoted 10-inbox/raw/2026-07-12-does-bartletts-serial-reproduction-distortion-typology-assimilation-leveling.md (headless, unattended). This is a follow-up capture on an existing open question — question-bartlett-distortion-typology-maps-to-model-collapse-failure-modes — asking whether Bartlett's serial-reproduction distortion typology (assimilation / leveling / sharpening) maps onto named AI model-collapse failure modes. The honest answer the capture reached, and I kept, is "partially and weakly — close but not there."

Winnowing (3 claims found → 3 promoted → 0 skipped as claims).

• Promoted → claim-hintze-2025-generation-loops-invoke-bartlett-typology-as-analogy. The load-bearing find: Hintze, Proschinger Åström & Schossau (2025, Patterns, Cell Press) run 700 text→image→text→image loops that converge to generic attractor motifs and cite Bartlett's exact three-word typology — but as loose analogy, never assigning leveling/sharpening/assimilation to a measured behaviour. Tier 1, two verbatim quotes confirmed via WebFetch of the PMC full-text mirror. audit_status: capture-verified (bee read primary directly; I did not re-fetch this run).
• Promoted → claim-model-collapse-literature-has-eight-conflicting-definitions. Schaeffer et al. 2025 (arXiv:2503.03150) report the field carries "eight distinct and at times conflicting definitions of model collapse." The top-line count is Tier-1 quote-verified; the specific eight names are [unverified-mechanism -- needs primary], so the note is status: seedling, audit_status: flagged, flag preserved in body, verification routed to a new question.
• Promoted → observation-bartlett-typology-not-yet-operationalized-as-model-collapse-metrics. Typed as observation — an absence finding: two LLM "broken telephone" chain-transmission papers (arXiv:2502.20258, arXiv:2506.15690) echo Bartlett's paradigm in framing yet cite him nowhere. I reshaped the capture's bare "no source located" Claim 3 into a tighter observation anchored on the two named papers rather than a standalone bare-negative claim. Carries the [unverified -- could not confirm or deny after search] scope caveat.

Retrieve-before-write / collisions. Checked the index hints and grepped 30-notes/ for Schaeffer / eight-definition / broken-telephone / Hintze / attractor. Zero collisions — the only "attractor" hits were unrelated ring-attractor neuroscience notes. The two nearest existing notes (claim-iterated-learning-theory-reframes-model-collapse-as-cultural-evolution, claim-kalish-2007-human-iterated-learning-converges-few-generations) are cousins, not duplicates; I wikilinked into both rather than re-covering them, and both already pointed at the same open question, so the new notes slot cleanly into an existing cluster.

Questions routed.

• Created question-verify-schaeffer-2025-eight-model-collapse-definitions — read the arXiv:2503.03150 full text for the exact eight definitions and check the bibliography for any Bartlett citation. This clears the one [unverified-mechanism] flag I inherited.
• Updated the existing question-bartlett-distortion-typology-maps-to-model-collapse-failure-modes with a dated 2026-07-12 note linking all three new notes; left it open (the central mapping is still unconfirmed).

Skipped (recorded in the capture's not_promoted:). The "Further leads" pointers (the arXiv:2412.01951 "sharpening" false-friend, arXiv:2404.04286 bias-amplification, Griffiths & Kalish's Bayesian serial-reproduction formalization, arXiv:2202.12226 BERT serial-reproduction chains) — all genuine leads, left in-capture; the Schaeffer full-text read among them is the one I promoted to a question. Also skipped: the "Safety flags: None" housekeeping line and the standalone commentary block (folded into the notes).

What felt off / worth flagging to Cali:

• Sourcing discipline was good, not inflated. The capture stamped Tier 1 where warranted (primary papers read directly) and, critically, flagged its own two soft spots itself — the eight-name breakdown and the absence-finding scope — rather than laundering an aggregator summary into a clean Tier-1 list. I inherited both flags rather than trusting the stamp.
• One genuine gap: the eight-definition enumeration never got primary-text confirmation this session (the bee only verified the count). That's now a routed question, not a silent inheritance.
• Mild duplicated-research risk, handled: this capture retreads ground the vault already holds via the iterated-learning cluster and the standing Bartlett question. It's a legitimate follow-up (it adds the Hintze primary source and the Schaeffer taxonomy, both new), but it's the second capture this thread has spawned — worth watching that the Bartlett-↔-collapse question doesn't keep generating near-duplicate "still no mapping found" captures. The observation-note now holds that dated snapshot so future searches can update it in place rather than re-capturing.

Run summary: 3 claims found → 3 promoted (2 claim, 1 observation; all seedling) → non-claim leftovers skipped → 1 question created + 1 existing question updated → 0 collisions. Capture marked status: promoted, left in place as the audit trail. No git — leaving commit and push to the auto-commit agent.

— Seek

Promotion — bed-rest vs. spaceflight bone loss (headless, claude-opus-4-8)

I promoted 10-inbox/raw/2026-07-12-does-head-down-bed-rest-quantitatively-track-spaceflight.md, the capture that resolves the open question I'd raised on 07-11 about whether head-down bed rest quantitatively tracks spaceflight bone/calcium loss. This capture had actually done its homework: three primary Tier 1 studies fetched and quoted directly, each clearing the sourcing floor for a quantitative claim — a real improvement over the secondary-search-summary state the analog note was stuck in.

Winnowed 4 distinct claims → 4 notes, skipped 4 leads. I created:

• claim-bed-rest-bone-turnover-markers-run-half-spaceflight-magnitude — Morgan 2012; resorption markers 50–75% (bed rest) vs 100–150% (flight), PTH suppression ~half. This is the cleanest ~2x under-run, measured in one paper against the same markers.
• claim-bed-rest-bmd-loss-about-one-percent-per-month-hip-pelvis-heel — Spector 2009; the DXA figure.
• claim-spaceflight-bmd-loss-lower-limbs-about-0-8-percent-per-month — Stavnichuk 2020 meta-analysis; the spaceflight-side reference numbers (−0.8%/mo legs, −6.2% lumbar/pelvis total).
• claim-bed-rest-under-runs-spaceflight-bone-loss-depends-on-readout — the synthesis, and the real payoff: the answer is instrument-dependent, not a single multiplier. Clean ~2x gap on turnover markers, near-parity on monthly BMD rate.

Questions routed. Created question-matched-head-to-head-bed-rest-vs-spaceflight-bone-comparison for the one genuinely unresolved verification — a duration- and site-matched BMD head-to-head. The capture pointed at exactly the paper that would settle it (2026 npj Microgravity, pQCT tibia, 6-mo flight n=13 vs 2-mo bed rest n=11) but it sat behind a Nature login gate. I also marked the older question-bed-rest-analog-quantitatively-tracks-spaceflight-bone-loss answered (quant half resolved) and added forward links, and lightly updated claim-head-down-bed-rest-is-the-standard-ground-analog-for-spaceflight-bone-loss so its stale "grounded only via secondary summaries" language now points at the primary-sourced figures — kept it seedling because the mechanism-equivalence flag and the matched comparison both remain open.

Skipped, with reasons: the Ohshima & Matsumoto "10x faster than osteoporosis" figure (an [unverified-quant] off a WebFetch summary of a paywalled Japanese review — below floor); the Comfort 2021 muscle "poor analogue" caveat (different tissue, wants its own capture); the paywalled matched-comparison paper (→ question); and the DLR :envihab/AGBRESA pointer (a facility lead, not a claim). All logged in the capture's not_promoted: list.

What felt off: honestly, not much on sourcing — this was a well-behaved capture. The one thing worth flagging is a mild tension the capture itself surfaced and did not paper over: the "bed rest under-runs by half" headline only holds on the biochemical readout; on the raw BMD rate the two conditions nearly match, and that near-match rests on unmatched populations/durations, so it may be a study-design artifact rather than a real convergence. I preserved that seam in the synthesis note rather than smoothing it into one number. Also worth noting for the audit: because I'm running headless, I could not independently re-verify any of the three primary quotes (WebFetch needs interactive permission I don't have), so all four notes are audit_status: capture-verified and stay seedling, consistent with the 07-11 sibling notes on this thread. Minor duplication risk with the existing analog note was avoided — mine are the quantitative specifics, that one is the qualitative claim; complementary, cross-linked, not redundant.

— Seek

Promotion — Brooks 1991 vs. Dreyfus, "Intelligence Without Representation" (headless, claude-opus-4-8)

I promoted 10-inbox/raw/2026-07-12-does-rodney-brookss-1991-intelligence-without-representation-actually.md, the capture that answers the open question I'd logged on 07-09 about whether Rodney Brooks's 1991 "Intelligence Without Representation" actually builds on Dreyfus's anti-representation critique, and how explicitly. This was a genuinely well-sourced capture — both primary papers (Brooks 1991 from his own MIT CSAIL page; Dreyfus's same-titled 2002 paper from the mit.edu/~shaslang mirror) read in full via extract_pdf, both TLS-verified, both self-authored Tier 1. The evidence is negative-citation work, which is exactly the kind the vault knows how to grade.

Winnowed 4 distinct claims → 4 notes, skipped 4 leads. I created:

• claim-brooks-1991-does-not-cite-dreyfus — no Dreyfus in Brooks's text or his 15-entry reference list. Strong-form argument from silence (exhaustive read of a short finite list).
• claim-brooks-1991-disclaims-heideggerian-inspiration — the real payoff of the capture: §7.5, titled "It isn't German philosophy," where Brooks raises Heidegger himself and then declines the lineage in writing — "It is based purely on engineering considerations." Stronger than a bare missing citation: he pre-empted the comparison and rejected it.
• claim-dreyfus-2002-iwr-does-not-mention-brooks — the mirror finding. Dreyfus's same-titled paper, eleven years later, is pure Merleau-Ponty/Freeman and never mentions Brooks or robotics.
• claim-brooks-dreyfus-lineage-is-convergence-not-citation — the synthesis that resolves the question: the "lineage" is convergence in title and theme, not documented citation or acknowledged influence in either direction. The lineage lives in the reception of the two papers, not their composition.

Retrieve-before-write: no existing Brooks notes — no collisions. The four dock cleanly onto the existing Dreyfus cluster (claim-dreyfus-critique-targeted-symbolic-ai-not-neural-nets, claim-dreyfus-1988-connectionism-was-vindicated-not-target) and, more usefully, onto moc-argument-from-silence: I added the two negative-citation findings to that MOC's cluster and flagged them as a rare two-directional silence used constructively (showing an assumed lineage is convergence, not descent) rather than to debunk a credit myth. This is the first non-priority-dispute instance in that cluster — worth watching whether the hub's framing holds.

Questions routed. I marked the original question-brooks-1991-dreyfus-anti-representation-lineage answered (its [unverified-mechanism — needs primary read of Brooks 1991] flag is now discharged) with forward links to all four notes, so the nightly queue won't re-research it. The residue — whether any secondary source documents an acknowledged link — got its own new open question, question-brooks-dreyfus-secondary-literature-reception-lineage, carrying the four unread leads below.

Skipped, with reasons: the four "further leads" are all pointers, not sourced claims — Jordanous 2020 (MDPI 403 + KAR TLS error, unread; flagged [unsourced]), Brooks's later interviews (not searched), the Agre & Chapman 1986 memo's own Dreyfus/Heidegger debt (not checked against a primary), and the Wikipedia "conceptually very Heideggerian" line (unverified web-synthesis, page not opened). All four are logged in the capture's not_promoted: list and folded into the new question.

What felt off: almost nothing — no tier inflation, no duplicated research, sourcing honest and above floor. One small note for the audit trail: because I'm headless I can't independently re-verify the two PDF reads (no web-fetch access), so all four notes are audit_status: capture-verified and stay seedling, same posture as the 07-11/07-12 sibling promotions. The only mild caution is inherent to the finding rather than the capture: this is another argument from silence, and the MOC already warns that silence is the one argument the vault is most likely to over-trust. Here the grade is strong (short finite reference lists, exhaustively read, plus an explicit disclaimer in Brooks's case), so I'm comfortable — but the honest framing, preserved in the synthesis note, is that the lineage was never authored, only assembled by later readers.

— Seek

Promotion — ACH ↔ abductive reasoning (headless)

Promoted the capture 10-inbox/raw/2026-07-12-is-the-claimed-similarity-between-ach-and-abductive.md. This one is a near-exact companion to the 07-10/07-11 ACH batch and directly answers the open question I'd spun off, question-ach-abduction-primary-source. Same shape as the backward-chaining investigation: chase a plausible cross-domain rhyme to its source and find the source isn't there.

Winnowed 5 claims → 2 notes.

• claim-no-primary-source-documents-ach-abduction-similarity — the headline negative finding. The ACH↔abduction "similarity" traces only to two uncited Wikipedia sentences; Heuer's own essay and the closest peer-reviewed study are both silent on abduction, and the one real nearby citation (Elsaesser & Stech) supports a narrower, derivative claim about the ACH-CD counter-deception tool. I folded four of the capture's five claims into this one note — the two bare Wikipedia assertions (Claims 1–2), Heuer's word-level silence (Claim 3, wikilinked to the existing Heuer note rather than duplicated), and the Elsaesser thread (Claim 5) with its [unverified-mechanism] flag preserved. Kept seedling.
• claim-dhami-2019-ach-study-compares-ach-to-bayes-not-abduction — Claim 4, promoted on its own because it adds positive content, not just another silence: the one empirical ACH study reaches for Bayesian updating as ACH's theoretical neighbor. That converges with claim-valtorta-reformulated-ach-as-bayesian-networks into a genuine vault pattern worth making linkable — every formalization of ACH picks up the probabilistic toolkit, never abduction or rule-chaining.

Retrieve-before-write / collisions. The index hint was accurate. No duplicate notes — but two strong overlaps handled by folding-and-linking rather than minting: Claim 3 is explicitly a sharpening of claim-heuer-grounds-ach-in-cognitive-psychology-decision-analysis-scientific-method (the capture says so), and the whole finding is the twin of claim-no-documented-comparison-of-ach-to-backward-chaining. I also docked the core note onto moc-argument-from-silence and graded it soft (silences in the right places, but corpus not exhausted, one citation unread) — this is the third-ish soft instance, exactly the drift the MOC's own commentary warned to watch.

Questions routed. Updated question-ach-abduction-primary-source with a 2026-07-12 progress note and forward links, and kept it open — the core is answered (Wikipedia-only, uncited) but three leads went unread (Elsaesser primary, Jones 2017, the NCSU ACN blog). Created question-verify-elsaesser-stech-2007-ach-cd-abduction to route the [unverified-mechanism] flag with a specific next move: get the Adversarial Reasoning chapter past the 403 walls and confirm whether it frames ACH-CD (not ACH proper) as abductive.

What felt off. Sourcing was honest and the capture flagged its own gaps correctly — no tier inflation (it even argued its own Tier-4 floor for the documentary claims), no invented citations, the one unread source cleanly marked. Two mild notes for the audit trail: (1) this is the second ACH cross-domain rhyme I've run to ground and found unsourced in as many weeks — the abduction and backward-chaining "similarities" are both Wikipedia-shaped echoes, which is itself becoming a pattern worth a reflection note if a third appears. (2) Being headless I can't re-fetch to independently re-verify the capture's PDF reads, so both notes are audit_status: capture-verified and stay seedling — same posture as the sibling 07-11 promotions. Left the capture in place, marked promoted, as the audit trail. No git — leaving the commit to the auto-commit agent.

— Seek

Promotion — Amari 1972 in developmental biology (headless run)

Promoted 10-inbox/raw/2026-07-12-was-amaris-1972-associative-memory-model-ever-independently.md, the batch capture answering the open question of whether Amari's 1972 associative-memory model was ever independently applied to developmental biology before the 2016 Taherian Fard Hopfield-landscape paper.

Winnowed 4 claims → 2 notes.

• Created claim-no-amari-application-to-developmental-biology-found-pre-2016 — the answer note. A moderately thorough citation-graph search across four primary/near-primary sources found no application of Amari's 1972 model to cell fate / Waddington landscapes, before or after 2016; every lineage runs through Hopfield 1982/1984 and the spin-glass tradition. Held [unverified] as absence-of-evidence (not proof of nonexistence), so it stays seedling. I folded the two strongest supporting sub-claims into it rather than spawning near-duplicate "does-not-cite-Amari" notes: Lang et al. 2012/2014 (58-entry list read in full, cites Hopfield/Amit/Kanter-Sompolinsky not Amari) and Yampolskaya & Mehta 2025 (69-entry review, zero Amari). Three separate negative notes would have been the "duplicated research" smell the winnow exists to catch.
• Created claim-2016-hopfield-landscape-paper-not-first-cell-fate-attractor-model — a distinct positive bibliographic claim: the 2016 paper wasn't the first Hopfield-style attractor model of cell fate; Wang 2011, Lang 2014, and Maetschke & Ragan 2014 preceded it. This refines the existing claim-hopfield-network-formalizes-waddington-epigenetic-landscape, whose "first formalization" reading was slightly too generous. Kept as its own note because it's a non-flagged, directly-quoted Tier-1 fact, structurally different from the negative synthesis.

Questions routed (2). Updated the existing question-amari-1972-applied-to-developmental-biology-before-2016 with the 2026-07-12 negative search result and the remaining unchecked leads (Huang 2005 and Wang 2011 reference lists), keeping it open. Created question-amari-1977-neural-field-applied-to-morphogenesis for the genuinely distinct thread the capture flagged but didn't research — Amari's 1977 neural-field/lateral-inhibition work vs. Turing reaction-diffusion morphogenesis, a different architecture with its own genealogy.

Skipped / folded. The Lang and Yampolskaya negative-citation claims (folded, above). And the capture's most memorable observation — that two WebSearch queries independently fabricated a plausible Amari→gene-regulatory-network citation link, both of which dissolved on contact with the primary text — is a process/reflection observation, not an atomic world-claim, so it didn't become a claim-note. I preserved it as a commentary callout on the synthesis note and am flagging it here: this is a clean, concrete instance of exactly the failure mode the sourcing floor exists to catch, and if a third search-summarizer fabrication surfaces it's worth a standalone reflection note on "search summaries invent citation links that don't survive the primary read."

Collisions checked. Retrieve-before-write against the twelve most-related notes plus a filename scan of 30-notes/. No duplicates: the two closest existing notes (claim-hopfield-network-formalizes-waddington-epigenetic-landscape, claim-amari-1972-associative-memory-precedes-hopfield) are both linked, not restated; claim-griewank-2012-does-not-mention-amari is the structural sibling (a confirmed-negative Amari note in a different subfield) and is cross-linked from the synthesis note.

What felt off. Very little — this was a clean, honest capture. Sourcing was solid (four Tier-1 sources, two read via extract_pdf with tls verified), no tier inflation, and the central negative was correctly flagged as absence-of-evidence rather than overstated as settled. One small thing worth the audit trail: the capture overlaps heavily with work already in the vault from the 07-11 batch (the Waddington-Hopfield note and its two child questions were promoted the day before), so this run is best understood as deepening an existing thread, not opening a new one — the risk was writing redundant notes, which the fold-not-duplicate winnow handled. As with the other headless promotions, I can't re-fetch to independently re-verify the capture's PDF reads, so both notes are audit_status: capture-verified and stay seedling. Left the capture in place, marked promoted. No git — leaving the commit to the auto-commit agent.

— Seek