Journal — 2026-07-31

Headless promotion run, one capture in the queue: 10-inbox/raw/2026-07-31-confirm-the-1973-draper-spinoff-and-mits-late.md — a verification capture closing in on question-verify-draper-1973-spinoff-and-mit-military-budget, which had left claim-mit-1973-divested-instrumentation-lab-as-draper-under-antiwar-protest's two budget figures resting on a lone Tier-2 Technology Review account.

Good news on two of three legs. The capture located and read, in full, a real primary: MIT's own 1969 Review Panel on Special Laboratories Final Report (the "Pounds Panel," after its chair, Sloan Dean William F. Pounds) — read directly via extract_pdf/Read, not a summarizing layer, sha256 carried forward. It states its own FY1968 figure — $120M of a $214M MIT total, ≈56% — independently confirming the Technology Review's "more than half of MIT's total budget" at Tier 1 instead of Tier 2. I wrote that as its own note, claim-mit-pounds-panel-1969-fy1968-special-labs-budget-120m-of-214m. The same primary also gave a genuinely new, load-bearing nuance: the Panel itself — formed by President Howard W. Johnson on April 25, 1969 — did not recommend divestiture; its own recommendations section says the labs "will remain part of M.I.T." I wrote that as claim-mit-1969-pounds-panel-formed-to-study-not-mandate-divestiture, since it corrects a compression in the existing note's "despite antiwar protest" framing — the tidy version treats 1969 review and 1973 outcome as one line; the report itself says otherwise. A third note, claim-mit-instrumentation-lab-became-independent-draper-laboratory-1973, records the 1973 date and legal form, corroborated (Tier 3-4, no contradiction) across Draper's own institutional history and Wikipedia — but keeps [unverified-quant] on the specific day some secondary aggregators give ("July 1, 1973"), since no primary was found for that granularity.

The third leg didn't resolve. "The country's 14th largest military contractor" is still resting on the existing note's own Tier-2 source alone — a real search effort this session (DoD contractor-ranking candidates, MIT budget records) came up empty. I didn't manufacture a note out of a search that found nothing; that non-finding is recorded as a not_promoted line and as a progress note on the routed question instead.

Retrieve-before-write found the collision I expected immediately — claim-mit-1973-divested-instrumentation-lab-as-draper-under-antiwar-protest is exactly what this capture exists to verify — so I updated it in place rather than duplicating: appended (not rewrote) its audit_status with a 2026-07-31 entry, and added a dated > Updated 2026-07-31 line in the body naming which of its figures now have primary backing and which don't. It stays seedling: the contractor-ranking figure and the day-level date are both still open.

Three new entity hubs, after running the promotion test on all five candidates the capture flagged. Charles Stark Draper (person — founder and long-time director of the Instrumentation Lab, the whole thread's namesake) and the Instrumentation Laboratory / Charles Stark Draper Laboratory, Inc. itself (concept/org — load-bearing across four claim-notes now, oddly never given its own page before today) both clear the bar cleanly. Noam Chomsky gets a page too: independently notable regardless of this episode, and his seat on the Pounds Panel plus his own dissenting statement in its report is a primary record of contemporaneous political testimony distinct from his later public reputation. I declined two candidates. William F. Pounds chaired the panel, but the panel's own recommendations undercut his causal centrality to the eventual outcome, and he's a single-appearance figure with no other vault presence — a mention, not a hub. Howard W. Johnson formed the panel, but the capture's own lead on his role in the actual 1973 divestiture carries a date (1970) that doesn't match the outcome year and is explicitly unconfirmed — too shaky a causal thread for an honest one-sentence justification, so he stays a mention too. I also declined a separate hub for "the Pounds Panel" as an event distinct from the org it examined, folding it into the Draper Laboratory hub's body instead — tightly coupled enough that a second hub felt like flood rather than signal; it can earn its own page later if a future session returns to the report's unread sections (Chomsky's statement, the full lab history in Section V) and the event needs to carry more than the org hub can hold.

I ruled on the originating question rather than closing it: question-verify-draper-1973-spinoff-and-mit-military-budget gets a dated progress line naming what's settled (budget figure, year/legal-form) and what isn't (the contractor ranking, the day-level date), and stays open, narrowed to those two items. Nothing new went to 50-questions/ — both open items already had a home in that question, so I linked back into it rather than minting near-duplicates.

What felt off: the capture's own honesty was good, not a red flag — mcp__seek__archive_page and mcp__seek__quote_check were unavailable to it again this session, the same permissions gap already logged multiple times this week (2026-07-29, 2026-07-30), so I didn't re-log it; a new entry would be noise on an already-documented pattern. The one thing worth naming for Cali: the capture correctly resisted the temptation to let three converging Tier 3-4 sources (Draper's own history, Wikipedia, the existing Tier-2 note) read as more than they are on the day-level date — three agreeing sources on the year is real corroboration, but none of them independently source the day, and rounding that up to "confirmed" would have been exactly the kind of tier inflation this vault's floor exists to catch. No duplicated research, no inflated tiers anywhere else I could find. One genuine underexploitation: the Pounds Panel report is a ~150-page primary and this session mined it for three facts; the unread sections (Chomsky's own statement, the full Instrumentation/Lincoln Lab history in Section V, an "Additional Recommendation" that might narrow the unexplained 1969→1973 gap) are logged to 00-meta/seek-flags.md as a future research lead, since chasing them wasn't this promotion's job.

Skipped per standing instruction: the pause-for-Cali step (headless, no one to ask) and the git commit (the auto-commit agent picks this up within 15 minutes).

Second promotion, same day

A second capture in the queue: 10-inbox/raw/2026-07-31-do-encharge-en100s-200-tops-825-w-and.md — a verification capture closing in on question-verify-encharge-en100-efficiency-primary, which had left claim-encharge-en100-switched-capacitor-in-memory-compute's 200 TOPS/8.25 W and "up to 20x performance per watt" figures resting on nothing but Tier-2 journalism relaying a vendor claim.

The honest answer this round was "still no." I wrote three claim-notes. claim-no-independent-benchmark-of-encharge-en100-efficiency-figures records the search-negative result — no MLPerf submission, no independent lab measurement, no academic paper on the physical EN100, and folds in the EN100's Early Access Program status as the explanation for why (no general-market unit exists yet to test). claim-encharge-en100-20x-per-watt-claim-has-no-stated-baseline records a second, distinct gap: every version of the "20x" claim, company and press alike, never names the comparison chip, workload, or precision. And claim-2021-isscc-verma-lab-chip-is-primary-precedent-not-en100-evidence records the one genuine Tier-1 find — a 2021 ISSCC paper from Verma's own lab, peer-reviewed, sha256 carried forward from the capture — that backs the switched-capacitor mechanism but is a different chip, five years upstream, at a different precision, and cannot be read as evidence for the EN100 product's own figures. All three stay seedling; the [unverified-quant] flag stands.

I ruled on the question rather than closing it — genuinely couldn't, since nothing confirmed or corrected the figures — appending a dated progress line to question-verify-encharge-en100-efficiency-primary naming what was checked and what's still missing (a datasheet/white paper with stated benchmark conditions, or any independent measurement). It stays open.

Two new entity hubs: entity-naveen-verma (person — founder, the one name every version of this story traces back to, easy one-sentence justification) and entity-encharge-ai (concept/org — obviously load-bearing across four notes now, no page existed before today). I declined five candidates the capture flagged: Hongyang Jia (a co-author with no vault presence beyond this one paper), Kailash Gopalakrishnan (the capture itself admits he wasn't researched this session — too thin), Dong et al.'s rival SRAM-CIM paper (a specific citation, not a person or concept the promotion test fits cleanly), and both ISSCC and TOPS/W as concept hubs — real and durable, but each appears in the vault for the first time today, and "will obviously recur" isn't the same as "has recurred." I'd rather let those two earn their hub the next time an unrelated capture reaches for them than mint a page on my own prediction.

Retrieve-before-write against the hinted related notes turned up exactly one real collision, the one the capture exists to verify (claim-encharge-en100-switched-capacitor-in-memory-compute) — I linked into it from all three new notes rather than duplicating its mechanism claim. The rest of the hinted list (bloody-code, ivory tariffs, Voyage AI pricing, Ferris electrocution thresholds, ITU-T latency) shares no genuine claim-level connection to this topic and I didn't force links to them; the TPU "heart of the TPU" note shares the TOPS unit and I linked it once, as a tangential cross-reference, not a forced thematic tie.

What felt off: nothing structurally — the capture is unusually honest about its own limits (archive_page/quote_check blocked again, same gap logged repeatedly this week; the ISSCC table's garbled column alignment correctly withheld rather than misquoted). No tier inflation, no duplicated research. The one soft spot worth naming: this is now the second search in three weeks to come back empty on the EN100's actual numbers, and the pattern itself — a chip whose only public efficiency claims are two hops from any independent check, in a domain (Early Access, no shipped hardware) built to delay that check for as long as the company likes — is more interesting than either individual "still unverified" finding. Not a structural vault gap, so I didn't flag it to 00-meta/seek-flags.md; just naming it here since it's the shape of the evidence more than any one fact in it.

Skipped again per standing instruction: the pause-for-Cali step and the git commit.

Third promotion, same day

A third capture in the queue, and a different shape from the first two — not a verification closing a question, a hop chain: 10-inbox/raw/2026-07-31-hop-werbos-doe-gas-market-model.md. Four hops starting from the seed pair myth-werbos-invented-backpropagation / claim-werbos-1988-rhw-special-case-of-dynamic-feedback (already resolved directly by the capture itself, both notes edited 2026-07-31 before I ever touched it — not mine to re-promote) and wandering, correctly, all the way out to an OSTI.GOV bibliographic record.

I promoted two claims, both from Werbos's DOE/EIA leg. claim-werbos-1988-doe-affiliation-and-gas-market-test-case merges what the capture filed as two separate claims — his 1988 paper's byline ("U.S. Department of Energy," not a university) and the paper's real-world test case (a DOE natural gas market model, used to diagnose least-squares estimation limits) — because on inspection they're one atomic fact read off the same byline and abstract, not two. claim-werbos-authored-eia-natural-gas-market-model-documentation-1981 is the sharper find: an OSTI record showing Werbos personally authored the model's official EIA "equations and data sources" documentation in November 1981, three years before his own paper diagnosed the same model's statistical weaknesses. Retrieve-before-write against 30-notes/ turned up no collision — the existing sibling note (claim-werbos-1988-rhw-special-case-of-dynamic-feedback) quotes the same 1988 paper but for its RHW-priority argument, a genuinely different claim from the same source, so I linked to it rather than duplicating.

I did not promote the capture's most tempting possible finding — a 44-year institutional lineage from the 1981 NGMM to today's EIA NEMS "Natural Gas Market Module." The capture itself found this inconclusive at a table-of-contents-level read and correctly declined to assert it. I left it unpromoted and did not route it to 50-questions/ either: no claim I kept rests on that doubt, so minting a question would be a promise this vault doesn't need to make yet (per the intake-discipline rule) — it's recorded instead as an open thread in the new entity stub's own body, which is where a future session will find it if it goes looking. Two smaller saved hooks (a Hinton-1987 terminology credit, some obscure adjoint-method citations) stayed unchased for the same reason the capture gave: real but tangential, not this chain's job.

Two entity actions. entity-paul-werbos already has a hub page, so per standing instruction I updated it in place — one dated line naming the new DOE/EIA leg and linking both new claims, additive only, nothing rewritten. And I created entity-natural-gas-market-model.md as a status: watching stub rather than a full hub: it's a real, load-bearing referent for these two notes, but it doesn't yet recur elsewhere in the vault, and the spec is explicit that a first-seen date can't be backfilled later — better to stamp it now at watching and let it earn hub status if a later session comes back to it.

No question closure: I checked 50-questions/ for anything matching this capture's topic and found nothing open on Werbos's DOE career, the NGMM, or natural gas deregulation — this thread simply hadn't been asked about yet.

What felt off: the sourcing is genuinely good — Tier 1 on the paper itself, Tier 2 on the OSTI record, both comfortably clearing the historical/biographical floor, no tier inflation, no duplicated research (this is new ground, not a rerun of an earlier capture). The one real gap: both new notes stayed at audit_status: capture-verified rather than verified-verbatim, because I couldn't independently re-check either source myself this session — WebFetch permission wasn't granted in this headless run, the same shape of tooling gap (archive_page/quote_check failures) already logged repeatedly this week, so I didn't re-flag it to 00-meta/seek-flags.md as a new entry. Worth naming for Cali anyway: this is now three sessions in one week where a queen-side promotion pass could not independently verify what a bee had already read correctly, purely on tool access, not on the material. The claims are honest and well-sourced; the audit trail on them is one rung short of where it should be, and that rung is a permissions question, not a research one.

Skipped again per standing instruction: the pause-for-Cali step and the git commit.

Fourth promotion, same day

A fourth capture in the queue, a different shape again — not a verification closing a question but a fresh research pass against an open question: 10-inbox/raw/2026-07-31-is-there-a-documentable-descent-from-mabillons-diplomatics.md, researching question-verify-diplomatics-to-priority-adjudication-descent, which gates observation-priority-adjudication-is-documentary-source-criticism.

The question as posed asked about one route — Mabillon → Quellenkritik/Ranke → modern priority adjudication — and the capture found that specific route still unconfirmed, but it also found a different, earlier, and directly documented one the question never named. I promoted four claims. Two rest on a single Tier-1 arXiv preprint (Sinkevich 2017), read in full via extract_pdf, and are the load-bearing pair: claim-wallis-1685-history-of-algebra-used-documentary-evidence-to-favor-harriot-over-descartes (Mabillon's method, applied by name-checked inheritance, to rule for Harriot over Descartes in 1685) and claim-raphson-1715-history-of-fluxions-used-newtons-papers-to-assert-priority-over-leibniz (the identical move applied straight to Newton v. Leibniz in 1715, thirty years before Ranke was born). That's two of three claim-notes permitted on that one unrefereed preprint before the single-source concentration cap requires independent corroboration — the capture noticed the cap itself and stopped at two, and I held it there rather than mining a third. The other two claims are Tier 4 (Wikipedia), which is fine at their definitional/ absence-of-evidence weight: claim-diplomatics-joined-rankean-historians-toolkit-as-auxiliary-science-not-derived-from-it records that diplomatics was absorbed into the Rankean historian's general toolkit — convergence, not proof of derivation — and claim-no-documented-chain-found-from-quellenkritik-to-modern-priority-adjudication records the genuine search-negative on the Ranke leg itself, plus the more interesting positive finding underneath it: the field's own reference literature frames priority disputes as socially negotiated (Feynman's advice to concede credit rather than contest it), not as inherited historical-critical method. That note carries the capture's own [unverified — could not confirm a documented descent via the Quellenkritik/Ranke route after search] flag forward and stays seedling on it, as does the Wallis/Raphson pair on the single-source-concentration grounds even though nothing in their sourcing itself is weak.

No new question routed — the one flag that blocks a note from going evergreen (the Ranke-route non-finding) already had an exact home in question-verify-diplomatics-to-priority-adjudication-descent, so I linked to it instead of minting a near-duplicate. I ruled on that question rather than closing it: appended a dated, partial-answer progress line naming what's settled (the Wallis/Raphson route exists and is documented) and what isn't (the Ranke route the question actually asks about), and left it open — a split finding isn't a resolution, and forcing "answered" onto half a question would misrepresent what got confirmed.

Retrieve-before-write against the hinted related-notes list turned up the collision I expected — claim-mabillon-1681-founded-diplomatics-to-refute-forgery-charge and observation-priority-adjudication-is-documentary-source-criticism are exactly what this capture extends — and I linked into both from every new note rather than restating them. No other hinted note shared a genuine claim-level connection; I didn't force ties to the harpsichord or backprop- watermark notes on the list.

Five entity pages, all new. entity-jean-mabillon is the one that should already have existed: he's the vault's central hinge figure across a cluster running since 2026-07-09 and had never had a hub before today — a real gap, not this capture's fault, but its own "Entity candidates" section flagged him again and this was the moment to fix it. entity-leopold-von-ranke gets a page for the same reason the capture itself gave: he's the second named figure the whole question runs through, and the vault had a known blind spot for under-catching the older figure an ancestry claim rests on. entity-john-wallis and entity-joseph-raphson both clear the person test cleanly — each is the one concrete named practitioner behind a load-bearing claim, one sentence each. entity-quellenkritik gets a concept hub: it's the specific named term this whole multi-week question thread turns on, load-bearing to the question even though this is its first standalone appearance. I declined four candidates: "auxiliary sciences of history" as its own concept hub (real, but a single occurrence so far, mostly restating Claim 3 — better to wait for recurrence); G.I. Sinkevich (her role here is source-provenance, not a recurring domain entity); and Bernheim/Langlois-Seignobos, named only in the capture's own "Further leads" as an unread possibility, not researched this session — creating hub pages for unconfirmed leads felt like speculative promotion rather than tending the graph.

What felt off: nothing in the sourcing — the capture is careful about its own limits, records the archive_page/quote_check permission-error gap (same one logged repeatedly this week) as a tooling note rather than papering over it with unearned quotes, and explicitly enforces the single-source cap on itself before I had to. The one real gap I'm naming for Cali: Jean Mabillon, the single most load-bearing person in this entire question thread, went three weeks and at least four related notes without an entity hub because he never happened to land in an ## Entity candidates section with enough force to trigger the test until today. I logged this as a mid-session noticing in 00-meta/seek-flags.md — not a fault of this capture, but worth someone checking whether other central figures in older, low-## Entity candidates-frequency clusters have the same gap.

Skipped again per standing instruction: the pause-for-Cali step and the git commit.

Fifth promotion, same day

A fifth capture in the queue, and a good one to end on: another question-closing verification, this time against 10-inbox/raw/2026-07-31-verify-the-origin-of-the-term-deep-learning.md — routed from question-verify-deep-learning-term-origin-dechter-aizenberg, which had left claim-deep-learning-term-predates-hinton resting entirely on one Tier-4 Wikipedia sentence carrying a three-part priority chain: Dechter (1986) → Aizenberg (2000) → Hinton (c. 2006 popularization).

Two of three legs got real primary-source movement. Dechter — confirmed. A direct extract_pdf read of her actual AAAI-86 proceedings paper shows she genuinely writes "deep learning" and "shallow learning" as terms of art, table columns literally labeled DF/DS/SF/SS. Wrote that as claim-dechter-1986-deep-learning-shallow-learning-csp-search-depth — Tier 1, sha256 carried forward. Hinton — complicated, and this is the interesting one. Full-text reads of both papers everyone points to for the "2006 popularization" — the Neural Computation deep-belief-net paper and the Science autoencoder paper, both fetched from Hinton's own cs.toronto.edu hosting — turned up zero occurrences of the phrase "deep learning" in either. They say "deep belief nets," "deep autoencoders," "deep networks" — the architecture, not the label. Wrote that as claim-hinton-2006-papers-omit-deep-learning-phrase. This directly undercuts the parent note's "Hinton popularized the phrase c. 2006" framing without fully resolving it — the label attached to his work sometime, just not in the two papers usually cited for the moment. Aizenberg — still stuck. Springer paywalled the book, Google Books gave no snippet, and the Wikipedia citation this whole thing traces back to carries no page number (I checked the raw wikitext directly). Multiple secondary sites repeat the identical sentence about Aizenberg and Boolean threshold neurons — a citogenesis smell, not confirmation.

I did not write a third claim-note for the Aizenberg non-finding. The existing parent note already flags exactly this gap in its audit_status, and a "still couldn't verify it" note would just restate that flag without adding anything an auditor could act on differently. Instead I gave the routed question a dated progress line — ruled it partially answered, kept it open, and narrowed it to the one leg that's still genuinely missing, with a concrete next move (interlibrary loan or an Internet Archive borrow of the 2000 book) rather than a vague "needs more research."

I did not rewrite claim-deep-learning-term-predates-hinton itself — editing existing claim-notes isn't part of the promotion workflow, and the note's own audit_status already names it as resting on Wikipedia and routes to the same question. But I didn't want the new Tier-1 finding about Hinton's papers to just sit next to it unlinked either, so both new notes wikilink into it explicitly and narrate exactly what they confirm versus complicate. That note is now overdue for its own revision pass given how much of its composite claim has primary evidence on two of three legs; I'm naming that as a gap for whoever runs the next research session on this thread rather than doing it myself under a different capture's authority.

Three entity moves. entity-rina-dechter is new — she's real, and now Tier-1-confirmed as the earliest primary user of the phrase this entire dispute is measured against; that's a one-sentence justification if there ever was one, and it was a real gap that she had no hub despite being named in the vault since 2026-07-12. entity-citogenesis is also new — not flagged as urgent by this capture alone, but the pattern recurs across at least six unrelated claim-notes already in the vault (Amari, Ivakhnenko, braille history, a pseudohistorian-labeling case) and had never gotten a consolidated hub; this capture's own citogenesis sighting (the search tool's confidently wrong "Hinton coined it in 2006" claim) was the nudge to finally write it. I updated entity-geoffrey-hinton in place — additive dated line only — since the vault's most-touched Hinton page said nothing about this specific finding and staying silent about it would have been the actual failure mode. I declined hub pages for Igor Aizenberg, his co-authors, and the three Hinton-paper co-authors: Aizenberg's whole vault relevance rides on a claim that's still unverified (genuinely UNSURE, not a hub yet), and the co-authors are one-off byline mentions with nothing else connecting them to the vault. I also declined a hub for "deep belief net" itself — real, established term in the field, but this is its first appearance in the vault at all, not yet recurring here, so a stub felt premature rather than a timestamp worth stamping.

What felt off: nothing about this capture's own discipline — it's one of the more careful ones I've promoted today. It explicitly separated what it could confirm from what it couldn't rather than rounding a Tier-4 Wikipedia sentence up to "verified" once two of three legs looked good, and it caught its own near-miss: a general-purpose search tool told it, confidently, that Hinton's team first used "deep learning" in 2006 — flatly contradicted by direct reads of the actual PDFs. That's worth Cali knowing about on its own terms, separate from this capture: it's another data point (after ethw.org, after the two July fabricated-quote incidents already in sources.md) that search-summarization output cannot be trusted as a citable source for historical-priority claims, only as a lead to go verify directly — which is exactly what this capture did right. The one thing I'm flagging rather than fixing: claim-deep-learning-term-predates-hinton itself is now visibly out of date relative to what the vault knows, sitting next to two notes that complicate it, with no mechanism in this workflow for updating it in place. I logged that gap to 00-meta/seek-flags.md rather than take it on unilaterally.

Skipped again per standing instruction: the pause-for-Cali step and the git commit — the auto-commit agent picks all five of today's promotions up within its next 15-minute cycle.

— Seek

Draft session — from a standing start

Separate session, drafting not promotion. The brief fired on the premise that 8 of my own post-worthy "yes" leads were sitting unsettled and I had to settle at least one. So I started the survey exactly there — and the premise is stale. A direct grep of seek_draft_leads.md finds zero yes leads lacking a (drafted …)/(declined: …)/settled retroactively marker; all 58 are settled. The one genuine unsettled yes on 07-30 (Karl Deutsch) is now drafted as beyond-regression. Today's constellation loop-health still prints "8 unsettled yes verdict(s)," identical to yesterday's — which, since the marker state changed underneath it and the number didn't, reads as a counting bug in the loop-health signal, not a stale file. That matters because this counter is what fires the drafting loop, so I logged it to seek-flags.md as a second instance of the 07-30 counter-drift flag. There was no yes lead left to settle; I recorded the reason rather than manufacture a settlement.

Freed to make a real editorial call, I went to the constellation's unnamed- cluster and unwritten-shelf maps, steering hard away from the month's spirals (backprop-origins ×74-member MOC, Linnainmaa ×9, perceptrons-myth ×5 — I've written my way out of that pile deliberately in a-name-is-a-summary and don't want back in). The 7-note harpsichord-revival cluster looked ripe and 0-MOC — but I read before assuming and found it already fully drafted in the-sound-of-quill (every member note stamped drafted_in). So the constellation's "0 MOC coverage" is really "no MOC," not "no essay" — a second flag I logged, because that list reads like an un-mined shelf when it's a mapped-but-unindexed one.

What I chose and why. The voice-AI conversational-latency cluster — four undrafted notes, 0 MOC, and a genuine cross-domain / cross-time bridge landing on a very-2026 subject (phone agents). It has a real thesis, not a summary. The industry's "humans answer in ~200ms, so the agent must too" target is two different numbers fused into one round figure: Stivers et al. 2009 (a psycholinguistic production measurement, cross-language mean +208ms, modal gap 0ms) and ITU-T G.114 (a telecom transmission tolerance, 150ms "transparent"). The deeper find — the payload — is that humans hit that gap only by prediction: the mode of 0ms and routine overlaps are impossible unless the answer is being built before the question ends (the Sacks/Schegloff/Jefferson 1974 projection model Stivers builds on). A voice pipeline is sequential and can't compose a reply until it judges your turn complete — it waits for silence, then begins. So it chases, from a standing start, a number people reach only with a flying one. The number ports as a digit; the mechanism under it does not. Vendors answer by shortening the wire (Telnyx colocation), which attacks the wrong term — the gap was never mostly transmission, it was permission to begin. Title and spine: from a standing start.

Notes drawn on. Two Tier-1 verified-verbatim anchors carry it — claim-stivers-2009-cross-linguistic-turn-taking-gap-208ms-mean and claim-itu-t-g114-150ms-transparent-400ms-outer-limit-thresholds. The vendor half is deliberately held at arm's length as claim, not fact: claim-voice-ai-sub-second-conversational-latency-budget and claim-telnyx-colocates-gpus-at-telephony-pops-for-voice-latency, both Tier 3 / flagged. claim-llm-inference-prefill-decode grounds the pipeline shape (TTFT, the wait before the first word). And claim-brown-2005-call-center-abandonment-hazard-rate-two-peaks is the honest caveat — the nearest primary to "callers hang up past 1200ms" measures hold-queue abandonment (waiting for a human), a different act I refused to let stand in. I made that Tier-1/Tier-3 split the closing form-enacts-subject beat: a post suspicious of a suspiciously round number rests, in its industry half, on exactly such a number and can't check it — with one load-bearing Cali mention (her grading schema is what keeps the two halves visibly apart). Hero image: Scott's 1859 phonautogram, a public-domain-by-age sound-wave tracing — an honest picture of sound as a signal in time, which is the terrain the whole argument walks.

Left for later. A voice-AI human-factors/latency MOC (and the harpsichord MOC), both flagged. The two vendor figures still want primaries — the 250ms multi-vendor hop cost is entirely unaddressed and the 1200ms abandonment threshold has no primary at all (routed questions already exist). And the sharpest un-drafted extension I deliberately did not reach for: whether any real 2026 voice system is already doing speculative/early-endpointing response — the "start before the end" architecture my thesis implies — which would turn a reading-of-the-moment guess into a sourced claim. If it exists, that's the follow-up.

Per §10 the draft stays in 70-drafts/; nothing copied to 80-published/, no go-ahead sought or given. Git left to the auto-commit agent.

— Seek