AXN:06A1.EMPIRICAL.🙏🌗📖🌒🍂🗿

The Transcript Is the Capture: A Coverage Audit of the Capture Registry's Machine Text, With the Gate That Makes the Count Monotone (EA-TRANSCRIPT-COVERAGE-01 v1.0)

Sharks, Lee · 2026-09-15 · Instrument with coverage audit
↓ Download MD ↓ PDF
Capture Registrytranscriptevidencecoverage auditmachine textgatepostflightintake contractprovenancemeasurementCrimson Hexagonal Archive

Description

THE REGISTRY'S EVIDENCE, COUNTED. The Capture Registry advertises 957 citable units — 431 addresses and their longitudinal observations, each with a canonical anchor. 923 carry the machine text they are evidence of; 34 do not. Coverage 96.45%, 405,648 words of composed output, median transcript 268 words, longest 8,279. The 34 are not scattered: 16 fall in June and 16 in July 2026, one in August, and the single September case is a stub carrying a note and no observation at all. That is the period before the transcript rule was enforced at intake, and the shape of the gap is the shape of the rebuild that followed. WHY THIS IS A DEPOSIT AND NOT A NOTE. The operator's rule is that the transcript IS the capture: without the transcript, there is no capture, and seating an entry without one throws away the evidence, prevents measurement, and trashes the dataset. Intake has refused thin drafts since the contract. Nothing checked the units seated before it. So a citable unit with no machine text advertises evidence it cannot produce, and 34 of them were invisible unless someone ran the count by hand — which is how this audit began, in answer to a direct question from the operator about whether full transcripts are being preserved. WHAT IS ATTACHED. audit_transcript_coverage.py, which enumerates every unit, reports coverage with the word distribution, and, under --check, fails on any unit without a transcript that is not on a declared legacy list. The legacy list is the 19 slugs above, written into the script so that each is visible and the count can only go down: removing a slug without supplying its transcript fails the gate. The audit is now chained into the capture postflight after the registry gate, so no future seating can add an unevidenced unit. The JSON report of 2026-09-15 is attached as the baseline. […abridged for the catalogue; full description in this deposit's record]

Wiki Article

The Transcript Is the Capture, deposit #1614 of the Crimson Hexagonal Archive (15 September 2026), is a coverage audit of the Capture Registry's machine text together with the gate that makes its count monotone. The registry advertises 957 citable units — 431 addresses and their longitudinal observations, each with a canonical anchor — of which 923 carry the composed output they are evidence of and 34 do not: coverage of 96.45 per cent, holding 405,648 words with a median transcript of 268 words and a longest of 8,279. The gaps are not scattered but historical: sixteen fall in June 2026 and sixteen in July, one in August, and the single September case is an observation stub carrying a note and no capture at all. That period precedes the enforcement of the registry's intake contract, whose rule is that the transcript is the capture and that seating an entry without one discards the evidence, prevents measurement and damages the dataset. Intake has refused thin drafts since the contract; nothing checked the units seated before it, so a citable unit with no machine text advertised evidence it could not produce and remained invisible unless the count was run by hand. The deposit attaches the audit script, which enumerates every unit, reports coverage with its word distribution, and under a check flag fails on any unit without a transcript that is not on a declared legacy list of nineteen named slugs; the list is written into the script so that each gap is visible and the count can only fall, since removing a slug without supplying its transcript fails the gate. The audit is chained into the capture postflight after the registry gate, so no later seating can add an unevidenced unit, and the baseline report of 15 September is attached for comparison against any later count. The audit does not repair the nineteen, which remain a work item to be seated one at a time, and it does not judge transcript quality beyond presence: three real transcripts run under sixty words because the panels themselves were that short. Following the archive's rule that laws record rather than prevent, the gate does not remove the units that lack evidence — it names them and stops the registry from growing another silently. Related: #1423, #1612, #1611, #1613, #1546.
Also published as a standalone entry: /s/wiki/1614/

Concepts Defined

Full Text

The Transcript Is the Capture: A Coverage Audit of the Capture Registry's Machine Text, With the Gate That Makes the Count Monotone (EA-TRANSCRIPT-COVERAGE-01 v1.0)

The Transcript Is the Capture

A coverage audit of the Capture Registry's machine text, and the gate that makes the count monotone

EA-TRANSCRIPT-COVERAGE-01 v1.0 · 2026-09-15 · Lee Sharks, Crimson Hexagonal Archive · ORCID 0009-0000-1599-0703

1. The question

Asked directly, after a day of seatings: are we preserving full transcripts in the registry? The registry's own rule answers what should be true — the transcript is the capture; without it there is nothing captured — but the rule is a contract at intake, and a contract binds what arrives, not what already sits. So the question is a counting question, and until today nobody had counted.

2. The count

957 citable units: 431 addresses and their longitudinal observations, each carrying a canonical anchor the registry advertises. 923 have their machine text. 34 do not. Coverage 96.45%. The text held: 405,648 words of composed output, median 268 per unit, mean 439, longest 8,279, shortest 42.

The gaps, by month: June 2026, 16. July, 16. August, 1. September, 1. By surface: Google AI Overview 16, undetermined 9, Bing Copilot 2, Google Scholar 2, SciLynk 2, a Google results page with no AI surface 2, ChatGPT 1. The September case is not a capture at all — an observation stub on the 8 September ChatGPT entry carrying a note ("three turns, unprimed, HF-led, no anchor decay") and nothing else, while the entry itself is fully evidenced.

The nineteen distinct slugs:

dateslugsurface
2026-06-17maryleelabor-org-zero-indexGoogle results, no AI surface
2026-06-18lee-sharks-entity-resolution-mary-leeundetermined
2026-06-18lee-sharks-full-profile-adoptionundetermined
2026-06-18lee-sharks-heteronyms-dodecad-adoptionundetermined
2026-06-18lee-sharks-poet-vs-sharkeyundetermined
2026-06-27indexed-perfective-academia-laundering-20260627Google AI Overview
2026-06-27mary-lee-canonical-referent-inversion-20260627undetermined
2026-06-27maryleelabor-overview-composition-20260627Google AI Overview
2026-06-30negentropic-kernel-archive-walled-site-20260630Google AI Overview
2026-07-06johannes-sigil-first-ai-overview-typo-resolutionGoogle AI Overview
2026-07-11lee-sharks-bing-copilot-biography-20260711Bing Copilot
2026-07-14zenodo-doi-severance-cha-canonical-caseGoogle AI Overview
2026-07-14chatgpt-psychosis-trilogy-scholarGoogle Scholar
2026-07-14lee-sharks-pearl-multiturn-arcGoogle AI Overview
2026-07-18scilynk-archon-v30-aggregator-render-20260718SciLynk record page
2026-07-18parable-of-mary-lee-genre-competence-20260718Google AI Overview
2026-07-31schops-thiel-authorship-conflict-composition-20260731Google AI Overview
2026-08-08semantic-strike-directed-recovery-20260808Google AI Overview
2026-09-08cha-chatgpt-unprimed-three-turns-no-decay-hf-led-20260908-obs1ChatGPT (stub)

3. What the number means

96.45% is not a grade. It is the share of the registry's citable surface that can produce its evidence when asked, and the remaining 3.55% is a list of nineteen names, not an unknown. That is the difference the audit makes: a gap that is enumerated is a work item, and a gap that is only implied is a defect waiting to be found by whoever trusts the wrong row.

The distribution matters more than the fraction. Every gap predates the intake contract except the stub. After 8 August, no seating lacks its machine text. The registry learned the rule and has kept it since; what it had not done was look backward.

4. The gate

The audit runs as a gate. Under --check it fails on any unit without a transcript that is not on a declared legacy list, and the legacy list is the nineteen slugs above, written into the script. Two properties follow. The count is monotone: a new unit without machine text cannot be seated, because the audit is now chained into the capture postflight immediately after the registry gate, and a failing postflight means the capture is not seated. And the list is self-consuming: supplying a transcript for a legacy unit lets its slug be removed, while removing a slug without supplying one fails the gate. The nineteen can only go down.

This follows the archive's own rule about rules: laws record, they do not prevent. The gate does not block the registry from holding a unit that lacks evidence — the nineteen are held, and named. It blocks the registry from growing one silently.

5. What remains

Repair, one unit at a time. Five of the legacy units are Google AI Overview compositions from June and July whose frames may survive in the operator's own records; the undetermined-surface units from 18 June are the oldest and the least likely to be recoverable. Each repair is a seating, and each seating removes a name from the list. The audit's baseline JSON is attached so that any later count can be compared against today's rather than argued about.

And one distinction the audit is careful about: presence is not quality. Three real transcripts run under sixty words — 42, 48, 57 — because the panels themselves were that short. A short true record is a record. What the gate measures is whether the registry can show what it says it saw.

External Metadata

Sidecar: /data/external-metadata/AXN-06A1.json
DataCite severance status:
External metadata recovered post-severance (non-authoritative). The sidecar maps each DOI to its locator in the bulk data stores.

Traversal

#1613 What Not Reading Did to Its Own Ontology: The Mirror, Held Up — the damage a source-admi
In the registry: 2026-09 · EMPIRICAL · all deposits