AI Overview Capture Registry — EA-WG-CAPTURES-01 v12.167 — 633 semantic addresses, 811 captures, 2026-06-07 to 2026-10-12

AI Overview Capture Registry — EA-WG-CAPTURES-01 v12.167 — 633 semantic addresses, 811 captures, 2026-06-07 to 2026-10-12

AI Overview Capture Registry — EA-WG-CAPTURES-01 v12.167

633 semantic addresses · 811 captures · 84 dates · 2026-06-07 to 2026-10-12

Lee Sharks · ORCID 0009-0000-1599-0703 · 12 October 2026 · supersedes v10.9 (#1461)

1. The citable unit

ONE SURFACE + ONE SEMANTIC ADDRESS + ONE DATE = ONE CAPTURE.

• The capture is citable at captures/{slug}/. Cite it for what a named system did.
• The address is citable at addresses/{slug}/. Cite it for what the address does across surfaces or over time.

Each address has had its own page since 2026-09-25. That page is the unit a fetcher can read whole, which the gallery is not. An address's slug is inherited from the first capture seated at it and is never renamed.

2. State at v12.167

Sections, by address count:

Surfaces, by capture count:

Authentication state: signed in 223; signed out and incognito (both spellings) 251; incognito 124; signed out 74; unrecorded 91; other recorded states 48.

Evidence: 792 captures carry a verbatim transcript and 328 carry a frame. 505 are scored for PER, at a mean of 0.506; at v10.9 the mean was 0.603.

3. The data flow

Since 2026-08-17 (ratified, MANUS), data/EA-WG-CAPTURES-01.json is the single writable authority. The rebuild store (rebuild/capture-registry/EA-WG-CAPTURES-01-REBUILD.json) is sealed as the rebuild-era source, last written 2026-08-18. This corrects v10.9's account of the data flow, which named the rebuild as canonical; that was accurate at its date.

The seat runs in this order:

1. scripts/capture_intake.py --seat

2. scripts/build_capture_gallery.py

3. scripts/seat_capture_postflight.py, which links, bakes, gates (the runtime gate executes the page scripts), builds the semantic-address layer and publishes addresses/{slug}/, then syncs and cites.

Each batch of intake is kept as rebuild/capture-registry/intake-YYYYMMDD-*/.

4. Files

The core data bundle is EA-WG-CAPTURES-01-v12.167.tar.gz (18,662,466 bytes; SHA-256 ce2c510cc2bd91926cebd71c094c5a13f1f14198a04ea44609477b2e40305548). It holds:

• data/EA-WG-CAPTURES-01.json, the canonical store;
• rebuild/capture-registry/: the pipeline documents, the intake directories and the sealed rebuild store;
• data/capture-deposit-links.json;
• data/semantic-addresses.json;
• datasets/capture-registry/;
• captures/index.html.

manifest.json lists 478 files with SHA-256, bytes, media type and role: the eight core files, the bundle, and the 469 capture frames (179.2 MB), which are manifested at their canonical paths under data/captures/{slug}/ rather than duplicated. This follows the registry's rule: images are held there, and galleries display them rather than host them. Total manifested: 269.4 MB.

5. Version chain

v6.0 (#825) → v6.1 (#1397) → v7.0 (#1398) → v7.2 (#1396, #1401) → v8.3 (#3) → v8.9 (#936) → v10.9 (#1461) → v12.167 (this deposit)

Between v10.9 and v12.167 the registry went from 328 addresses and 470 captures to 633 and 811. The intervening minor versions are recorded in the canonical file, not as separate deposits.

6. What changed in the instrument since v10.9

1. Canonical authority moved from the rebuild store to the projection (2026-08-17).

2. Address pages. Every address has a page of its own (2026-09-25).

3. Surfaces widened. At v10.9, ChatGPT accounted for 3 captures; it now accounts for 167. The registry also holds Google AI Mode, Grok, Bing Copilot, Perplexity, Claude and Qwen.

4. Source chips are data. Each chip is recorded with its relation to the archive and carries into the PER notes. What Enters Composition Through the Cards (#1612) profiles source admission across June–September.

5. The transcript is the capture. #1614 audited transcript coverage and installed the gate that keeps the count monotone.

6. The record-page run (2026-10-12). 43 deposits, from #1421 to #1675, were captured at their own record pages' Ask ChatGPT prompts, signed out and incognito. The author was retained in all 43, the archive in 27, the identifier in 22. Mean PER was 0.22, with one score excluded (see §7).

7. Known defects at v12.167, recorded as data

Defect counts, as the registry declares them:

Declared but not held. Two addresses declare more observations than they hold, so the registry counts 813 observations and holds 811:

• operative-semiotics-20260610 declares 4 and holds 3.
• operational-semiotics-bing-copilot-20260823 declares 2 and holds 1.

PER against its markers. Twelve captures carry all four PER markers true with PER 1.0:

• Ten were seated on 2026-08-13. The marker convention of that seating is not yet verified.
• Two were seated this month, under the convention where true means retained, and are inconsistent with it: dedication-what-syllogizing-can-divide-chatgpt-20261011 and missing-translator-ask-chatgpt-20261012.

Surface not recorded. Four captures carry no surface label at the capture. They are counted here as unrecorded and not assigned.

None of these is repaired in this snapshot. Repair is its own mode.