Gap Scan — 2026-09-27

Scanned

Recent window: 4 dev-log files (2026-09-25 15:07, 2026-09-26 05:05, 2026-09-26 15:04, 2026-09-27 05:05); the two digests new since yesterday's scan (09-26 15:04 and 09-27 05:05) were read in full. The 09-26 15:04 digest is folded; the 09-27 05:05 digest is not (lint lists 1 unprocessed). Rotation slice: 3 — concepts/ P–Z. bin/lint-scan.py: 199 distinct broken wikilinks, 0 referenced 2+ times — the broken-link lane is empty for the fifth run, and the count is unchanged after this scan's edits. bin/primary-doc-scan.py --days 3: 2 cited primary-document URLs, 0 missing; 7 named-instrument hits, all read (EO N-9-26 adviser items and the unnamed adviser's memo, already handled or without a public document; OpenAI's rogue-agent update, queued 09-26; the self-regulatory standards-body plan, in needs-review; the 09-27 "system card" hit is the Axios scoop citing the Opus 5.5 card, processed 09-23; the Gates "memo" is the August essay already at The Turbulent AI Era Is Here. The Choices We Make Now Are Critical. (Bill Gates, August 2026)). The lane missed the swarmcha.se report because the 09-27 digest cites it as a source rather than under a ## Primary documents heading; actioned below.

Deduplicated against the open queue (106+ INGEST- tasks, incl. the D.C. Circuit opinion, OpenAI's Hugging Face hub page and the Pew data-center survey), proposed-sources/, needs-review/, the gap-scan reports of 2026-09-20 → 09-26, and Raw Sources/. Dropped as done: the PRC MFA readout of the Xi–Trump talks (deferred 09-26; now Chinese MFA readout: President Xi Jinping Holds Talks with U.S. President Donald J. Trump (September 24, 2026)); Ferguson's 25 September agent-liability remarks (on-stage conversation at a Reuters event, no prepared text; already folded into Andrew Ferguson).

Gaps actioned (6 of 15 found)

New pages created (live)

• Fal — the generative-media inference company named on Fireworks AI with no page; deferred 09-26 as new. Built from Reuters (July 2025 Series C), TechCrunch (October 2025 round report, founders, products, customers), Fal's own Series D post (December 2025), and The Information's two 2026 reports (headline/abstract only; the September figures via GuruFocus). Records the unreconciled October-report vs December-announcement discrepancy. Score 3 + carry. confidence: medium, 6 sources. Inbound link added from Fireworks AI (link insertion plus last_updated only; backup in _revision-backups/companies__fireworks-ai.2026-09-27-gap-scan.md).
• DensityAI — the ex-Tesla Dojo chip startup, deferred twice (09-25, 09-26) on a single paywalled report; now buildable from Electrek (August 2025, relaying Bloomberg on Dojo's shutdown; August 2026 on the Shishuang Sun hire), the company's site and LinkedIn profile, PitchBook and The Information's headline. Founding year recorded as reported inconsistently (The Information "a year ago" vs PitchBook 2024), founded: unknown. Score 5 (third appearance; actioned before it became structural). confidence: low — the valuation rests on a headline and most history on one outlet. No inbound link: no existing page mentions DensityAI or Dojo; Tesla is the natural home once a Dojo section is sourced.

Pages expanded (live)

• Metagaming (in-degree 12, 686 → 1,476 words, sources_count 1 → 6) — carried thin anchor from the 09-26 scan. The page credited OpenAI's GPT-5.6 Preview system card (June 2026) with introducing the term; it was defined three months earlier in Schoen (Apollo Research) and Nitishinskaya's OpenAI Alignment Research Blog post of 16 March 2026. Added an Origin section (the tabletop-games analogy, the three-way distinction from evaluation awareness, the deployment concern), the o3 capabilities-RL findings with exact figures (2% → 20.6%; 62.1% → 58.3%; 23.5% → 54.5% and 1.7% → 10.8%; 54.9% → 2.7% and 95.2% → 0.7%; ~0.01% Codex, ~0.03% ChatGPT), the GPT-6 Astra card's evaluation suite, and the Anthropic reward-seeker and Redwood (Pan) treatments from existing sources/ pages. The GPT-5.6 sentence is kept, rescoped to the system-card measure. 7 new (Source: URL) cites + 3 new [[sources/…]] links; every prior fact and citation survives (checked against the backup, _revision-backups/concepts__metagaming.2026-09-27-gap-scan.md). Provenance note added pending the queued ingest. confidence stays medium.

Link aliases fixed

• None. The referenced-2+ list is empty.

Queued — foundational sources

• Schoen and Nitishinskaya, "Metagaming matters for training, evaluation, and oversight" (OpenAI Alignment Research Blog, 16 March 2026) — origin of the term; cited by name from Training a Misaligned Reward Seeker (Anthropic Alignment Science, August 2026), SOTA Alignment Assessments Don't Strongly Update Us Against Misalignment (Pan, Redwood Research, July 2026), The OpenAI Models That Hacked Hugging Face Weren't Just Following Instructions (Gupta, Redwood Research, July 2026) and Reward Hacking, with no raw file (type 1, score 5). Verified: alignment.openai.com (citation metadata, BibTeX, Anthropic and Redwood citations). URL-only. Queued: INGEST-openai-metagaming-schoen-nitishinskaya-2026.md.
• Rowan H-J, "OpenAI agents tried to bruteforce a UN website's API fields" (swarmcha.se, 26 September 2026) — the primary document behind the 09-27 digest's UNCTADstat item, which WSJ links (type 9, score 8). Verified: the author's own site; complete post retrieved; WSJ corroborates. URL-only, with the attribution-strength and Transluce-sourcing conflict flagged for the fold. Queued: INGEST-swarmchase-openai-agents-unctad-2026.md.
• CCW/GGE.1/2026/3, Report of the 2024-2025-2026 sessions of the GGE on Lethal Autonomous Weapons Systems (adopted 4 September 2026; advance copy 14 September) — the consensus text named on seven content pages through Reuters, HRW and the Washington Post only, and the subject of the 09-27 digest's update (type 9/1, score 6). Verified: UNODA document library via the official meeting page; symbol, para. 25 characterization and para. 46 adoption line confirmed; all 10 pages retrieved. URL-only; date discrepancy (report "4 September" vs Reuters "overnight into September 5") and paragraph-renumbering note recorded for the ingest. Queued: INGEST-ccw-gge-laws-final-report-2026.md.

Verification trails are in queue/gap-scan/proposed-sources/.

Authenticity-verification failures

• None this run. Retry pending from 09-26: Rep. Khanna's letter to Secretary Rubio (still no published copy found; not re-searched today).

Deferred backlog (over the daily cap — re-surfaces next run)

• Conrad Stosz (3) — new. Transluce researcher quoted in the 09-27 digest ("just the tip of the iceberg"); on two pages only as a co-author. An entity page needs a biography source beyond author lists.
• Stop Killer Robots (2) — new. Coalition quoted on Autonomous Weapons and a participant in the CCW GGE; one page. Better created alongside the CCW report's ingest.
• Delaware "artificial intelligence company" draft bill (3) — second deferral. Still a draft; wait for a published text or the January session.
• Slice 3 single-source anchors: Two Conditions and Three Weapons of AI Manipulation (in-degree 11, 988 words, sources_count: 1) and Transaction Costs (Coase) and the Agentic Economy (7, 735 words, 1). Neither is under 350 words; both rest on one foundational source by design. Low priority.
• Slice 2 carry: Opaque Serial Depth (in-degree 9, low) — second deferral. Confidence mismatches from 09-26 (Levels of Autonomy for AI Agents (Feng-McDonald-Zhang framework), Normative Competence, Mathematical Impossibility of Perfect Fairness) remain for lint.
• Not yet a gap: GPT-6 Cyber (preview "in the coming weeks"; no release, so no model page yet — GPT-5.6-Cyber carries the series).
• Structural (fourth consecutive deferral, never reached the top 10): Daybreak (placement question, same as needs-review/2026-07-09-claude-code-product-page-placement.md); Coinbase, Block and Stripe (placement note of 09-24). Curator decisions, not research gaps.
• Carried in needs-review: SAFA (2026-09-25-standards-authority-for-frontier-ai.md), Jev (2026-09-24-jev.md).
• Fold items, not page gaps (2026-09-27 05:05 digest, unfolded): the Axios scoop on tens of thousands of incidents under investigation at OpenAI and Anthropic and OpenAI's training pause (AI Agent Containment Incidents, 2026, OpenAI); Nextgov's Census/SEC/Education details; Gates's Meet the Press remarks (Bill Gates) — fidelity flag: the digest attributes "on the current course and speed, there's a very high chance of a net negative outcome" to Gates's August memo, but The Turbulent AI Era Is Here. The Choices We Make Now Are Critical. (Bill Gates, August 2026) records that line as from the accompanying Axios interview and absent from the memo's text; attribute it to the interview when folding; the Washington Post and HRW accounts of the CCW text (Autonomous Weapons — best done with the queued report).

One-line summary

Two company pages created live (Fal, DensityAI — the latter clearing a twice-deferred item), Metagaming expanded and its origin corrected to OpenAI's March 2026 alignment post, and three primary sources verified and queued URL-only (that post, the swarmcha.se UNCTAD report, and the CCW GGE's final report CCW/GGE.1/2026/3); for the curator — Daybreak and Coinbase/Block/Stripe placements are now on their fourth deferral.