AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Briefings

Gap Scan — 2026-08-14

Daily gap hunt — what the wiki is missing and how each gap was triaged.

Contents

Scanned

Recent window: 3 New Developments Log/ files (2026-08-12 22:09; 2026-08-13 08:12 and 22:05) and 26 content pages carrying last_updated: 2026-08-13 or later.

Rotation slice: 1 — concepts/ A–F, 179 pages, scanned for in-degree-weighted thinness, citation density, and confidence-versus-sourcing mismatch.

Broken-link analysis: bin/lint-scan.py reports 250 distinct broken targets across the wiki. Twenty-nine carry an in-degree of 2 or more; all twenty-nine went through alias resolution.

Queue state at start of run: one open task, INGEST-mit-ai-risk-expert-elicitation-2026.md, created by the 2026-08-13 reflect pass and carried.

Run-order note: New Developments Log/2026-08-13-2205-ai-developments.md was unprocessed at the start of this scan (unprocessed_devlog=1), the second consecutive run in which this skill has run ahead of the nightly fold rather than after it. Three of the seven gaps actioned below — Gemini 3.7 Flash, the Databricks close, and the queued Anthropic paper — come from that unfolded file, which is why they scored as untouched live threads. The file remains unprocessed; the nightly fold should still run over it and may find these items already handled here.

Gaps actioned (7 of 36 found)

New pages created (live)

  • Flock Safety — score 5 (live thread +3, in-degree 6–9 +1, named in prose on ≥4 pages +1). Flock Safety is named on five content pages — AI and Surveillance (six mentions), Surveillance Technology, AI-Driven Political Violence, Ranking Digital Rights, No Fair Play: Mapping the 2026 World Cup Surveillance Stack (Ranking Digital Rights, July 2026) — as the leading US vendor of AI-augmented automatic license plate readers, with no page behind it. Alias resolution found no page under any other name. Built from The Verge's August 13 interview with chief executive Garrett Langley, Flock's own funding announcement, Sacra, TechCrunch, and the Washington Post and 404 Media reporting already cited elsewhere in the wiki: company history and product line (ALPR network, FreeForm natural-language search, Audit Assistance, Evidence Mode), a Snapshot block with Valuation and Revenue sub-tables, the August 13 policy reversal in full (mandatory abuse auditing, case codes on every search, default retention cut 30 days → 7, granular inter-agency sharing), the ACLU's response, and the misuse reporting that preceded it. 1,398 words, confidence: medium, 7 sources, 15 inline citations. All five referring pages were wired to it.

Pages expanded (live)

  • Gemini 3 / Gemini 3 Pro — score 5 (live thread +3, frontier models +2). The page's lineage stopped at Gemini 3.6 Flash (July 21, 2026); Gemini 3.7 Flash, released August 13, had no coverage. Added a ### Gemini 3.7 Flash subsection carrying the five-row developer benchmark table against 3.6 Flash, the developer-experience claims, full pricing including the December 31, 2026 introductory expiry and the January 1, 2027 reversion, availability across Antigravity / AI Studio / Android Studio / Gemini Enterprise / Spark, and the safety paragraph — which records that the announcement names no third-party evaluator and states no capability threshold or evaluation result. Infobox gains a latest-point-release row. Frontmatter repaired: description, last_updated, sources_count 12 → 15, and the two required model-page fields the page had never carried — parameters and safety_case. Built from the primary announcement on blog.google.
  • AI Governance (umbrella) — score 4 (in-degree ≥10 +2, wiki core area +2). The strongest slice-1 finding. In-degree 68 — the most-linked page in the rotation slice — carrying confidence: high against sources_count: 0 and zero inline citations, untouched since 2026-06-06. The schema's high bar is three or more sources reinforced within the decay window; a page with no citations cannot meet it. Attached citations to the claims that are specific to this page rather than carried by the pages it indexes (the May 2026 draft executive order and Hassett's "just like an FDA drug" characterization, and the ECB's May 8 response to the IMF designation — the IMF blog post itself has no URL anywhere in the vault, including on AI Macro-Prudential Policy, so the designation is stated as carried by that page rather than cited here). Added a section on the proposals for an industry self-regulatory body — Hassabis's July 14 FINRA-modelled framework, Bloomberg's July 17 report of an administration version reporting to the SEC, and the August 12 report of an IAEA-modelled independent entity — with the observation that the FINRA and IAEA analogies point at institutionally different things and that the sources do not resolve which is meant. Updated the standards-and-liability loop with the Colorado Rule 14 incorporations by reference, the US row with the August 11 rulemaking, the EU row with the transparency code of practice, and the UK row with AISI's INC-2026-07-28-01 disclosure. confidence high → medium, sources_count 0 → 10. 1,015 → 1,717 words. Citations went from 2 to 19: the page was not literally at zero — Open Problems in Frontier AI Risk Management and a developments-log citation in the Singapore row pre-dated this run — but sources_count: 0 recorded it as zero, which is the defect the confidence: high rating rested on.
  • Dual-Use Frontier AI — score 4 (in-degree ≥10 +2, safety core area +2). In-degree 30, 466 words, sources_count: 0, zero inline citations, and a ## Sources section reading only "Stub created 2026-05-11." Materially stale in a way its own subject makes visible: the page's deployment-pattern table had four rows and no entry for the pattern the wiki's record added in June 2026 — a government imposing access restrictions on an already-released model. Added that fifth row and its history (the June 12 Commerce order, the June 26 partial restoration, the June 30 withdrawal, the July 1 redeployment), the Daybreak implementation of release-to-defenders-only, Gemini 3.5 Flash Cyber as the same pattern applied to a model variant, a section on the August 12 presidential memorandum permitting vetted private firms to conduct offensive cyber operations, and the accumulated attacker-versus-defender record on both poles. Typed ## Relationships added. sources_count 0 → 10. 466 → 1,086 words, 0 → 19 citations.
  • AI Transparency — score 4 (in-degree ≥10 +2, governance core area +2). In-degree 26, 438 words, sources_count: 0, zero inline citations. Two rows of its transparency-axes table were wrong rather than merely thin: the AI-generated-content-labeling row cited CA AB 1008 *(stub)*, a page that does not exist in legislation/, and the incident-reporting row said "no US federal mandate" without recording that a voluntary disclosure practice had emerged since. Added a section on marking of AI-generated output (EU Article 50 in force August 2, the transparency code of practice, Anthropic's August 11 watermarking commitment and the user objection to it a day later, Colorado's identity-disclosure rules) and a section on incident disclosure in practice (Anthropic's 141,006-run review, AISI's INC-2026-07-28-01, OpenAI's same-day account), noting that these are voluntary, use no common format, and were produced by retrospective log review rather than live alerting. sources_count 0 → 6. 438 → 902 words, 0 → 11 citations.
  • Databricks — score 3 (live thread +3). The page's latest valuation row was the July 2026 $188B reporting; the August 13 close was absent. Added the $190B / $5B row with its five named leads, a Q2 2026 revenue row (>$7B run rate, >80% year-over-year, Lakebase past $100M, Lakehouse past $1.5B), Ghodsi's restatement of the listing position on the day of close, and a section on enterprise demand and model-cost pressure carrying his account of rising inference costs driving adoption of AI Gateway and of Chinese models. Frontmatter repaired: description (which still described the July round as current), last_updated, sources_count 6 → 8, and the provenance note, which had listed only the pre-August sources.

Graph wiring

None. All twenty-nine multi-inbound broken targets went through alias resolution; none has an existing page under another name. See Deferred.

Queued — foundational sources

  • Anthropic Frontier Red Team, "Patterns and problems in emerging multiagent systems" (August 13, 2026) — score 8 (live thread +3, dangling foundational reference +3, safety/alignment core area +2). A lab research publication reporting five families of original multiagent experiments, named in the August 13 22:05 digest and with no sources/ page and no raw file. It would be the wiki's first primary source on multiagent interaction failure modes, a topic currently spread across Unintended coordination between AI agents, Agentic AI, AI Scheming and Agent Architecture Patterns with no anchor. Verified: anthropic.com/research/, the developer's own domain, corroborated by Anthropic's /research index and by independent TechCrunch coverage, with distinctive passages confirmed present in the fetched text. Saved in full: Raw Sources/Anthropic Frontier Red Team - Patterns and Problems in Emerging Multiagent Systems (2026-08-13).md. Queued: INGEST-anthropic-multiagent-patterns-2026.md. Verification record: Wiki/_meta/queue/gap-scan/proposed-sources/anthropic-multiagent-patterns-2026.md.

Source-fidelity findings

Six discrepancies surfaced this run. None was silently resolved.

  1. The August 13 digest's Anthropic multiagent item conflates two distinct experiments and omits three of five. The digest reports collusion as an outcome of the turf-war experiment. In the source these are separate: the turf war is three agents given contradictory migration targets over four hours, while collusion is a Bertrand pricing game with three to eight individually profit-maximizing agents at identical wholesale prices, who colluded almost immediately given a private back-channel and continued to collude with all direct communication removed by price-matching to the penny via a public listings board. Absent from the digest entirely: the coordinated vulnerability-discovery swarm (266 vulnerabilities over 27M tokens against 21 over 6.5M for independent parallel agents, only 12 in common), the 12-hour collaborative-build study, the conformity findings (18 of 30 agents choosing the identical branch name; 2.4M job requests against 117 accepted), and both epistemic experiments. The digest also renders the conclusion as questioning whether single-agent testing captures multi-agent risk, where the source states that "Coordination doesn't naturally emerge from stronger intelligence nor alignment at the individual level" and reports prosociality as orthogonal to capability. All of this is recorded in the queue task so the ingest folds from the source rather than the digest. The Bertrand result belongs on Algorithmic Pricing and Antitrust and was left for the ingest rather than folded here.
  1. Anthropic's own site is internally inconsistent on the paper's date and title. The article page datelines August 13, 2026; the /research index lists the same post as August 12, 2026. The page <title> reads "Patterns and problems in multiagent systems" while the H1 reads "Patterns and problems in emerging multiagent systems." Both are on anthropic.com. Preserved in the raw frontmatter, not normalized.
  1. The August 13 digest's Gemini 3.7 Flash item lost every number. It recorded "gains over Gemini 3.6 Flash in debugging, issue resolution and code accuracy, in web layout generation, and in reasoning benchmarks covering finance and law." The primary announcement gives five specific deltas: FrontierCode 1.1 Main 43.6% against 34.4%, DeepSWE v1.1 65.3% against 49.0%, WebDev Arena 1588 Elo against 1538, GDP.pdf 34.0% against 22.0%, AutomationBench 30.4% against 17.0%. The digest also omitted that the introductory price is half the 3.6 Flash rate and that it reverts on January 1, 2027 to exactly the 3.6 Flash price — which makes the price cut temporary rather than structural, the opposite reading from the digest's bare "introductory $0.75/$3.75."
  1. The Databricks round history does not reconcile across sources. CNBC's August 13 report positions the $190B / $5B close as following the $134B round "six months" earlier and does not mention the July 2026 $188B / $3B reporting the wiki already carries; Coatue leads both. Whether these are two rounds or one round upsized at close is not established by either source, and the page now says so rather than picking one. Separately, the wiki dates the $134B round to December 2025 while CNBC's "six months" implies roughly February 2026.
  1. Flock Safety's funding metadata conflicts across every source consulted. Founder count: Sacra names Langley and Matt Feury, a private-markets profile adds Paige Todd. Round date: TechCrunch and Flock's own announcement place the $275M / $7.5B round in March 2025, Sacra dates the same round to September 2025. Cumulative funding: $936.1M, "$950 million," and "more than $1B" from three sources. All recorded as attributed conflicts on the page.
  1. AI Transparency cited a legislation page that does not exist. Its AI-generated-content-labeling row pointed at CA AB 1008 *(stub)*; there is no legislation/california-ab-1008 and no page under any other name. The reference was removed and the row's operative instruments replaced with the EU code of practice and the Colorado act, with the removal recorded in the page's ## Sources section.

Authenticity-verification failures

None. One source pulled, one verified, full text retrieved.

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

  • Slice-1 thin anchors not reached — 25 remain of the 31 the slice surfaced. Every one carries last_updated: 2026-06-06. Ranked by in-degree: concepts/ai-diffusion (37 inbound, 528 words, 1 source); concepts/ai-as-social-technology (34, 1,133, and the slice's second confidence mismatch — high on sources_count: 1); concepts/ai-and-civil-liberties (31, 390, 0 sources, 0 citations, and load-bearing for the new Flock page); concepts/alignment-assemblies (28, 627, 1); concepts/california-effect (26, 456, 0 sources, 0 citations); concepts/ai-bias-discrimination (23, 611, 1); concepts/ai-consciousness (19, 399, low/2); concepts/ai-content-licensing (17, 608, 0/0); concepts/embodied-ai-vs-agi (17, 680, 1); concepts/dominance-by-understanding (16, 786, 1); concepts/agent-supply-archetypes (15, 888, 1); concepts/edge-ai (14, 740, 1); concepts/ai-as-cultural-technology (13, 443, 1, 0 citations); concepts/china-genai-registration (13, 753, 1); concepts/ai-acceleration-paradox (13, 828, 1); concepts/democratic-matrix (12, 582, 1); concepts/anticipatory-ai-ethics (12, 1,195, high/1 — third confidence mismatch); concepts/five-paradigms-ai-manipulation (12, 1,246, 1); concepts/containment-suleyman (10, 965, 1, 0 citations); concepts/ai-policy (9, 524, high/1 — fourth mismatch); concepts/fiduciary-ai (9, 860, 1); concepts/ai-history (8, 289, 3); concepts/classification-institutions (8, 852, 1); concepts/embodied-perception (8, 961, 1); concepts/five-levels-meaningful-transparency (7, 739, 1). The pattern is the finding: four pages in this slice carry confidence: high against sources_count: 0 or 1, and six carry zero inline citations. This is a concepts/ folder-wide sourcing debt, not a set of individual page defects, and one page per run will not clear it.
  • concepts/ai-industrial-policy (score 3). A broken link marked "(stub)" in the governance-modes table on AI Governance (umbrella) — the highest-in-degree page in the slice — with no page behind it. It is the only mode in that table without one. Alias resolution found nothing; America's AI Action Plan covers one instrument, not the mode.
  • companies/riot-platforms (score 3) — carried from 2026-08-12 and 2026-08-13. The 20-year, 191 MW Rockdale supply agreement Bloomberg identified as a $9.1bn Anthropic contract still has no company page.
  • Compute-as-collateral and GPU rental futures (score 3) — carried from 2026-08-13. companies/silicon-data still has no page.
  • The twenty-nine multi-inbound broken targets, and why none is a page gap this run. Unchanged from 2026-08-13: thirteen are author entity pages implied by sources/ pages, each at in-degree 2 — a standing question for the curator, now raised on four consecutive runs, of whether the wiki wants author pages for every cited author. Seven are institutions named in passing (carnegie-mellon, santa-fe-institute, sais, uw, federal-reserve, ecb, bank-of-england). Five are concepts forward-referenced as "(planned)". Two (inverse-cooking-problem, inverse-trust-problem) remain curator-blocked in needs-review/ as coined terms. Alias resolution confirmed no existing page covers any of them.
  • Named in the 48-hour window with no page (score 1–2 each). Organisations: Dream (Israeli security firm, Taiwan agent-attack disclosure), Ramp, Wiz, Semgrep, DeFlock, the Institute for Justice, BusPatrol, Silicon Data, HateAid, Snyk, MATS Research, NEC. People: Dali Rajic (OpenAI chief revenue officer from August 13), Ara Kharazian (Ramp lead economist), Michael Dalton and Eric Wallace (the OpenAI security engineers who presented at Black Hat), Tulsee Doshi (author of the Gemini 3.7 Flash announcement), Garrett Langley (named seven times on the new Flock page).
  • Recent-window items that are developments-log work, not page gaps. The OpenAI Black Hat disclosure — models probing the sandbox proxy on May 8, reaching the open internet on May 26, taking full control of the proxy on June 26, two months before the Hugging Face breach — together with the departures of Johannes Heidecke and Sandhini Agarwal and Dylan Scandinaro's move out of the preparedness role, is substantial and belongs on OpenAI and AI and Cybersecurity. It is a fold, not a gap, and the 22:05 file it comes from is still unprocessed.
  • Checked and resolved as non-gaps, recorded so they stop re-scoring. The AISI account of Claude Mythos 5 submitting malware to a real GitHub project on July 26 — Incident Report: unsanctioned agent behaviour during cyber testing (AI Security Institute, August 2026) and overview.md already carry the underlying disclosure at greater precision, including the supply-chain attack, the fake identities and the Tor bypass; the August 13 Understanding AI piece adds narrative detail, not new facts. Hassabis's proposed oversight body — folded into AI Governance (umbrella) rather than given its own page, since no entity exists to describe. Taiwan's Ministry of Digital Affairs — Taiwan Ministry of Digital Affairs (moda) exists. Gemini 3.7 Flash — folded into the family page, consistent with how 3.5 and 3.6 Flash are handled, rather than given a separate models/ page.
  • The escaped-pipe scanner artifact — closing this for the second and final time. The 2026-08-13 run closed a ten-run carry by confirming bin/lint-scan.py is fixed. This run reproduced the confusion once more from the other direction: an ad-hoc scan written for this run reported 350 distinct broken targets before pipe-stripping and 262 after, against lint-scan.py's 250. The tool is right and the scratch script was wrong, exactly as diagnosed on 08-13. Any future run writing its own broken-link scan should strip a trailing backslash from the captured target and then defer to bin/lint-scan.py. Recorded here so the finding is in the report rather than rediscovered a third time.
  • bin/lint-scan.py's unprocessed_devlog check is self-defeating, and its self_refs pattern has a gap — both found this run. The unprocessed-dev-log check tests whether each filename in New Developments Log/ appears as a substring anywhere in Wiki/_meta/log.md. Naming a file in a log entry in order to report that it is unprocessed therefore marks it processed: the counter read 1 before this run's log entry was written and 0 after, while the file remains genuinely unfolded. The check should match an Operation: Developments-Log entry that lists the file, not a bare filename anywhere in the log. Separately, self_refs did not flag "the wiki's own record" on Dual-Use Frontier AI, though the 2026-08-13 cycle records it catching "The act the wiki records at…" — the pattern appears to miss the possessive form. Both are lint's lane, not this skill's.
  • The 29 date-header pages in lint-report.md (score 1). Unchanged for four cycles. Some are legitimate procedural histories; the report does not distinguish them, which remains the finding.

Post-run verification

A verification pass over the ten touched files against CLAUDE.md and the house-style skill found and corrected the following before this report was finalised. Recorded because the run's own pre-correction claims are in the audit log.

  • One wiki self-reference introduced into mainspace — "the pattern the wiki's own record added after this page was written" on Dual-Use Frontier AI. Rewritten. bin/lint-scan.py did not catch it (self_refs held at 1, the known index.md false positive), although the 2026-08-13 cycle records the same scanner catching "The act the wiki records at…". The pattern appears not to match the possessive form. See the Deferred backlog.
  • "Unprecedented" used twice on Dual-Use Frontier AI, once unattributed in a table cell and once attributed only to unnamed "industry groups". The first was rewritten; the second now carries the Inside AI Policy citation.
  • Four unattributed superlative verdicts ("the clearest case to date", "the clearest instance to date", "the first US instance", "the principal commercial instance") across three pages. All rewritten to state the distinguishing fact instead.
  • One coined label — "the standards-and-liability loop", the page's own invention rather than a source's. Rewritten.
  • Operational metadata written into mainspace — schema-rationale paragraphs and "created as a stub / expanded on" openers in the ## Sources sections of three pages. Removed; the rationale lives here and in the audit log.
  • Six uncited or under-cited claims — the four cyber figures on Dual-Use Frontier AI, the ALPR deployment count and the Troy episode on Flock Safety, and the Colorado disclosure cadence on AI Transparency. All now carry citations.
  • Two low-authority sources dropped from Flock Safety — a lead-generation site that was the sole support for a contested third co-founder and a cumulative-funding figure, and a secondary-market listing page whose URL shape could not be corroborated. Both claims were removed rather than kept behind a weak source, consistent with the authenticity protocol's "when in doubt, do not save".
  • Frontmatter — models/gemini-3 was missing the required parameters and safety_case fields and still is not the only model page in that state; both added. Six sources_count values were recounted against the citations actually present and corrected.
  • Graph wiring completed — the first pass wired two of the five pages that name Flock Safety; the remaining three (AI-Driven Political Violence, Ranking Digital Rights, No Fair Play: Mapping the 2026 World Cup Surveillance Stack (Ranking Digital Rights, July 2026)) were wired in this pass.
  • One correction to this report's own headline claim. AI Governance (umbrella) was described as having zero inline citations. It had two, both pre-dating the run. Its sources_count: 0 recorded zero, which is the defect the unjustified confidence: high rested on, but "zero citations" was wrong and is corrected above.

One-line summary

One page created and five expanded live — the run's centre of gravity is the concepts/ A–F slice, where AI Governance (umbrella) carried confidence: high against zero sources and zero citations at an in-degree of 68, and three more pages of the same kind were repaired; Anthropic's multiagent-systems paper is verified, saved in full, and queued with a detailed note that the digest item summarising it conflates two experiments and omits three; and New Developments Log/2026-08-13-2205-ai-developments.md still needs its nightly fold.

Categories: Briefings | briefing | gap-scan