AI use disclosure

1 · The headline boundary

2 · Models used

3 · Coverage Games disclosure

The quarterly Coverage Games protocol measures inter-rater agreement on a stratified sample of coverage cells. The Q2 2026 event used Claude Sonnet 4.5 as the "independent second classifier" alongside one human classifier (the founder). This is disclosed honestly rather than buried: a single human + an LLM second classifier is materially weaker than the protocol's target of 3–5 independent expert classifiers, and we will not run another Coverage Games event without at least 2 named human classifiers (the editorial-board recruitment dependency in /wiki/editorial-board gates this).

The Q2 disagreement matrix and recusal decisions are documented in docs/coverage-games-2026-Q2.md (public repo). Read alongside this page for the full audit trail.

4 · The AGI Social Scientist research engine

The AGI Social Scientist (AGI SS) research engine is the AI-native research-production layer of Policy Window. It uses Claude for paradigm-classification extractions (Narrative Policy Framework, Multiple Streams Framework, Institutional Analysis & Development, Punctuated Equilibrium, Diffusion), evidence mapping, claim extraction, regulatory-gap identification, regulatory-transfer assessment, and briefing-draft composition. The link from /dashboard/services previously cycled through "Open AGI Social Scientist workspace" (early framing) → "Open Research Workspace" (iter-311 humility downgrade) → and is now restored to the AGI SS framing as of iter-313 because the defensive scaffolding around it (this section, §4.1-§4.4 below; the validation layer at §4.5; the charter §7.4-§7.7 disclaimers) makes the elevation defensible.

4.1 · What the AGI SS does

Evidence mapping; claim extraction; regulatory-gap identification; policy-transfer assessment; semantic-diff generation; comparative analysis; structured-case timeline construction; codebook drafting; uncertainty reporting. It works on primary sources and catalog rows; it does not invent.

4.2 · What AGI SS proposes vs what humans approve

Every AGI SS output is a proposal, never a publication. Approval is a human decision in every case: editorial-board approval for catalog content (additions, edits to existing rows, methodology changes); on-demand expert validation-bench approval for research-module outputs (evidence packs, case studies, sprint deliverables). Both approval paths are recorded as an ApprovalDecision row with a named human reviewer. No daemon, scheduler, or webhook ships AGI SS content to /wiki or to paid-research-module customers without a human in the loop.

4.3 · When human researchers are used

The on-demand expert validation bench (see /wiki/services) handles validity-critical tasks AGI SS cannot reliably perform on its own:

• Expert and stakeholder interviews
• Primary data collection
• Qualitative validation of AGI SS claim extractions
• Source-context checks (was the regulator quoted in context?)
• Methodological review (does the AGI SS evidence map use a defensible inference chain?)

Interview-based work follows consent-by-default, declared retention, anonymisation, and a written protocol (see charter §7.7). No covert observational research; no human-subject work without an approved protocol.

4.4 · What is never automated

• Final editorial approval (always human)
• Legal-advice statements (never produced; see charter §7.4)
• Lobbying outputs (never produced; charter §7.5)
• Court-citation-grade claims (never asserted; the catalog is a structured-lookup layer, not a legal authority)
• Auto-publication of catalog data, coverage classifications, and any surface other than the three disclosed exceptions below (human-gated unless excepted). EXCEPTION 1 — the §7.10 prose tiers (adopted 2026-06-13): abstract / explainer prose tiers on wiki article templates MAY be AI-authored, AI-reviewed, and published without human review — machine-labelled via the data-prose-provenance attribute (the visible per-tier label was retired 2026-07-06 under the §7.9 disclosure-presentation revision, Option A), citation-gated, kill-switch-reversible, disclosed at charter §7.10 + on this page. EXCEPTION 2 — §7.11 AI-curated instruments (adopted 2026-06-15): a small, conspicuously-labelled set of instruments MAY be added and their coverage classified by AI without prior human review, at reduced confidence, citation-gated, kill-switch-reversible, disclosed at charter §7.11. EXCEPTION 3 — §7.12 AGI Social Scientist as primary catalog author (adopted 2026-06-16): AUTHORIZES catalog content (instruments, topics, concepts, benchmarks, coverage classifications, prose) to be AI-authored as the PRIMARY path without prior human review — at reduced confidence, citation- and provision-supports-gated, conspicuously labelled "AI-authored (§7.12)", freeze-to-snapshot kill-switch, named-editor override; the human authoring paths retire only after the engine is wired and proven (delete-last). FIRST EXERCISED 2026-06-16 — CA-SB-243 (Companion Chatbots) is the first §7.12 AI-authored published instrument; not yet at scale. Surfaces other than wiki prose and catalog content still require a human in the loop.

4.5 · Defensive scaffolding

The technical guardrails: verifyDraftIsGrounded() is called defence-in-depth from both src/lib/services/approval.ts and src/lib/services/publication.ts before any draft can leave the workspace. Workspace outputs cannot become catalog rows without an editor explicitly adding them to the typed catalog. Substrate-side events flow through /api/webhooks/agi-ss with HMAC verification + ±5-minute replay-window enforcement. Every AGI-SS-touched surface carries the "AI-assisted" badge.

5 · What we do not do with AI

• We do not auto-publish anything. Every publication transition is recorded as an ApprovalDecision with a named human reviewer. No daemon, scheduler, or webhook ships content to /wiki or /partners without a human in the loop.
• We do not synthesise survey respondents. "Synthetic respondents" are not used to fill evidence gaps; cells without primary sources are marked silent rather than hallucinated.
• We do not generate persuasion content. Briefing composition is for partner evidence packets; campaign messaging, political microtargeting, and covert-influence content are bright-line out-of-scope (see /wiki/charter §7).
• We do not issue reproducibility verdicts using AI. The forthcoming /wiki/reproducibility-policy forbids public "failed-reproducibility" verdicts without human replication review.
• We do not scrape facial images or harvest personal data. The catalog is about instruments, topics, benchmarks, concepts, debates, organisations— not individuals.

6 · The structural enforcement

Under charter §7.9 AI-assisted article prose is permitted — but the §7.9 guardrails (named-editor review, provenance disclosure, unchanged citation integrity, no auto-publication) are not just a policy promise. They are backed by structural properties of the codebase:

• Article templates (src/lib/wiki/templates/{instrument,topic,benchmark,concept}.tsx) render only fields from typed catalog rows. AI-assisted prose enters only through a named editor committing it to those typed catalog fields — there is no model-output sink that bypasses the editor.
• The catalog files (src/lib/international-governance/instruments.ts, src/lib/wiki/concepts.ts, etc.) are version-controlled TypeScript modules committed by humans. Every change — including the drafting-provenance label on any AI-assisted row — shows up in git log.
• With two disclosed exceptions, there is no code path that auto-publishes article prose, and under §7.9 every catalog change is a human commit under named-editor review. The first exception is charter §7.10 (adopted 2026-06-13), under which the abstract / explainer prose tiers on article templates may be AI-authored, AI-reviewed, and published without human review — gated in code by a citation-faithfulness check (reviewCleared()), a machine-readable data-prose-provenance="autonomous-7.10" signal (the visible “no human review” render label was retired 2026-07-06 under the §7.9 disclosure-presentation revision, Option A), and a kill-switch. §7.10 applies to those prose tiers only. The second exception is charter §7.11 (adopted 2026-06-15): a small, conspicuously-labelled set of AI-curated instruments may be added — and their coverage classified — by AI withoutprior human review, at reduced confidence and kill-switch-reversible. Charter §7.12 (2026-06-16) further makes the AGI Social Scientist the primary catalog author (AI-authored at reduced confidence, human-overridable) — FIRST EXERCISED 2026-06-16 with CA-SB-243 (Companion Chatbots), the first §7.12 AI-authored published row. The rest of the catalog remains human-authored pending the engine's expansion, and surfaces other than wiki prose and catalog content remain human-gated.
• The grounding gate verifyDraftIsGrounded() runs at both approval.ts and publication.ts (defense-in-depth). A briefing draft that cites an unverified claim cannot be approved or published.