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 internal research workspace

Policy Window's signed-in dashboard surfaces a research workspace that uses Claude for paradigm-classification extractions (Narrative Policy Framework, Multiple Streams Framework, Institutional Analysis & Development, Punctuated Equilibrium, Diffusion) and briefing-draft composition. The link from /dashboard/services was previously labelled "Open AGI Social Scientist workspace" (a separate adjacent project). As of iter-311 it is labelled "Open Research Workspace" to avoid the "AGI publishes papers" weak-version frame.

Outputs from the research workspace are partner deliverables, not public-wiki articles. They go through mandatory human approval + lobbying-risk screening before any external delivery (see src/lib/services/approval.ts and src/lib/services/publication.ts which both call verifyDraftIsGrounded() defense-in-depth). Workspace outputs cannot become catalog rows without an editor explicitly adding them to the typed catalog.

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

The "no LLM article prose" commitment is not a policy promise — it is a structural property of the codebase:

• Article templates (src/lib/wiki/templates/{instrument,topic,benchmark,concept}.tsx) render only fields from typed catalog rows. There is no "ai-generated" field, no model-output sink, no LLM-output template branch.
• The catalog files (src/lib/international-governance/instruments.ts, src/lib/wiki/concepts.ts, etc.) are version-controlled TypeScript modules edited by humans. Every change shows up in git log.
• Iter-311 removed the /wiki/generate stub and the NEXT_PUBLIC_WIKI_AI_GEN feature flag entirely — structural temptation closed. There is no code path that would let a future contributor flip a flag under deadline pressure and start generating article prose.
• 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.