# Policy Window — full-context AI-assistant guide

> The machine-readable AI-governance catalog. (The team's longer-form strategic positioning frames this as "AI-native AI-governance research infrastructure" with three pillars — see § Three pillars below; the public surface deliberately leads with the artefact rather than the platform claim, for credibility with the current state of the evidence base.)

This is the full-context AI-assistant ingestion guide for Policy Window. For the curated short-form version, see https://policywindow.org/llms.txt.

## Catalog scope (snapshot)

- 45 governance instruments across jurisdictions: EU, US, UK, CN, JP, SG, IN, BR, ASEAN, African Union, plus international (G7, OECD, UN, Council of Europe) + industry voluntary codes (Anthropic RSP, OpenAI Preparedness, DeepMind FSF, Meta Frontier AI Framework)
- 23 contested topics (foundation models, biometric ID, employment AI, training data, redress, transparency, compute reporting, etc.)
- 10 capability benchmarks
- 32 concepts (frontier-tier, ASL-3, scalable oversight, alignment, etc.)
- 525 coverage cells (instrument × topic, with type / confidence / pinpoint citation)
- Literature & evidence base: academic + grey-literature sources (peer-reviewed, preprints, working papers, research-institute, think-tank, civil-society, standards, incident databases) linked to the contested topics they bear on; AI-drafted ≤1-line findings are labeled. Machine-readable at https://policywindow.org/wiki/literature.json

## Three pillars (strategic positioning)

1. **Primary-source-cited, machine-readable AI-governance catalog** — the free public surface. Every claim cites a primary source; article prose may be AI-assisted under charter §7.9 (named-editor-reviewed, provenance-disclosed), and some abstract/explainer prose tiers are AI-authored + AI-reviewed and published without human review under charter §7.10 (conspicuously labelled, citation-gated, kill-switch-reversible) — catalog data + coverage classifications stay human-gated, EXCEPT conspicuously-labelled AI-authored instruments at reduced confidence (additive under §7.11; AI-authored as the primary path under §7.12 since 2026-06-16, first row CA-SB-243), all kill-switch-reversible. CC BY 4.0 + CC0 citation graph.
2. **AGI Social Scientist research engine** — AI-native research-production layer that produces evidence-mapping + claim extraction + regulatory-gap identification + comparative analysis as PROPOSALS for human review. Substrate-side; details at https://policywindow.org/wiki/ai-disclosure §4.
3. **On-demand expert social-science validation layer** — trained researchers review research-module outputs for validity-critical claims (claim-extraction fidelity, source-context check, methodological review, expert interviews) before publication. Aspirational scaffolding — pages ship; live bench recruits per commission.

## Operating charter — bright lines (canonical)

Policy Window will not do the following — these are bright lines, not aspirations:

- §7 — Lobbying on behalf of any client, paid or unpaid
- §7 — Political microtargeting, persuasion campaigns, or covert-influence operations
- §7 — Issuing public "failed-reproducibility" verdicts on specific papers without human replication review (see https://policywindow.org/wiki/reproducibility-policy)
- §7 — High-stakes individual decision support (criminal sentencing, asylum determination, medical triage)
- §7 — Auto-publication of any output to a public surface without an audited human approval (as of 2026-06-13, §7.10 carves out one scoped exception: the abstract/explainer wiki prose tiers, AI-authored + AI-reviewed and published without human review, conspicuously labelled + kill-switch-reversible — absolute for every other surface)
- §7.9 / §7.10 — Auto-publishing AI-assisted article prose outside the two governed regimes: §7.9 permits it under named-editor review + provenance disclosure + the citation-integrity check; §7.10 (adopted 2026-06-13) permits the abstract/explainer prose tiers to be AI-authored + AI-reviewed and published without human review under a conspicuous label + citation gate + kill-switch. Every claim still cites a primary source; outside those regimes, auto-publishing prose remains a bright line
- §7.4 — Legal advice (outputs are evidence references, not legal counsel; consult qualified counsel for jurisdictional opinions)
- §7.5 — Lobbying automation (no bulk regulatory comments, no synthetic public-consultation submissions, no astroturf campaigns — at any scale, for any client)
- §7.6 — Covert persuasion or political microtargeting (all editorial output is signed + dated)
- §7.7 — Unapproved human-subject research (Layer-5 expert interviews require informed consent, anonymisation, declared retention, and a written protocol)

Full charter: https://policywindow.org/wiki/charter

## AI use disclosure (§4 canonical)

The AGI Social Scientist research engine 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, policy-transfer assessment
- Briefing-draft composition (always human-reviewed before publication)

Every AGI SS output is a PROPOSAL, never a publication. Approval is a human decision in every case:
- Editorial board approves catalog content
- On-demand expert validation bench approves research-module outputs
- Both approval paths recorded as ApprovalDecision rows with named human reviewer

What is NEVER automated (with one disclosed §7.10 exception, below):
- Final editorial approval of catalog data + coverage classifications (always human)
- Legal-advice statements (never produced; §7.4)
- Lobbying outputs (never produced; §7.5)
- Court-citation-grade claims (never asserted)
- Publication of public-wiki article-body catalog content (AI-assisted prose is permitted under charter §7.9 with named-editor review + provenance disclosure, and every claim still cites a primary source)

The one §7.10 exception (adopted 2026-06-13): the abstract/explainer prose tiers on the four wiki article templates MAY be AI-authored, AI-reviewed, and published WITHOUT human review — gated by a code-level citation-faithfulness check, a conspicuous "no human review" label, and a kill-switch. It applies to those prose tiers only; catalog data, coverage classifications, and every other surface remain human-gated.

Full AI disclosure: https://policywindow.org/wiki/ai-disclosure

## Methodology (canonical Three Rs framing)

1. **Reproducible** — render-deterministic by construction. Articles render deterministically from typed catalog rows; no random sampling, no fitted model. Article prose may be AI-assisted under charter §7.9 (named-editor-reviewed, provenance-disclosed, every claim still cites a primary source). (Cell-level research reproducibility is the Replicable axis; see Coverage Games.)
2. **Replicable** — Coverage Games inter-rater protocol re-runs each quarter; multiple human classifiers independently re-classify a stratified sample of coverage cells. Q2 2026 event was 1 human + 1 LLM (process shake-out, not peer-review-grade estimate). Q3 2026 gated on ≥3 named human classifiers.
3. **Robust** — per-cell confidence tier (high / medium / low / pending) disclosed on every coverage row. Citation chain to primary source on every cell.

Full methodology: https://policywindow.org/wiki/methodology

## §11 Limits and blind spots (canonical honest disclosure)

- **Instrument-centric ontology** — refusal politics, labour strikes against algorithmic management, abolitionist demands sit outside the instrument frame and are not catalogued
- **Coverage-depth asymmetry by jurisdiction** — Western instruments (EU AIA, US EO-14110, UK, OECD) carry ~40% more coverage cells than Global South instruments (India DPDP, Brazil PL 2338, AU AI Strategy, ASEAN AI Guide)
- **Frame favours regulator anxiety over harm narratives** — topic selection currently centres what regulators have authority over; welfare-system automation, child-protective-services ML, immigration-enforcement AI, gig-worker algorithmic management are under-catalogued (named at https://policywindow.org/wiki/for-advocates §3)
- **Three Rs framework excludes situated knowledge** — lived-experience testimony, oral histories of algorithmic harm, and refusal narratives do not fit the cell-grid format
- **Editorial board geographic concentration** — 1 of 6 slots filled, by the founding editor with no disclosed Global South AI-policy expertise
- **English-only** — articles render only in English; no translation roadmap commitment yet
- **"Neutral" framing has politics** — charter §7 commits to not authoring advocacy content; this forecloses being a partisan resource for communities in struggle against a particular governance regime

## Audience guidance (who PW serves well + less well)

**Well-served**:
- AI-governance researchers needing cross-jurisdiction structured comparison
- Journalists needing lead-generation + comparative framing + named-expert contacts
- Procurement evaluators needing institutional-readiness checklist + machine-readable export
- Civil-society advocates needing coverage-gap identification + topic-proposal channel
- Graduate-school instructors needing sample syllabi + assignment ideas
- AI agents needing structured, agent-readable regulatory substrate
- Open-knowledge / library systems needing FAIR-shaped metadata

**Less well-served**:
- US state-law preemption analysis (drafted; pending editorial verification)
- Real-time regulatory tracking (PW updates on editorial cadence, not minute-by-minute)
- Industry-deep capability benchmarks (CSET / Stanford HAI win for depth)
- Algorithmic-management / gig-worker / labour-displacement (v2 roadmap)
- Non-English-fluent users (no translations yet)
- Court-citation-grade authority (charter §3 explicit disclaimer)

## How to cite Policy Window (canonical form)

> Policy Window. (YYYY). *Article title*. Retrieved YYYY-MM-DD from https://policywindow.org/wiki/[slug]?asOf=YYYY-MM-DD

The `?asOf=YYYY-MM-DD` parameter pins the URL to a specific historical snapshot stored in ArticleRevision. When the snapshot exists, a green banner confirms; when no snapshot was captured for that date, an amber banner discloses live-fallback.

8 reference-manager export formats per article (APA, Chicago, Harvard, BibTeX, RIS, CSL-JSON, OSCOLA, Bluebook). DataCite XML at `https://policywindow.org/wiki/[slug]/datacite.xml` pre-stages Q4 2026 Zenodo DOI registration.

## Machine-readable data surfaces (canonical)

- **Full catalog JSON**: https://policywindow.org/wiki/catalog/json (CORS-open, 5-min cache, CC0 citation graph; now includes the `literature` evidence array)
- **Literature & evidence-base JSON**: https://policywindow.org/wiki/literature.json (academic + grey-literature evidence linked to topics; CC0; AI-drafted findings flagged `aiGenerated`)
- **Catalog CSV** (one row per cell): https://policywindow.org/wiki/catalog/csv (CSVW metadata: https://policywindow.org/wiki/catalog/csv-schema)
- **JSON-LD @context**: https://policywindow.org/wiki/catalog/jsonld-context
- **RDF Turtle export**: https://policywindow.org/wiki/catalog/rdf (owl:sameAs triples to Wikidata + ELI + CELEX)
- **Per-cell JSON anchor**: https://policywindow.org/wiki/catalog/cells/[instrument]/[topic]
- **Iframe citation card per cell**: https://policywindow.org/embed/cell/[instrument]/[topic] (oEmbed JSON sibling at .../oembed.json)
- **Editorial-board JSON sidecar**: https://policywindow.org/wiki/editorial-board.json
- **Funding JSON sidecar (per quarter)**: https://policywindow.org/wiki/funding/[year]/[quarter].json
- **OpenAPI 3.1**: https://policywindow.org/api/openapi
- **DataCite XML per article**: https://policywindow.org/wiki/[slug]/datacite.xml (DataCite 4.4)
- **Citation-manager formats per article**: https://policywindow.org/wiki/[slug]/citation.bib | citation.ris | citation.csl-json
- **RDF vocabulary**: https://policywindow.org/vocab + per-term resolvers https://policywindow.org/vocab/[term]
- **RSS changelog**: https://policywindow.org/wiki/changelog/feed
- **Permanent snapshots**: any URL accepts `?asOf=YYYY-MM-DD`
- **Sitemap**: https://policywindow.org/sitemap.xml
- **Robots**: https://policywindow.org/robots.txt (declares https://policywindow.org/sitemap.xml)

## New audience-specific dashboards + per-instrument explorers

- **Silent-cells (catalog gaps)**: https://policywindow.org/wiki/silent-cells
- **Cross-instrument enforcement timeline**: https://policywindow.org/wiki/enforcement-timeline
- **NCA regulator briefing console**: https://policywindow.org/wiki/regulators
- **Per-instrument provision explorer**: https://policywindow.org/wiki/[slug]/provisions
- **Per-instrument vendor disclosure form**: https://policywindow.org/wiki/[slug]/disclosure-form
- **Per-instrument harm-narratives (civil-society)**: https://policywindow.org/wiki/[slug]/harm-narratives
- **Citation guide**: https://policywindow.org/wiki/citing-us
- **Citation-reach transparency dashboard**: https://policywindow.org/wiki/citation-reach
- **Agent templates index**: https://policywindow.org/wiki/templates (Claude Project, Custom GPT, Gemini Gem, NotebookLM, Microsoft Copilot)
- **Charter §7 revision draft (open public comment until 2026-06-30)**: https://policywindow.org/wiki/charter/revisions/2026-Q3

## Licensing (canonical)

- Article content: CC BY 4.0 — share + adapt with attribution
- Source code: MIT
- Citation graph (structured citation triples): CC0 1.0 per the I4OC (Initiative for Open Citations) norm — the COVERAGE matrix IS a citation graph; CC0 keeps it ingest-friendly for OpenCitations / CrossRef / academic-aggregator pipelines

Full licensing: https://policywindow.org/wiki/charter §5

## Contact + corrections

- General inquiries + paid services: hello@policywindow.org
- Privacy / data-subject requests (GDPR Art. 15/17/20): privacy@policywindow.org
- Editorial corrections: file via the "Report a problem" link on any article page; or GitHub issue via the public-repo path documented in /wiki/charter §6

## Last reviewed

This llms-full.txt content is generated from the current catalog state on each request. The canonical source pages (linked throughout) are the authority — when this file disagrees with them, the rendered HTML pages win.