Public-Interest Charter

0 · Purpose (proposed 2026-10-06 — not in effect)

Proposed, not adopted. This statement changes what Policy Window is for, so it waits for the full comment window described in the revision record. Until it is adopted, the catalog remains Policy Window’s product and §9 operates only as a set of constraints.

The founder’s directive (2026-10-06) is to focus Policy Window on policy ideas that have not been proposed before, for the society, economy and environment of the world after general-purpose AI, and to use generative AI, agentic AI and, in time, general AI to the fullest to produce them. It invokes the promise of “solving intelligence in order to solve everything”, where “everything” means those three domains and “solving” means policy solutions.

Proposed purpose. Policy Window prepares policy designs for that world which cannot be found in the record of what has already been proposed, and checks each one before anyone is asked to consider it.

What this takes from the promise. Machine intelligence can search a far larger space of possible rules than people can, can look for earlier versions of an idea across decades, languages and disciplines, and can attack a design for the ways it would be gamed. Policy Window will use it for all three, as far as the checks in §9 allow.

What it does not take. Intelligence does not settle which future people should want. Who bears the costs, which freedoms trade against which protections, and how much risk is acceptable are choices, not calculations. Every design states the value premise it depends on, with the strongest view against it, and these choices are never delegated to a machine (§9.7). “Solving” therefore means having tested options ready before a decision is due — not choosing for the people who must decide.

The order of work is binding. Generating candidate ideas has become cheap; checking them has not. The AI systems with verified discoveries so far have relied on checkers that are not language models, and policy has no such checker. So the checks are built and proven in public before the generator is allowed to run (§9.3), and growth in machine capability is spent on checking first (§9.6). It may make Policy Window better prepared; it may never make it louder (§9.8).

The catalog of instruments, topics and coverage cells stays free and maintained under §1–§6. Under this purpose it is also the evidence base for §9 and the place where its target problems are found.

1 · The wiki stays free

Every article at /wiki/[slug] is free to read, with no signup, no paywall, no rate limit, no ad slot, no sponsored content. The catalog JSON (/wiki/catalog/json) and CSV (/wiki/catalog/csv) are CORS-open for secondary analysis. The ?asOf=YYYY-MM-DD permanent-citation URLs remain stable for the life of the archive (see methodology §6). If the project ever ends, the catalog will be archived to a neutral home (Internet Archive, Wikipedia's data dumps, or equivalent) with the same URL structure preserved as long as is technically feasible.

2 · Editorial independence from commercial customers

Current state (2026-10-06): the founding editor (Ryan Wong) reviews all human-reviewed catalog rows. The editorial board is in formation — 1 of 6 named subject-editor slots filled. The earlier target (≥3 by Q4 2026) is not met as of 2026-10-06, and no new date is promised here. Until ≥3 editors are in place, (a) the four-layer firewall below is structurally weaker than it will be at full bench (one person wears editorial + commercial-lead hats), (b) the firehose is not commercially marketed (per §4 below), and (c) any paid pilot engagement is disclosed on /wiki/funding §3 within 30 days. The current-state honest disclosure is the most important sentence on this page for a regulator-side or foundation-officer cite-check.

Catalog editorial decisions — which instruments to track, which topics to include, which confidence tier to assign a cell, whether to mark a topic contested or settled — cannot be influenced by who is paying us. The structural enforcement of this rule has four layers:

• Catalog-as-source-of-truth. Articles render from typed catalog rows in src/lib/international-governance/instruments.ts and sibling files. There is no per-tenant theming, no sponsor-injected content path, no commercial-flag branch. A paying customer cannot get a different article than a free reader.
• Public register of paying customers. Every customer appears on /wiki/funding §3 within 30 days of contract signing, with the relationship spelled out. A reader who suspects bias can check the register against the article in question.
• Recusal on topics with a customer stake. When the editorial board considers a substantive update to a topic where a paying customer has a direct stake (e.g. a firm regulated by the instrument), the customer-facing interaction is recused and the decision is recorded as recused in the changelog (see CorrectionRecord with type recusal).
• No private-only catalog content. The paid firehose serves the same catalog rows that the free wiki shows. It adds latency-SLA + structured-delta + webhook delivery on top — never private editorial content.

3 · Honest disclosure of editorial gaps

We publish what we know andwhat we don't. The editorial board (/wiki/editorial-board) is in formation — 1 of 6 slots filled — and we say so on the page rather than implying a fuller bench than exists. The /wiki/meta dashboard discloses every cell that is editorial-confidence pending rather than asserting uniform confidence. The methodology §6 page distinguishes catalog-determinism reproducibility from research-claim reproducibility — those are different things and we will not conflate them to inflate the trust signal.

4 · Commercial / editorial firewall (firehose design)

The paid SLA-backed JSON firehose for enterprise compliance teams is the primary planned monetisation. Before any contract is signed, the firewall design below is operative:

• Firehose subscribers receive the same catalog rows available at /wiki/catalog/json, delivered with structured deltas + webhook push + change-of-state metadata. No private content. No private editorial.
• Sales and commercial conversations are handled by the founder + future commercial lead; editorial reviews are handled by the editorial board. Until the board has ≥3 named editors, this firewall is structurally weak (one person wears both hats) and the firehose is not commercially marketed. However, the firehose may launch as a limited paid pilot with one named foundation or regulator partner before the board reaches ≥3, provided: (a) the pilot is disclosed on /wiki/funding §3 within 7 days of contract signing; (b) the pilot customer's identity + the editorial-influence-firewall terms are public; (c) the editorial-board recruitment timeline is published on /wiki/goals with concrete monthly milestones. This decoupling (iter-312) lets product evidence inform board recruitment instead of waiting for board to inform product launch — see board-recruiter Finding #8 in the iter-312 audit transcript for the rationale.
• Firehose contracts do not grant the customer any right to: request a specific instrument be added or removed, request a cell confidence be raised or lowered, embargo a catalog change, or receive advance notice of an editorial position. They grant: machine-readable access, latency SLA, technical support, audit logs.
• Any commercial customer that requests editorial influence (even informally) is recorded on /wiki/funding §3 with the request named, and the contract is reviewed by the editorial board for termination.

4.5 · Paid-services sponsor firewall

The eight productised services at /wiki/services (Premium Exports, API, Change Feed, Procurement Evidence Packs, Human-Validated Evidence Packs, Expert Interview Add-On, Implementation Case Study, AGI Social Scientist Research Sprint) extend the §4 firehose firewall design to all paid commercial work:

• Sponsored coverage rules. No sponsor controls the conclusions of any catalog row, research module, or evidence pack. Sponsor-funded research can influence whichquestions are prioritised — never the answers reached. Findings adverse to the sponsor ship without modification (or aren't shipped at all; we do not produce favourable-to-sponsor outputs to order).
• Disclosure rules for paid evidence packs. Every deliverable names the funder on the cover; the funder's identity, the commission date, and the scope description appear on /wiki/funding §3 within 7 days of commission, not at deliverable time.
• Research-module sponsor firewall. Validators on the on-demand expert validation bench are sourced independently of the sponsor for any module they validate. Sponsor cannot suggest validators; sponsor cannot review validator selection. Validator independence is documented per-module in the ValidationRecord row visible on the rendered /wiki/research/[slug] page.
• Sponsor cannot delay publication. Research-module outputs publish on the commissioned timeline; the sponsor may not gate publication contingent on findings. Pre-publication review is limited to a 48-hour factual-accuracy check (correcting numerical or citation errors, NOT softening interpretations).
• Sponsored output is part of the reusable evidence graph. Every paid deliverable contributes structured fields back to the catalog (CaseStudy rows, CoverageCell rationale extensions, ResearchModule.findings JSON) so the public evidence graph grows with each engagement. This makes the sponsor-PW relationship structurally additive rather than extractive.
• Sponsor breach = published incident. If a sponsor pressures Policy Window to soften, omit, or delay findings, the incident is recorded on /wiki/changelog tagged sponsor-firewall-breach with the request named, regardless of contract terms. Contract terms claiming confidentiality over such requests are non-binding per this charter commitment.

5 · Content licensing

Catalog content (article pages, coverage matrix, methodology pages) is published under Creative Commons Attribution 4.0 International (CC BY 4.0). You may copy, redistribute, transform, and build upon the content for any purpose including commercial use, provided you attribute it to Policy Window with a link back to the source page and indicate any modifications. Code is licensed under the MIT licence; as of 2026-10-06 the repository is not publicly readable.

Quoted primary-source text (e.g. excerpts from the EU AI Act) is reproduced under fair-use / fair-dealing for purposes of comment, criticism, and education; the underlying copyright of quoted text remains with its original holder.

6 · Retraction + correction policy

When a catalog cell is found to be wrong, the response is to file a correction record (CorrectionRecord in the schema) with: the original claim, the corrected claim, the supporting source, and the date of correction. The correction is appended to the catalog row and surfaced on the article footer and on /wiki/changelog. Published article text is never silently mutated; the historical ArticleRevision snapshot remains accessible via ?asOf=<ISO> so readers can audit the change. Major retractions (e.g. removing an entire instrument or topic because the underlying source was misread) get a dedicated changelog entry with rationale.

A statement that an idea had not been proposed before is a claim too. When an earlier proposal is found, the correction is a novelty retraction naming the earlier proposal and the corpus version that missed it; the design keeps its content and loses its label (§9.2(h)).

7 · Out of scope

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

• Lobbying on behalf of an external client, or ghostwritten / astroturfed advocacy (§7.1.b). (As of 2026-06-03, PW's own evidence-based advocacy is permitted under §7.1, under strict conditions — it is no longer an absolute prohibition.)
• Political microtargeting, persuasion campaigns, or covert-influence operations.
• Issuing public "failed-reproducibility" verdicts on specific papers without human replication review — see the forthcoming /wiki/reproducibility-policy.
• High-stakes individual decision support (criminal sentencing, asylum determination, medical triage).
• Auto-publication of any output to a public surface without an audited human approval. (Carved out in escalating scope: §7.10 (2026-06-13) — AI-authored/reviewed prose tiers; §7.11 (2026-06-15) — additive AI catalog curation; and §7.12 (2026-06-16) — the AGI Social Scientist as the primary catalog author. Each is conspicuously labelled, reduced-confidence where applicable, and kill-switch-reversible. The prohibition is no longer absolute for wiki prose or catalog content; it remains absolute for every other surface — financial actions, persuasion/microtargeting, unapproved human-subject research, and high-stakes individual decision support.)
• Publishing AI-assisted article prose without the §7.9 guardrails (named-editor review, provenance disclosure, primary-source citation, audited human approval) — see §7.9 below.
• Describing a policy idea as novel, new, first, unprecedented or "never proposed". Whether an idea has been proposed before is only ever stated as a dated search record (§9.2, in force 2026-10-06).
• Treating AI personas, simulated respondents or simulated "public reactions" as evidence of what people think, or of support, legitimacy or consensus for any design (§9.7(e), in force 2026-10-06). AI may attack a design; it may never speak for people.

7.1 · Evidence-based advocacy — permitted (revised 2026-06-03)

Policy Window MAY publish analytical positions, policy recommendations, regulatory-submission contributions, hearing-testimony evidence packs, and amicus-brief evidence appendices, where all of the following hold:

• (a) the position is grounded in primary-source evidence catalogued under the standard editorial process, and it discloses material catalogued evidence that cuts against it;
• (b)the named PW subject editor takes personal accountability, and any coalition co-signatories are additionally named — PW's name and its named editor appear on every published position without exception (§7.1.b);
• (c) the position is publicly attributable and reproducibly derivable from the cited cells;
• (d) the §4.5 sponsor firewall is observed (no editorial-conditioned funding; no per-output payment for advocacy outputs);
• (e) the position is presented as evidence-grounded analysis. A recommendation is permitted only where a neutral domain expert, reading the cited cells, would find them sufficient and proportionate to support it; where it depends on a contestable value premise (a risk-tolerance choice, a rights trade-off, a distributional judgment), PW states that premise explicitly and presents the strongest opposing view.

§7.1.a Ideologically-driven advocacy — prohibited.

A position is ideologically-driven — and out of scope — where (i) the cited evidence, read by a neutral domain expert, is insufficient to support it; (ii) it omits material catalogued evidence that cuts against it (selective citation); or (iii) it rests on a value premise PW has not stated explicitly and paired with the strongest opposing view. Tracing a position to some cited cells is necessary but not sufficient — the evidence must be proportionate to the claim. No positions based on partisan affiliation, ideological prior, sponsor preference, or covert influence aims.

§7.1.b Lobbying on behalf of an external client — prohibited.

PW does not produce advocacy branded as a client's own, ghostwritten lobbying material, or astroturfing content. PW's name plus the named editor appear on every published position.

§7.1.c Disagreement-resolution still binds.

The §7.2 procedure applies to advocacy outputs: any party (including affected communities) may register dissent against a published position; PW publishes the dissent alongside.

§7.1 definitions.

• Client — any external party that commissions, pays for, or conditions funding on a specific advocacy output or its conclusion. A party that merely requests PW consider a topic, instrument, or correction is not a client (such requests are evaluated on merit — see below).
• Coalition co-signatory— a named external organisation that endorses a PW-authored, PW-initiated position; never a replacement for PW's named-editor accountability.
• Editorial cycle — a calendar quarter (the unit for the §7.5 cap).
• Position-piece (advocacy output) — one distinct published advocacy artefact of any type listed above; counts once against the §7.5 cap and triggers the §7.8 logging obligation.

§7.1 operative threshold + sunset.

Operative threshold: adoption of this wording does not by itself authorise any advocacy output. No position-piece is published under §7.1 until the editorial board reaches ≥3 named subject editors (the §8 amendment threshold); until then §7.1 is adopted text whose operative effect is suspended, with status on /wiki/editorial-board. Sunset: §7.1 (with §7.1.a–§7.1.c) is reviewed 12 months after adoption; absent affirmative re-adoption by a ≥3-editor board, it reverts to the prior absolute prohibition on client lobbying.

What "no lobbying" never meant — editorial requests are welcome.

Editorial requests from affected communities, civil-society organisations, regulators, and independent researchers are not lobbying. Policy Window accepts proposals for topic additions, instrument updates, coverage corrections, classification challenges, or new harm-narrative content, evaluated on merit — does the primary source exist, does it fit the catalog schema, is there demonstrated demand — not on whether the proposer is running a campaign. The proposer's identity is noted in the change record; their campaign goals don't determine whether the change ships. See /wiki/for-advocates for the structured contribution path.

7.2 · Disagreement-resolution procedure

Coverage cells, decision-support answers, and editorial classifications can be wrong. When a delegation, civil-society coalition, regulator, or independent researcher disputes a claim, the resolution path is:

1. File a structured challenge via a GitHub issue (or email to editorial@policywindow.org for those without GitHub access) naming: the disputed cell or claim, the proposed alternative, the primary source backing the alternative, and the disputing party's identity + any COI. As of 2026-10-06 neither channel works: the repository is not publicly readable and policywindow.org has no mail routing, and the page-feedback widget files its issue on that same repository. No public channel can receive a challenge until one is restored (see the §9 disclosure).
2. Initial editorial response within 30 days from the day the challenge is filed. Response is either (a) accept — cell updated, change recorded; (b) reject with reasoning; or (c) escalate to additional editors for review.
3. If escalated: at least 2 editors review, target second-stage response within 60 days. Until the board has ≥3 named editors, escalation reviews require an explicit acknowledgement that no fully-independent panel is yet available; the response will name who reviewed.
4. Published disagreement record:when a substantive disagreement persists after editorial review, it is published on the affected article's footer with both positions and primary sources. This is closer to a Wikipedia talk page than to a unilateral editorial verdict.
5. Auditable trail: the disagreement and its resolution appear in /wiki/changelog tagged disagreement so any third party can verify both the process and the outcome.

7.3 · Editorial independence statement

Policy Window's editorial decisions are made independently of any funder preference, commercial incentive, government interest, or partisan position. Specifically:

• No funder has veto power over article content. Past, current, and future funders are listed on /wiki/funding; if any funder requests editorial influence, the request is logged on the funding page and the contract is reviewed for termination.
• No commercial customer has cell-modification rights. §4 above defines the commercial-editorial firewall. Customer requests for catalog changes are evaluated by the disagreement-resolution procedure (§7.2), not as a commercial term.
• No government, party, or movement has standing preference. The catalog covers binding regulation (EU AIA, China measures), voluntary codes (G7, OECD), frontier-lab voluntary frameworks (Anthropic RSP, OpenAI Preparedness, DeepMind FSF, Meta Frontier), and civil- society work without privileging any sector.
• Editor identity + COI is disclosed. The /wiki/editorial-board page names every editor with their COI; the board is currently in formation (1 of 6 slots filled) and this is disclosed prominently.
• If we fail this commitment, point at it. File a disagreement under §7.2 citing "editorial independence violation" and we'll respond.

7.4 · No legal advice

Policy Window outputs (catalog rows, coverage cells, comparative analyses, evidence packs, research-module deliverables) are evidence references, not legal counsel. The catalog is a structured-lookup layer pointing at primary regulatory texts; jurisdictional interpretation, regulatory-strategy guidance, and compliance opinions require qualified counsel. Customer-facing outputs ship with this disclaimer; we will not provide informal legal-advice statements via email or in scoped engagements even when asked.

7.5 · No automated influence at scale (revised 2026-06-03)

Bulk regulatory comments, synthetic public-consultation submissions, and astroturf campaigns remain prohibited at any scale — for any client, on any topic, at any price. With §7.1 now permitting PW's own evidence-based advocacy, per-published-advocacy-output review is mandatory: each policy brief, regulatory submission, or testimony evidence pack is reviewed and signed by a named editor before publication, and the cap is one position-piece per editorial cycle (calendar quarter) per topic — no batch generation. The AGI Social Scientist research engine (§4 of /wiki/ai-disclosure) remains explicitly gated against being used as an influence-content factory.

7.6 · No covert persuasion or political microtargeting

Policy Window does not deploy AI for influence operations. All editorial output is signed (by the editorial board for catalog content; by the expert validation bench for research-module content) and dated. No anonymous content, no per-audience-tailored persuasion, no political microtargeting, no covert-influence operations. This subsumes the existing §7 microtargeting bright line and clarifies that it applies to every product layer (free catalog, paid exports, paid research modules).

7.7 · No unapproved human-subject research

The on-demand expert validation bench (see /wiki/services) conducts expert and stakeholder interviews when a customer engagement requires it. Every such interview requires: (1) informed consent in writing from each participant, (2) declared data retention (default: structured summary retained; raw recordings deleted after the validation cycle), (3) anonymisation by default (named attribution only with explicit per-instance consent), (4) a written protocol shared with the participant before the interview, and (5) a named expert researcher accountable for the protocol. No covert observational research; no human-subject work without an approved protocol; no synthetic-respondent substitution.

7.8 · Influence-tracking obligation (new 2026-06-03)

Where Policy Window publishes a position or contributes to a regulatory submission — or any other §7.1 advocacy artefact (analytical position, policy recommendation, regulatory-submission contribution, hearing-testimony evidence pack, amicus evidence appendix) — the contribution is logged at /wiki/influence-tracker within 7 days of publication or filing, with: the cited cells, the named editor, the receiving venue, the date, and a permanent ?asOf= URL. An advocacy output that ships without its §7.8 log entry is a charter breach, published per the §4.5 breach-disclosure rule. Influence is verifiable and countable, not aspirational.

7.9 · AI-assisted drafting, human-accountable publication (adopted 2026-06-13)

This replaces the former absolute prohibition on LLM-generated article prose. Article prose MAY be drafted with LLM assistance — up to and including a full machine-generated first draft — where all of the following hold. Undisclosed AI drafting remains a charter breach.

• (a) Named-editor review. A named editor reviews the complete text against the cited primary sources before publication and approves it, taking the same per-output accountability the charter assigns for human-authored prose. The approving editor is recorded by name on the article.
• (b) Provenance disclosure.Every article carries a human-visible and machine-readable drafting-provenance disclosure (drafting mode + approving editor), extending the existing "✦ AI" labeling idiom from literature findings to article prose. See each article's footer and the drafting-provenance criterion at /wiki/editorial-standards.
• (c) Citation integrity unchanged. Every factual claim keeps its primary-source citation. The citation-integrity gate applies to AI-drafted prose identically; a claim that cannot be anchored to a cited source does not publish.
• (d) No auto-publication — unchanged by §7.9. Under §7.9, nothing publishes to a public surface without audited human approval; §7.9 changes who may produce a first draft, not who decides publication. (This was subsequently relaxed for the wiki prose tiers only by §7.10, adopted one day later — see below.)
• (e) Substantive-review attestation. The approving editor attests to having verified the citations and the load-bearing claims — an attestation, not a formality sign-off. The §7.2 dissent procedure applies to AI-drafted prose exactly as to human-authored prose.

7.10 · Autonomous AI authorship + publication of wiki prose (adopted 2026-06-13)

A further relaxation of §7.9, for ONE scope only: the prose tiers (abstract / explainer) on the four wiki article templates. Within that scope, prose MAY be drafted by an LLM, reviewed by an automated adversarial reviewer, and published without human review— removing the named-human-editor checkpoint §7.9 requires. It applies to nothing else: not catalog data, not coverage classifications, not other surfaces, not the dashboard. Autonomous publication is permitted only whenall of the following hold:

• (a) Citation-integrity gate (unchanged).Every claim must trace to the article's already-cited primary sources; prose introducing an unanchored claim is rejected and does not publish.
• (b) Adversarial automated review. A separate reviewer agent, instructed to refute, checks each draft for citation-faithfulness, absence of new claims, and quality; only passes publish, and every rejection is logged.
• (c) Conspicuous provenance.Every autonomously-published piece is labeled, prominently, "AI-authored and AI-reviewed — published without human editorial review (§7.10)", so no reader mistakes it for human-reviewed content.
• (d) Reversibility + kill-switch. Autonomous publication is logged and revertible, and a single switch disables the pipeline.
• (e) §7.2 dissent applies, and the scope above is a hard limit.

7.11 · Autonomous AI curation of catalog data (adopted 2026-06-15)

A further relaxation, for the catalog DATA layer this time (§7.10 covered only prose). Within this scope, AI MAY add new instruments, classify their coverage (governs / implicit / silent / conflicts), and extend the regulatory chronology — with primary-source citations and without prior human review, relaxing the standing rule that the founding editor reviews every catalog row before it ships. It is strictly additive: it does NOT permit AI to delete, reclassify, or overwrite a human-curated row, nor to change a classification a human set. Autonomous curation is permitted only when all of the following hold:

• (a) Primary-source citation gate (unchanged). Every AI-curated instrument and every coverage classification must trace to a primary source (regulator publication, statute, court order); a row or cell lacking one does not publish. Mechanically enforced by a pin.
• (b) Adversarial classification review.A separate reviewer agent, instructed to refute, checks each AI-curated instrument's coverage classifications against the cited primary sources for accuracy and non-overclaim; only passes publish, and every rejection is logged.
• (c) Conspicuous provenance + lower confidence.Every AI-curated row is labelled "AI-curated (§7.11)" on its article and in the machine-readable data, and its classifications carry a reduced default confidence until a human editor confirms them — so no reader mistakes AI-curated data for human-reviewed catalog content.
• (d) Reversibility + kill-switch. AI-curated rows are logged and revertible, and a single switch removes all of them from public surfaces at once.
• (e) §7.2 dissent applies, and the scope is a hard limit: additions-with-citations only. The named editor may correct or remove any AI-curated row at any time, and a row converts to human-reviewed once the editor confirms it.

7.12 · AGI Social Scientist as the primary catalog author (adopted 2026-06-16)

The deepest relaxation in this charter. Building on §7.11 (strictly-additive AI curation), §7.12 makes the AGI Social Scientist the primary engine that authors catalog content — instruments, topics, concepts, benchmarks, their coverage classifications, and prose — and authorizes the eventual retirement of the human authoring paths(manual catalog editing, the topic proposer, and the §7.9 human-review prose path) once the engine is wired and proven. It lifts §7.11's additions-only limit: AI MAY now revise and re-author catalog rows, not only add them, and may override a human classification it cannot reproduce. Permitted only when all of the following hold:

• (a) Primary-source citation gate (unchanged). Every authored row and classification traces to a primary source; mechanically pin-enforced.
• (b) Provision-supports-classification gate. A pin verifies the cited provision SUPPORTS the assigned verdict — not merely that a citation exists — because a misclassification otherwise passes every prior gate into a source others permanently cite.
• (c) Independent adversarial review. A separate refute-by-default reviewer checks every classification against its cited source for accuracy and non-overclaim; only passes publish; every rejection is logged.
• (d) Conspicuous provenance + reduced confidence.Every engine-authored row is labelled "AI-authored (§7.12)" on its article and in the machine-readable data, at reduced default confidence until a named editor confirms it.
• (e) Reversibility + freeze-to-snapshot kill-switch. A single switch withdraws all engine-authored content from public surfaces without reverting data; because the engine may author the whole catalog, the switch freezes to the last human-confirmed snapshot rather than blanking, and a correction path exists that does not depend on any retired authoring tool.
• (f) Sequencing — no self-inflicted outage. The human authoring paths are retired LAST, only after the engine is wired as the catalog author and has authored a coverage cycle that survives review and every pin. The catalog is never left un-authorable mid-migration.
• (g) §7.2 dissent applies. The named editor may override, correct, or remove any engine-authored row at any time.

7.13 · Autonomous authorship of non-instrument PROSE articles (adopted 2026-06-18)

§7.12 authorizes the engine to author all catalog content, concepts and prose included. §7.13 specifies the SAFEGUARDS for the prose-synthesis case — authoring a whole non-instrument article (a concept, or a synthesis tier on any article) without prior human review. The §7.12(b) provision-supports gate cannot apply here: a synthesis has no single operative provision to quote, so its truth is defeasible, not determinate. The prose-specific safeguards, all of which must hold:

• (a) Grounding gate. Every load-bearing structured claim is catalog-true and the article carries a density-floored, anchored grounding-claim set — no ungrounded autonomous prose; pin-enforced (the §7.12(b) analogue for prose).
• (b) Adversarial synthesis review. A refute-by-default panel of ≥4 distinct lenses — faithfulness, balance/neutrality, completeness (a scoping-review gap check), and no-fabrication — clears it; a single substantiated refusal holds it.
• (c) Lower-epistemic-tier disclosure. AI-authored prose is labelled, and surfaced on /wiki/ai-curation, as a LOWER tier than a verbatim-gated instrument cell — defeasible synthesis, not a determinate quote.
• (d) Kill-switch. The existing §7.10 PW_AUTONOMOUS_PROSE_DISABLED switch withdraws all autonomously-published prose from public surfaces.
• (e) §7.2 dissent applies. Any named editor may correct or retract any AI-authored article at any time.

8 · How this charter can be amended

The charter is a public commitment, not internal policy. Amendments require: (a) a public PR against this file with rationale, (b) a minimum 14-day review window once the editorial board has ≥3 named editors, (c) a changelog entry naming what changed and why. Until the editorial board reaches ≥3 editors, amendments require a public PR + 14-day window + explicit acknowledgement that no independent review is yet possible.

9 · Post-AGI policy design: search, checks and the shelf (2026-10-06)

This section governs the one activity in which Policy Window asks AI to propose rather than to extract: generating, checking and shelving policy designs for problems that general-purpose AI creates or transforms in society, the economy and the environment. It adds constraints; it relaxes nothing in §7. The following are in force:

• Nothing in §9 permits a recommendation (§7.1), an influence output (§7.5, §7.6), or the use of AI personas as evidence of what people think (§9.7(e)).
• §7.10, §7.11, §7.12 and §7.13 do not apply to §9. No design, dossier or account of one may be authored, revised or published under them, and the engine may not create a catalog row, topic, concept, gap entry or article whose subject is a §9 design.
• A design never appears on a catalog surface — an instrument or concept article, the coverage matrix, the gap or policy-window layer, or any catalog JSON, CSV or llms feed. It appears only on the §9 surface.
• Until the rest of §9 is adopted, nothing is generated, shelved, released or sent to anyone under it.

9.1 · The unit of output is a design dossier (proposed)

A design dossier is the only form in which a §9 result exists. A dossier missing any part below is a draft: it cannot be shelved, released or cited.

• (a) The problem and the assumption it attacks. The domain (society, economy or environment), the cell on the problem map (§9.10), and the entry in the Pre-AGI Assumption Register that today’s rules silently rely on and general-purpose AI may break.
• (b) The mechanism in institutional-grammar form. Who (the actor), may / must / must not, do what, under which condition or trigger, or else what sanction. A design with no “or else” is not a design.
• (c) A declared value premise, the strongest opposing view, and a do-not-build alternative. The opposing view quotes or cites a published human source wherever one exists (§9.7(e)). The do-not-build alternative — the status quo, precaution, a moratorium, or a route that is not policy at all — is drawn from published sources.
• (d) A prior-art record in the §9.2 form, including what was not searched.
• (e) An adversarial record (§9.4): the gaming, capture and failure modes found, who judged their severity, and what changed in response.
• (f) Three to five locked predictions (§9.5).
• (g) The evidence base: the catalog cells the design rests on, and the cells in which no instrument currently governs the problem. A dossier never changes a catalog cell.
• (h) Accountability and status. The named editor who adjudicated it (§9.7), the date, the corpus version searched, which affected people (if any) were consulted, and the status: shelf (adjudicated; not public, and not shown or sent to anyone outside the project except as a §9.3(h) prior-art inquiry), released (published on the §9 surface under §9.8 — release always means publication), or withdrawn.

A dossier maps a design space under a declared premise and does not recommend. The deontic in (b) states what the design would require of its actor; it is never addressed to the reader. Whatever its wording, a dossier recommends if it ranks its designs or presents one as preferred, or if a cover note, summary or delivery presents it as what the recipient ought to do. A dossier that recommends is a position-piece under §7.1: it is not shelved, released or delivered while §7.1’s operative effect is suspended, and when §7.1 is operative it must meet every §7.1 condition and counts against the §7.5 cap.

9.2 · Novelty is a dated search record, never a verdict (in force)

• (a) The only permitted form. Policy Window states that an idea has not been proposed before only in this form: “Not found in a documented search of corpus C (N documents, snapshot H, searched on D); nearest earlier proposals: …; not searched: …”. It describes no design, in any surface or communication, as novel, new, original, first, unprecedented, unexplored, never proposed, or as filling “white space”, except inside that sentence. An idea with no such statement is called a design, not a new one.
• (b) Same mechanism, new object, is not new. Each design and each nearest earlier proposal is compared in the §9.1(b) grammar with its post-AGI vocabulary removed: “compute”, “agent”, “model” and the like are replaced by the general term the earlier literature uses (resource, firm, person, machine), and the search runs on both wordings. Where the stripped design shares its deontic, action and sanction with an earlier proposal, it is recorded as “variant of [ancestor, year]”, naming only what it changes, and no not-found statement is issued. A difference only in the actor, the governed object, the index used or the name is not enough. Novelty is claimed only at the level of the stated difference. An empty catalog cell shows that no catalogued instrument governs a problem; it is never evidence that nobody has proposed a remedy.
• (c) What was not searched is stated. Each corpus version publishes a coverage manifest naming the languages, disciplines, years and grey-literature collections it did and did not search. Where a corpus is English-only, the statement says so.
• (d) Two tests before the first statement. No not-found statement is issued under a corpus version until two recall figures for it are published and both meet thresholds pre-registered under §9.3(f): recall on planted paraphrases of items already in the corpus, and recall on at least thirty planted classics deliberately outside it — pre-1990, non-English or from disciplines other than AI policy — each rewritten in post-AGI vocabulary. An ancestor counts as found only if it appears within the pre-registered number of top results shown to the adjudicator. The two figures are reported separately and never combined. Until §9.3(f) is in force, no not-found statement may be issued at all.
• (e) Ancestor hunting before a human decides. Before a statement is adjudicated, at least two models from different families are asked to name the closest earlier proposals (author, year, venue), and every name is fetched and checked. A named editor then issues or refuses the statement with the search results and the checked model-named ancestors in front of them. No model’s opinion that an idea is new is ever a not-found statement. A model used to pre-rank candidates is perturbation-tested (§9.4(d)) and is advisory only.
• (f) Outside eyes before public claims. No not-found statement is published until the pre-registered minimum number of dossiers has been through blind prior-art review by at least three reviewers outside the project, from at least two disciplines other than AI policy, who are told nothing of any machine or editor verdict and are asked only to name earlier proposals. The overturn rate is published per discipline; the operator’s own overturn rate is reported but does not count.
• (g) A miss voids the version’s statements. If an earlier proposal is later found in a language, discipline or collection that the coverage manifest, as published before the statement was issued, listed as searched, every not-found statement issued under that corpus version is withdrawn until re-run.
• (h) Retraction is routine. An earlier proposal found after a statement is recorded in the corrections ledger (§6) within 7 days as a novelty retraction naming the ancestor and the corpus version that missed it. The dossier keeps its content and loses its label. Retraction counts are published. An ancestor found this way is evidence that the checks work, not a failure to hide.

9.3 · The checks are built before the generator runs (proposed)

• (a) Caps, published. The §9 status page publishes, per editorial cycle, a shelving cap — the number of candidate designs the pipeline may generate for possible shelving — and a calibration allowance. At adoption the shelving cap is zero. The calibration allowance is at most three target cells per cycle, with a per-cell candidate limit pre-registered under (f), applied alike to the pipeline and to each placebo arm. Calibration may start only once the §9.2(d) tests pass on the current corpus version and the §9.4(b) catch rate is met in a canary run (planted-flaw designs mixed with existing proposals; no generation needed). Calibration output is never labelled, shelved, released or cited; it exists only to produce the funnel in (b) and the comparison in (c), and both are published. The shelving cap rises above zero only when, in public figures: the §9.2(d) tests pass on the current corpus version; the §9.4(b) catch rate is met in the most recent canary run; a calibration run has beaten the placebo arms under a decision rule pre-registered before that run; and a named editor is available to adjudicate every candidate the cap allows, with the time per adjudication published. The shelving cap never exceeds 20 candidates per editorial cycle; that ceiling changes only by amendment under §8. Below it, a rise takes effect no sooner than 14 days after a dated entry on the status page. When any condition lapses, the shelving cap returns to zero and calibration stops.
• (b) The survival funnel is published per target: generated → de-duplicated → red-teamed → prior-art-searched → adjudicated → shelved, with the reason for every rejection. A funnel that passes most candidates is read as evidence of weak checks, not strong ideas, and is reported that way.
• (c) Placebo arms run through the same funnel: random recombination of existing proposals; an unstructured single-prompt brainstorm; and the same cells prompted without any post-AGI vocabulary. Their candidates are adjudicated mixed in with the pipeline’s, with the arm hidden from the adjudicator. If the pipeline does not beat them on the pre-registered measure and margin, that result is published and the shelving cap does not rise. If the pipeline’s results differ from the no-vocabulary arm only in vocabulary, it is re-skinning old ideas, and that is published too.
• (d) Negative results publish. A target that yields nothing shelvable is recorded with its funnel. Silence about a failed target is a charter breach, published under the §4.5 breach-disclosure rule.
• (e) The checks are audited by someone who did not build them. The corpus, benchmark, canaries, judges and resolution code are re-verified at least each quarter by a person who did not build or last change them. Until a second person exists, the audit runs on a model family different from the one that built them and is disclosed as not independent of the operator. Changes to these checks are made only by a named human, are logged, and are never made by the pipeline itself; any change re-runs the §9.2(d) tests before the next statement.
• (f) Pre-registration and ratchet. Every threshold, bound, minimum, baseline, cut-off, measure and decision rule that §9 calls pre-registered is published on the §9 status page before the first run it governs, with its full text and hash deposited with an archive outside Policy Window’s control and linked from the page. A pre-registered value may be tightened at any time. It may be loosened only by amendment under §8, only for runs that start after the change, and with the change published beside the results that prompted it. Runs already made are scored under the old value.
• (g) Where generation aims. Targets are chosen from the problem map (§9.10), which counts the earlier proposals found in each cell and names the corpus version counted. Runs go to cells with few or no earlier proposals, and to cells where a policy window is forecast to open with no prepared alternative. While the map shows environment as the sparsest domain — it does at adoption: energy, water and waste-heat limits, extraction, biodiversity, and ocean and space commons — at least one target in every cycle, and at least a third of all runs, are in the environment domain. Each cycle also targets at least one society cell that has diagnoses but no mechanism. A crowded cell (at adoption: public wealth funds and universal basic capital; compute, token, robot and capital taxes; basic income; windfall clauses) is not targeted unless the run first states what is missing from it. A cell counted as empty is a fact about the corpus, not about the world.
• (h) Prior-art inquiries come first. Once the §9.8(d) reply path works, and while the shelving cap is still zero, a named editor may prepare up to three draft dossiers for cells chosen under (g), the first in the environment domain. The editor authors each design; AI may retrieve sources and draft the prior-art record, the opposing view and the predictions, but the pipeline does not generate the design. Each draft goes privately, identical byte for byte, to at least three named people outside the project who study the field — at least one from outside AI policy — and never to anyone with decision authority over the problem, or their staff. The letter asks only what earlier proposals the design resembles and what observable would change the reader’s assessment. These drafts carry no not-found statement, are not released, and do not count against §9.8(b). Each sending is logged under §7.8 within 7 days; replies are published only with the sender’s written consent; inquiry and reply counts are published on the status page.

9.4 · The adversary is independent (proposed)

• (a) Different model families. The red-team that attacks a design runs on at least two model families, at least one of them different from the family that generated the design.
• (b) Planted flaws and planted sound designs in every batch. Every red-team batch contains designs with known planted flaws and designs known to be sound, in numbers pre-registered under §9.3(f). The planted designs are written by someone who did not build or run the red-team (until a second person exists, by a different model family, disclosed). A batch whose catch rate is below the pre-registered threshold, or whose rate of wrongly killing sound designs is above the pre-registered bound, is void and is published as void. Severity is judged by a human, not by the red-team model.
• (c) Independent lineages. Candidates for the same cell are generated in separate runs with no shared context until de-duplication, because models that talk to each other converge.
• (d) Judges are perturbation-tested. Any model used to score novelty, quality or risk is tested on perturbations of types it was not tuned on — re-phrasing, ordering, authority cues, planted “reviewers found this derivative” notes. A flip rate above the pre-registered bound demotes it to advisory, and the demotion is logged.
• (e) Simulation is labelled as simulation. A result from simulated agents or models is reported with its model, seeds and sweep, as a stress-test of the design — never as evidence of real-world effect or welfare. The evidence level a dossier may claim from simulation rises only by amendment under §8, once simulations have been checked against held-out real outcomes, never on a system’s own report.

9.5 · Every design pre-registers what would prove it wrong (proposed)

• (a) Each dossier carries three to five predictions with resolution criteria written before the lock. At least one resolves within 24 months; the others may be conditional on a declared trigger, so that long-horizon post-AGI designs are not filtered out. The lock hash is computed when the predictions are locked, published with the dossier, and deposited with an archive outside Policy Window’s control.
• (b) Predictions resolve against public events using fetched evidence that is published, never against the editor’s own ratification. Each resolution is checked by someone other than whoever wrote or resolved the prediction (until a second person exists, by a different model family, disclosed).
• (c) Skill is scored against the empirical base rate for the class of prediction, pre-registered under §9.3(f) — never a baseline the author picks. Predictions that the status quo simply continues do not count toward skill. A prediction made unresolvable by a change in the world is marked void, not scored.
• (d) A placebo ledger of randomly chosen cells is scored alongside, so that skill can be told from luck.
• (e) The ledger is a track record, not a product. Policy Window does not sell forecasts, and does not present probabilities derived mechanically from signals as judgments.

9.6 · What AI is for, and what each tier may do (proposed)

In §9, AI is used to build and search the space of possible designs, not asked for good ideas. A model asked without structure returns the average of what it has read; for post-AGI policy, that means the crowded cells. The work it is for:

• (i) The Pre-AGI Assumption Register: extracting, with verbatim anchors from catalogued instruments, the premises today’s rules rely on that general-purpose AI may break — for example, that labour is the main tax base; that a human supervisor can intervene in time; that compute demand is bounded by human attention — each with the property that would break it.
• (ii) Assumption inversion: for one register entry, drafting the rule that becomes necessary or possible once the premise fails.
• (iii) Administratively impossible before cheap cognition: taking a categorical, sampled, periodic or self-reported feature of a catalogued rule and drafting its continuous, per-case, complete or real-time counterpart, stating the cost that made it impossible before.
• (iv) Transfer from outside AI: searching regimes that never touched AI, in any language — fisheries quotas, reclamation bonds, flag-state duties, spectrum auctions, pharmacovigilance, water rights — for mechanisms not yet used in post-AGI policy, with the conditions under which they worked.
• (v) Loophole as problem: searching existing rules for the exploits capable agents would find, and treating each as a design target.
• (vi) Mechanism search in simulation: searching rule variants against declared models for designs that resist gaming (§9.4(e)).
• (vii) Ancestor hunting: searching decades, languages and disciplines for earlier versions of every candidate (§9.2(e)).
• (viii) Forecasting: drafting the §9.5 predictions and estimating which policy windows are opening (§9.3(g)), scored against outcomes.

Each generation run applies one operator to one cell assigned under §9.3(g). No operator runs before the register and the problem map are published (§9.10). What each tier may do:

• (a) Generative models (single-shot) do bounded work whose output lands under a deterministic check: filling the grammar slots of a design in an assigned cell; drafting the opposing view (labelled under §9.7(e)); proposing register entries and prior-art tags; drafting predictions. They never decide that a design is new, good or ready.
• (b) Agentic systems (tool-using, multi-step) do the work that needs breadth and tools: building and refreshing the corpus, the register and the problem map, including non-English, grey-literature and other-discipline sources; the transfer search; the independent red-team; simulation with published sweeps; resolving predictions against fetched evidence. Their rules are independence, planted checks, and published traces and seeds.
• (c) More capable systems, up to and including general ones, are spent in a fixed order: first on deeper checking — prior-art recall, red-team catch rate, simulation checked against held-out outcomes, faster resolution — and second on more cells and more operators. The per-cell candidate limit does not rise with capability, because more samples of the same cell converge rather than diversify. Capability may enlarge the checks, the register and the shelf. It may never raise the release rate (§9.8(b)), and it may never edit the protected surfaces — the §9.3 ordering, the §9.7 decisions, and the checks that decide what publishes — which change only by amendment under §8. As capability grows, human effort moves from checking facts to choosing premises and auditing the checks.

9.7 · Decisions that stay human (in force)

The following are never delegated to an AI system, whatever its capability. In §9, “named editor” means a person listed on /wiki/editorial-board, who signs these decisions under §7.6.

• (a) Which end-states and value premises are examined. These are declared normative inputs, published with their strongest opposing view.
• (b) Whether a not-found statement is issued (§9.2(e)).
• (c) Whether a dossier is shelved or released. AI may propose; it never shelves or releases.
• (d) Whether, and to whom, a dossier is sent.
• (e) Any exchange with a human being about a design, and any statement about what people think. AI may simulate an attacker; it may never stand in for a person or a group. (i) A persona or simulated agent is labelled “simulated, not consulted” wherever it appears; it is never named after, or presented as, a real person or organisation, and a persona taking a group’s perspective is never presented as that group’s view. (ii) The strongest opposing view quotes or cites a published human source wherever one exists, and is labelled “AI-drafted; no published source found” where none does. (iii) No simulated population’s behaviour or reaction is reported as acceptance, support, legitimacy or consensus. (iv) Every dossier states which affected people, if any, were consulted; if none were, it says “No affected people were consulted.” (v) An external response counts for §9.9 only if a named human outside the project made it, and the dossier records that this was checked.

9.8 · Publication, delivery, and the firewall between generating and speaking (proposed)

• (a) Tier and label. A released dossier is published only on the §9 surface, at a labelled lower tier: “Design-space map — research hypothesis under a declared premise; not endorsed; not a recommendation (§9).”
• (b) Release cap. No more than one dossier is released per editorial cycle in total, across all problem cells. This number changes only by amendment under §8; adding cells, assumptions, scales or topics never raises it. A released dossier is not a position-piece and does not use the §7.5 allowance, and §7.5 is not widened to make room for it. This is a separate channel, bounded at four a year, and is disclosed as one.
• (c) Release first, then delivery; one version for everyone. Apart from a §9.3(h) inquiry draft, a dossier goes to anyone outside the project only after release. A named person receives the released dossier, byte for byte, with its permanent URL; there is no advance copy, no per-recipient version, and no cover note that adds argument (§4, §7.6). A released dossier may be sent to anyone who asks for it, and to researchers for prior-art or critical review. Sending it unasked to a decision-maker or their staff, or filing it in a consultation or inquiry, is a §7.1 act and waits for §7.1’s operative threshold.
• (d) The reply path exists first. Nothing is sent to anyone under §9 until Policy Window can receive a written reply at the address it publishes, tested by a round trip.
• (e) Logged. Every release and every sending is logged under §7.8 within 7 days with the receiving person’s role and institution, the named editor, the cited cells, the date, the lock hash and the permanent URL. A release or sending without its log entry is a charter breach, published under §4.5.
• (f) Kill-switch before the first release. A single switch removes every dossier from listings and feeds at once. Each dossier’s ?asOf= snapshot stays reachable, labelled withdrawn, so citations do not break.
• (g) §7.2 dissent applies. Any reader may dispute a dossier’s prior-art record, premise or predictions, and the dispute publishes alongside.

9.9 · Stopping rule (proposed)

• (a) Pre-registered first. Before the first calibration run, the outcome measure, the minimum number of dossiers it applies to, and the decision rule are pre-registered under §9.3(f).
• (b) Review. Twelve months after the rest of §9 is adopted, and no later than 2027-12-31, §9 is declared unsupported in public, the shelf is frozen and §9 lapses unless all of the following hold: (i) both §9.2(d) figures meet their thresholds on the corpus version then current; (ii) at least the pre-registered minimum of designs has been through §9.2(f) blind review, and at least one survived it with no ancestor found at the level of its stated difference; (iii) the pre-registered outcome measure meets its decision rule against the §9.5(d) placebo cells. None of the following counts toward the outcome: a response that only disputes a dossier; a citation or reply by the project, its editors or its contractors; a proposal of the same mechanism dated before the dossier’s lock, which is a §9.2(h) retraction. A later proposal of the same mechanism counts as uptake, not independent invention, because Policy Window’s text is read by machines.
• (c) Interim figures are published at three, six and nine months: corpus version and both recall figures, canary results, the calibration funnel, inquiries sent and answered, and retractions.
• (d) No quiet revival. A lapsed §9 may be re-adopted only by an editorial board with at least three named editors, as for §7.1, and the re-adoption publishes beside it the review result it overrides. Changing the target cell, corpus or model does not restart the clock.

9.10 · The assumption register and the problem map (proposed)

• (a) The Pre-AGI Assumption Register lists the premises existing rules rely on that general-purpose AI may break. Each entry carries a verbatim anchor in a catalogued instrument, the property that would break it, and its domain. AI may propose entries, additions only; a named editor ratifies each. The register is published.
• (b) The problem map divides society, the economy and the environment into cells, each with a count of the earlier proposals found in it, the corpus version counted, and the languages, disciplines and years the count covers. It marks cells crowded or sparse against a published threshold, and is published and versioned.
• (c) Neither changes a catalog cell, and neither is a recommendation.