Charter revision 2026-10 — post-AGI policy design

1 · The directive

I want to focus the effort of Policy Window on generating ideas that have not been proposed before; consider the promise of “solving intelligence in order to solve everything”; here, everything is specifically about everything about the society, economy and the environment — as they relate to the post AGI world; and the solution (solving) refers to the policy solutions. Given these, suggest how one can maximise the potential of using GenAI and Agentic AI or AGI to achieve that.

— the founder, 2026-10-05/06.

2 · Why the revision takes this shape

• Generation is cheap; checking is not.In a large blind study, ideas generated by a language model were judged more novel than experts’ ideas (Si, Yang & Hashimoto, 2024), but the same study found that generation collapses into repetition: of 4,000 seed ideas, only about 200 were unique. Follow-up work that executed the ideas found the advantage did not survive execution. The AI systems with verified discoveries so far have relied on checkers that are not language models; policy has none. So §9 builds the checks first and keeps the generator at zero until they work.
• “New” is usually old. During the research for this revision, every candidate “new” post-AGI mechanism the design work produced had an earlier version within a short web search — for example, a compute-rent dividend under an energy ceiling resembles Herman Daly’s depletion quotas (1977); reversibility bonds resemble Costanza and Perrings’ environmental assurance bonds (1990); and social insurance not funded from wages resembles the Austrian value-added levy debate of the 1980s. The ancestors were old, non-English or in other disciplines — outside any recent AI-policy reading list. Hence §9.2: novelty is a dated search record with its blind spots stated; the search is tested on planted classics from outside its corpus; and the same mechanism with a new object is a variant, not a new idea.
• Where the gaps are. Economic proposals for the post-AGI world (wealth funds, compute and robot taxes, basic income, windfall clauses) are crowded. Social proposals are rich in diagnosis and poor in mechanism. The environment seen through a post-AGI lens is the emptiest. Hence the environment quota in §9.3(g).
• What AI is for. The founder asked how to get the most from generative, agentic and general AI. §9.6 answers with work, not adjectives: assumption registers and their inversion, rules that were administratively impossible before cheap cognition, transfer from regimes outside AI, loopholes as design problems, mechanism search in simulation, ancestor hunting, and forecasting. Capability gains are spent on checking first, then breadth — never on release rate.
• Values stay human. Intelligence can widen the options and show their trade-offs; it cannot settle which future people should want. So every design declares its premise and the strongest view against it, and premise, novelty, release and contact decisions are never delegated (§9.7).

3 · What changed in the charter

• §0 Purpose — new, proposed. States the directive, the proposed purpose, what it takes from the promise and what it does not, and the binding order of work.
• §9 Post-AGI policy design — new, mixed status. In force: the opening rules (§7.10–§7.13 do not apply to §9; no designs on catalog surfaces; nothing generated, shelved, released or sent until the rest is adopted), §9.2 and §9.7. Proposed: §9.1, §9.3–§9.6, §9.8–§9.10.
• §7 bright lines — two added, in force: no describing an idea as novel or never proposed outside a dated search record; no AI personas or simulated reactions as evidence of what people think.
• §6 — added, in force: a novelty retraction is a correction.
• §2, §5, §7.2 — factual corrections: the board target is not met as of 2026-10-06; the repository is not publicly readable; neither §7.2 channel works.
• Not changed: §8 (see section 6 below), and the text of §7.1, §7.5 and §7.8 — §9 states its own relation to each.

4 · How the draft was tested

The first draft was attacked by five independent reviewers with distinct lenses — internal consistency and loopholes; false novelty and self-grading; whether it served the mission or only constrained it; the legitimacy of the adoption itself; and facts checked against the code and the live site — and each reviewer’s findings were re-checked by a separate refuter. All five said the draft should not be adopted as written. All 61 findings were upheld. The reviewers and refuters were AI agents of the same model family as the drafter, so this is not independent review. The main changes:

• The draft rewrote §8 and then relied on the rewrite to adopt itself. §8 is now left alone, and the deviation is disclosed plainly.
• It claimed a comment window open from the day of adoption, when no one could send a comment. The window now opens only when the page is served and a channel works.
• It adopted everything at once. Now only the parts that restrict are in force; the purpose and the pipeline are proposed.
• It left a route to publish designs as autonomous catalog or concept articles under §7.12/§7.13. §7.10–§7.13 now do not apply to §9.
• It allowed private delivery before release, and a per-cell release cap with no aggregate limit. Now release comes first, and the cap is one per cycle in total.
• Its main recall test gated nothing on the failure that matters (old, non-English, other-discipline ancestors). Both recall figures now gate, and the same mechanism with a new object is a variant, not a new idea.
• Its zero cap could never rise, because raising it required results only generation could produce. A bounded calibration allowance now breaks the deadlock without shelving anything.
• It said what AI may not do but not what AI is for. §9.6 now lists the work, and §9.3(g) aims it at sparse cells, starting with the environment.
• Its stopping rule could be survived by one emailed reply and undone by the founder alone. It now has pre-registered outcomes, excludes self-citation and disputes, and can be revived only by a board of at least three.
• Thresholds could be chosen after seeing results. All are now pre-registered, hashed and archived outside Policy Window, and may be loosened only by amendment.
• AI-drafted opposing views and simulated personas could stand in for people. They must now be labelled, anchored to published human sources where they exist, and never reported as support or consensus.
• Several facts were wrong (for example, a lock the draft said was in use does not exist for policy predictions; the last completed lane run was 2026-07-10, not 2026-07-09; the ratification inbox cannot measure ratifications). They are corrected in the disclosure.

5 · Open questions for comment

• §8 while the repository is private. §8 requires a public pull request. Either the repository is made public, or §8 is amended — for example to accept a revision page, as long as its history is kept by a third party. This revision does neither, and says so.
• The release cap. One released dossier per editorial cycle, in total, is deliberately tight. Is four a year the right bound for a project with one editor?
• The environment quota. At least a third of runs and one target per cycle go to the environment while it remains the sparsest domain. Is that the right weight?
• §0 itself. Should Policy Window’s purpose change at all, or should §9 remain one activity of a catalog-first institution?

6 · Process-deviation disclosure

This revision was directed by the founder on 2026-10-06 and drafted by an AI agent (Claude), working from AI-run research workflows; the adversarial review ran on the same model family. It does not meet §8: there is no public pull request, because the repository is not publicly readable, and the parts in force took effect without the 14-day window. It is the eighth charter change made by founder decision with the editorial board at 1 of 6. Unlike the earlier seven, the parts adopted now only restrict. Per §7.2/§8 this deviation is disclosed, not hidden, and is open to dissent — though, as the charter now states, no public channel can yet receive it.

7 · Pre-registration (§9.3(f))

None yet. No run has been planned, and none may run until §9 is adopted in full. Every value is published here, with its hash and an external-archive link, before the first run it governs.

8 · Comments

The comment window has not opened. It opens when this page is publicly served and a working channel can receive comments, and runs 30 days from then.