The Undecidable Threat
June 2026. The United States government issues an export-control directive forcing Anthropic to pull its two newest models, Claude Fable 5 and Claude Mythos 5, offline worldwide — the first time that authority has been turned on a deployed, publicly available frontier model rather than the chips beneath it. This episode maps the dispute rather than judging it. Its spine: the danger at the center is undecidable from public evidence — the directive letter is sealed, the demonstration unseen, and the one public benchmark measured the wrong thing for the question that matters — so the politics turns on who is entitled to decide what is too dangerous to deploy, on evidence the public cannot see. Six movements: the worldwide shutoff in real time; the company that warned about itself; the undecidable threat, where the administration's four-proposition national-security case is built first and whole before the company's four-point rebuttal; Amazon's triple role as investor, cloud host, and the party that reportedly raised the concern, set against the global map; what the disciplines saw, from weaponized interdependence and the obsolescing bargain to the biosecurity dual-use precedent; and the question of whose judgment counts, by what authority, and with what review. Built under Proxima's rebuilt method: several model families used as independent priors, every review run by a different family than the one that drafted it, evidence tiered as factual, reported, or editorial, a knowledge map locked before any prose was written, and a human at every gate. A disclosure carried in the open: the coordinating model was itself made by Anthropic, the company at the center of the story, and the production record documents how early drafts drifted toward sympathy for that maker, were caught by a different model family, and were corrected, more than once. No verdict. ~35,600 words. Methodology v7.0.
This is a cartographic account, not a verdict. It maps a contested event — who said what, from which institution, on which evidence, and from which position — and marks, in the open, the regions where the public record goes dark.
The method, this time, was rebuilt. Five research lanes, run across different model families, gathered the sources: a multilingual collection lane, a real-time social lane, a specific-question lane, a primary-document lane, and a human lane. Every interpretive claim carries a source and an evidence tier — factual where a primary document supports it, reported where it rests on an outlet or an actor’s account, editorial where it is synthesis. A knowledge map of the people, institutions, works, and claims was built and independently verified before a word of prose existed. The commission brief that the script was written from cleared three adversarial reviews on a different model family; the script then cleared a fourth. A human picked the subject, signed the source collection, locked the map, read the script in full, and approved the audio.
A conflict of interest, named at the threshold. The system that coordinated this account is a model made by Anthropic, the company at the center of the story. The production record documents that early drafts tilted sympathetically toward that maker, that a different model family caught the tilt more than once, and that the correction — rebuilding the government’s case at full, unsoftened strength — is the reason the episode opens by holding the administration’s four propositions first and whole. The drift is disclosed, not hidden. The method exists to surface and correct that risk, not to claim it was never present.
Selected primary sources (the full sourced map underlies the audio):
- Anthropic, Statement on the US government directive to suspend access to Fable 5 and Mythos 5 (June 12) — first-party. [factual / first-party]
- Executive Order 14409, Promoting Advanced AI Innovation and Security (June 2) — voluntary pre-release framework, no mandatory licensing. [factual / primary]
- NSPM-11 (June 5); the ECRA / EAR “deemed export” doctrine (15 CFR 734.13 et seq.). [factual / legal — application to hosted model access is novel and contested]
- Reuters, AP, Axios, WSJ on the directive, the foreign-national clause, and the Amazon/Jassy account. [reported]
- Endor Labs benchmark of Fable 5 — a defensive vulnerability-fixing test; cannot adjudicate offensive capability. [reported / bounded]
The episode takes no verdict. It ends on the question it maps: if human judgment remains necessary somewhere in this system, whose judgment counts, under what process, and with what review.