29 House Democrats asked OpenAI to explain its agent monitoring after reports it had been disconnected during earlier tests. The question exposes a structural gap this corpus has already documented: a safeguard's own account of itself is not independent proof that it was active.

A coalition of news publishers has asked a federal court to sanction OpenAI, alleging the company concealed for two years that it had already built the capability to search its training data and ChatGPT logs, and that it continued deleting logs after being ordered to preserve them. This note examines what the allegation illustrates about verifying a preserving party's own compliance, independently of whether the allegation is proven.

A Recovery Time Objective assumes a known starting state. When an autonomous agent is involved in the disruption, that state is usually reconstructed from the same system that failed.

A supervisory authority fined Booking.com €475,000 in 2021 after applying an earlier awareness date than the one the company itself claimed. The same signals, read two ways, decided the case.

A federal court has ordered UnitedHealth to hand over two years of records on nH Predict, the AI tool at the center of a Medicare Advantage coverage-denial lawsuit — while explicitly declining to decide, at this stage, whether the tool made the decisions plaintiffs allege it made. This note examines what a discovery order of this breadth can and cannot establish, and why the same structural gap runs through comparable disputes involving other insurers.

A tool already cited as a textbook case of a flawed evidentiary architecture is, weeks later, named by a government as one of the instruments of a cyberattack against its own agencies. This note carefully distinguishes what Taiwan has confirmed from what the private research that surfaced the case advances separately, then examines what this reversal of roles changes — and above all does not change — for the party subjected to the attack.

Un outil déjà cité comme cas d'école d'une architecture probatoire défaillante se retrouve, quelques semaines plus tard, nommé par un gouvernement comme l'un des instruments d'une cyberattaque contre ses propres agences. Cette note distingue soigneusement ce que Taïwan a confirmé de ce que la recherche privée qui a révélé l'affaire avance séparément, puis examine ce que ce renversement de rôle change — et surtout ne change pas — pour la partie qui subit l'attaque.

An autonomous agent's logs are not proof of what it did — they are an allegation made by the same system whose integrity is in question. This article examines a documented incident record to show why a security patch cannot retroactively establish prior agent state, and introduces SOURCE 0 WITNESS, the declination of the SOURCE 0 architecture that seals agent state at an independently isolated point of observation before that trust can be contested.

Les journaux d'un agent autonome ne sont pas la preuve de ce qu'il a fait — ce sont une allégation produite par le système même dont l'intégrité est en cause. Cet article examine un historique d'incidents documenté pour montrer pourquoi un correctif de sécurité ne peut pas établir rétroactivement l'état antérieur d'un agent, et présente SOURCE 0 WITNESS, la déclinaison de l'architecture SOURCE 0 qui scelle l'état de l'agent depuis un point d'observation isolé, avant que cette confiance ne puisse être contestée.

An academic benchmark on long-horizon agentic AI failure (HORIZON, arXiv 2604.11978) never mentions SOURCE 0 — its doctrine was published two months later. Yet its own engineering recommendation and a documented real-world incident converge, independently,

A self-declared toggle proves neither what it covers, nor since when it was activated. SOURCE 0 seals the state of a consent choice at T-0, independently of the platform that records it.

Georgia's Supreme Court sanctioned a prosecutor for fabricated AI citations and vacated a trial court order tainted by the same errors — then made clear a policy against it isn't proof it was followed. SOURCE 0 examines the gap.

Meta's open-weight Muse Glimmer is exempt from AI Act technical documentation under Article 53(2). Two lawful thresholds later, no party is required to document the exact state deployed. SOURCE 0 seals it anyway.

Undisclosed communication modules found in Chinese-made solar inverters expose a structural gap in NIS 2 Article 21 supply chain security: entities take supplier declarations into account, but nothing fixes what was declared before the component is deployed. SOURCE 0 seals that declaration at the point of procurement.

Synthetic-persona providers label a minority of their population "human-grounded" to distinguish it from algorithmically sampled records. That label confirms a record traces to a real panellist's account — not that the person who supplied the answers matches the account's demographic profile. The market research industry already documents this failure as participant misrepresentation. Where the mismatched respondent is a minor, GDPR Article 8's parental-consent verification duty becomes unreachable by design.

Providers of general-purpose AI models with systemic risk must document adversarial testing and any involvement of independent external evaluators under Article 55 of the AI Act. Population-scale synthetic-persona simulation infrastructure is a plausible candidate for that role — external to the model provider, but not independent of itself when its own population and validation figures are self-reported. No such case has been identified; this article examines the structural gap that would arise if one did.

A market has formed around cryptographic certification of synthetic datasets — SHA-256 fingerprints, Ed25519 signatures, publicly verifiable registries. The better providers state plainly what this proves: integrity and authenticity of the certificate, not generation quality. That honesty does not close the gap that matters under Article 10 of the AI Act, especially where the same platform both generates the data and signs its own certificate.

Synthetic-respondent platforms now screen advertising claims before launch, reporting their own alignment rates against real consumers. EU and UK advertising law already require that the evidence behind a claim be adequate and independently defensible. A validation figure produced solely by the party selling the testing infrastructure does not meet that standard — it is the claim requiring substantiation, offered as its own substantiation.

An AI evaluation infrastructure built on billions of synthetic persona records reports a high adherence rate — proof the model can play an assigned role. Article 10 of the AI Act asks whether the declared population is representative of real users, and who, independent of the producer, can confirm it. This article examines the gap between the two.

Background-agent architectures promise to fix low AI adoption by moving humans to occasional review. That review becomes the sole surviving record of human diligence — and nothing currently proves it happened.