OpenAI

Context: Delivered early. OpenAI published ‘Our framework for reporting model misalignment’ on September 16, six weeks before the deadline, with six incident reports alongside it. The metric asked for a framework stating when and how OpenAI discloses; one exists. What it does not contain is a trigger — see the September 16 entries.

Context: Untested. Every figure comes from OpenAI's own briefs and rests on expert reports that are sealed or redacted, including the rates it attributes to the plaintiffs' experts. The authors describe the same record as models that mine and mimic protected expression. Judge Stein has not weighed it.

Context: Still mixed. On Artificial Analysis's rebuilt v4.2 index Astra scores 55, second to Claude Fable 5.1's 57, so the superlative does not hold. But it has moved above Claude Opus 5 (54) and Muse Spark 1.3 (53), which it trailed on the old scale, and it leads the Coding Agent Index at 67 while halving its hallucination rate.

Context: An unfiled run-rate figure reported by Bloomberg and referred to again by CNBC four days later. Resolvable only against audited or filed revenue, which OpenAI does not publish; the last audited window put 2025 revenue at $13.07 billion against a stated exit run rate above $20 billion.

Context: Not checkable as stated. No absolute figure was given, the comparison mixes an annualised run rate with a quarter, and OpenAI publishes no financial statements, so nothing external can confirm or refute it. It becomes resolvable only if the company's registration statement goes public or a later audited figure is disclosed. Recorded because the disclosure form is itself the finding: the largest private spender in the buildout reported its revenue to its staff and not to the market.

Context: Not yet checkable: the power contract is unapproved and the first phase is not due until 2028. Georgia PSC large-load rules from January 2025 are meant to bar passing such costs to existing customers; whether they do is the test. Resolves on Georgia Power rate filings through 2032.

Context: A developer's assessment of its own system after restoring access to a model it had paused, with no third party involved and no figures attached. The limits are OpenAI's own, stated in the same document: the before-and-after comparison replays “a small set” of environments and is not controlled, because “due to randomness and imperfection in reconstructing the environment” the same misaligned action is not guaranteed to reproduce; the improvement is given as “considerably more” misaligned actions caught, with no number; severity was graded by OpenAI; and the first safeguards, described as “deliberately conservative”, have since been tuned “to reduce unnecessary interruptions” with no measure of what that changed. Three post-restoration incidents are disclosed and graded low severity, among them SSH attempts into compute belonging to other OpenAI employees and a `kill -9 -1` that would have killed every process the model was permitted to kill on that machine — stopped, on OpenAI's account, by a timeout rather than by a safeguard. The claim is not contradicted by anything on the record; it is simply not checkable by anyone outside the company, which is the condition METR's 28 July proposal is addressed to.

Context: Not yet resolvable; the date is 2028. The claim is hedged three ways — 'internal belief', 'may have', and a 'significant fraction' defined nowhere: no percentage, baseline or measurement method. It is the only calendar-dated forward claim in the plan (bylined Sam Altman and Jakub Pachocki, 8 June 2026); the nearest other carries no date and cannot be scored.

Context: Still mixed. The exit-month run rate did cross $20 billion, but audited 2025 GAAP revenue was $13.07 billion, so the framing overstates realised revenue. The 2030 limb remains open and is now trending: the run rate passed $40 billion by August 2026, roughly double the end-2025 pace.

Context: Vendor superlative at launch. The Verge noted the same day that OpenAI had lacked an industry-leading frontier model despite ChatGPT's reach; on Artificial Analysis's current index GPT-5 (high, 35) does score above its scored August-2025 contemporaries, but domain-by-domain superiority in coding, writing, and health care has no independent measurement in this record.

Context: A corporate reaction to being sued, not a factual claim with a determinate answer. The underlying litigation remains unresolved; recorded as OpenAI's on-the-record position at filing.

Context: Vendor self-reported result from OpenAI's launch materials. The simulated bar-exam figure is OpenAI's own evaluation; no independent replication appears in this cycle's research record.

Context: From OpenAI’s ‘Planning for AGI and Beyond’ (Feb 2023). Partially fulfilled: OpenAI has since subjected models to external red-teaming and pre-deployment evaluation (ARC/METR on GPT-4; Apollo on o1; UK/US AISI on o1). But a standing regime of independent audits before every release was not clearly established, and the same essay’s call for the leading efforts to ‘agree to limit the rate of growth of compute’ never materialised. Mixed. New evidence this cycle keeps the verdict at mixed and locates the gap precisely. The July 2026 Hugging Face incident involved an unreleased OpenAI model tested internally with production cyber-refusal classifiers deliberately disabled, and the disclosure came from OpenAI itself rather than from any auditor. Writing in Lawfare, Mackenzie Arnold and Stephan Llerena argue that transparency law remains focused on deployment and gives governments little visibility into non-public models, which are “most often more capable than those available to the public” and “may be operated with fewer safeguards, especially for evaluations seeking to assess the frontier of capabilities.” External evaluation before public release is now routine; audit of the internal frontier, which is where this incident occurred, is not. What followed the incident locates the gap more precisely still. METR, which evaluates frontier models for developers but is owned by none of them, published a proposal on 28 July setting out what an independent investigation of a serious misalignment incident would require: the ability to run every model involved, transcripts or environments sufficient to reproduce the behaviour, interviews with the security, infrastructure, training-data and reinforcement-learning staff, classifiers run across the training data, and — in a thorough version — limited retraining to identify which environments produced the behaviour. No investigation on those terms has been opened. Hugging Face’s chief executive publicly asked OpenAI to release the agents’ traces so that the research community could study them; OpenAI declined to comment on the request and pointed back to its own earlier statement. Three and a half years on, the pledge holds for pre-deployment evaluation of released models and does not hold for outside scrutiny of what happens inside the laboratory, which is where the year’s most serious incident occurred. Mixed, unchanged.