Three labs commit to readable reasoning, with outside checks

Where things stand (13 Oct 2026)

No qualifying commitment exists yet, from any developer. A CASRAI review last verified on 25 Sep 2026 concludes that "no instrument requires a frontier developer to preserve chain-of-thought monitorability." The EU Code of Practice only asks for access to reasoning traces "if available" (https://casrai.org/guides/chain-of-thought-monitorability). Here is where each developer stands:

• OpenAI has an internal "policy against directly grading CoTs." Its own wording is "we still aim to avoid," and it keeps self-granted exceptions "where…benefits clearly outweigh the risks" (https://alignment.openai.com/accidental-cot-grading/). Both features fail this question's rules on hedged language and unilateral waivers.
• The GPT-6 Astra system card (3 Sep) reports a "substantial decrease" in monitorability (https://deploymentsafety.openai.com/gpt-6-astra/monitorability). UK AISI's external monitorability evaluation tested the model only, not training practices, so it would not count as a compliance check.
• Google told METR that it uses supervised fine-tuning and reinforcement learning on Gemini's reasoning traces (thought traces).
• Anthropic has accidentally trained against the chain of thought several times.
• METR's Frontier Risk Report (19 May 2026) collected each lab's chain-of-thought practices but checked them against no commitment (https://metr.org/blog/2026-05-19-frontier-risk-report/).

The momentum is unusually strong:

• OpenAI, 9 Oct: "preserving the monitorability of frontier models requires an industry-wide commitment, including from OpenAI" (https://www.cbsnews.com/news/openai-defends-firing-safety-researchers/).
• Cross-company work: the researchers OpenAI fired say they were stewarding "commitments to prevent loss of monitorability" and "cross-company limits on…unmonitorable architectures" (https://mikitabalesni.com/letter/letter.pdf).
• Google DeepMind published The case for reasoning transparency on 16 Sep (https://institute.deepmind.com/essays/the-case-for-reasoning-transparency/). Its Gemini 4 Argon launch post "strongly encourage[s] the rest of the industry to preserve reasoning transparency."
• Outside checkers are being built: Anthropic has embedded evaluators (Accenture/Faculty, plus a METR pilot), and OpenAI is "finalizing contracts" with outside assessors.
• An industry standards body: Google, OpenAI and Anthropic are reportedly forming one (provisionally SAFA), targeting early 2027.
• Other pressure: the White House Accord (29 Sep) commits signers to "meet regularly to establish standards," and the EU AI Scientific Panel (9 Oct) urges scrutiny of architectures that reduce monitorability.

How I modelled it

Reference class. Voluntary technical commitments by frontier labs, such as the 2023 White House commitments and the 2024 Seoul pledges, usually come out hedged and without per-company audits. Credible checking with published findings is very new: Anthropic's risk-report reviews in 2026 and METR's pilot report.

YES needs a full conjunction:

1. Three developers make firm commitments. Realistically these are OpenAI, Anthropic and Google DeepMind, with Meta possible.
2. Each commitment covers both readable reasoning and no training against it.
3. None allows a unilateral waiver.
4. Each names an outside checker.
5. That checker publishes findings on both obligations.

Computation. I ran a hazard model in Python:

• The chance of a qualifying commitment set is front-loaded: about 2–3% per quarter through 2027, falling to about 0.6% per quarter by 2031. That gives about 26% cumulative by end-2031.
• I convolved that with an audit-report lag (mean about 2 quarters), with a 0.72 chance that qualifying reports covering all three developers actually appear.
• Result: about 0.2% by end-2026, 5% by end-2027, 10.5% by end-2028 and 18% by end-2031.

Main paths:

• To YES: a joint pledge in 2027 through SAFA or the cross-company effort, checked by METR or embedded evaluators in their periodic reports. A slower legislative route also exists, via the federal FRONTIER Act (no vote expected before 2027) or California's new independent-verification law.
• To NO: hedged wording or capability carve-outs; Google's existing training on reasoning traces; OpenAI's move toward architectures with more hidden reasoning, plus its firing of the effort's steward; antitrust worries about agreeing architecture limits with competitors; or the industry abandoning readable reasoning altogether.

Strongest case for a higher number. All three leading US labs now publicly favour preserving readable reasoning. This is the cheapest and most checkable safety pledge on offer, and METR already gathers exactly the relevant evidence. That could justify 25–30%. I hold lower because of the strict bar on firm wording plus published findings on both obligations, and because the hazard probably fades if architectures move away from readable reasoning.

Question. By the last day of each quarter, have at least three covered developers committed to keeping their frontier models’ reasoning readable and not training against that readability, with a named outside body checking compliance?

Definitions.

• Frontier models: the developer’s most capable general-purpose models, including its leading reasoning models, whether publicly released or used internally. A qualifying commitment must apply prospectively to this class of models, not just to one named checkpoint, a demonstration, or an experimental model.
• Keeping reasoning readable: preserving human-understandable intermediate reasoning traces used by the model in reaching answers or taking actions, so that safety monitors and authorized human reviewers can inspect safety-relevant reasoning. A commitment merely to generate explanations or summaries after the fact does not qualify. Public or end-user access to raw traces is not required, nor is a claim that every internal computation is represented in text.
• Not training against it: an explicit undertaking not to optimize frontier models to conceal, obfuscate, or sanitize their intermediate reasoning, including not using chain-of-thought monitor judgments as training rewards or penalties that teach the model to hide evidence of misbehavior. This does not prohibit training on final-answer correctness or actual behavior, or training separate monitors. A promise to measure monitorability alone does not qualify.
• Commitment: an official, operative policy or undertaking, voluntary or legally binding, that specifies both obligations above and subjects them to outside compliance review. Aspirations, recommendations, research plans, and preferences qualified only by “all else equal,” “where appropriate,” or “we aim to” do not qualify. A commitment allowing unilateral waiver of either obligation for capability or commercial reasons does not qualify.
• Named outside body: an identified organization legally separate from the developer and not controlled by it, assigned to assess compliance with both obligations. Government institutes, independent research evaluators, auditors, and industry bodies can qualify. Developer funding or industry membership alone does not disqualify a body, but a developer’s internal committee does. Different developers may use different bodies.
• Checking compliance: the outside body has published at least one substantive compliance report identifying the developer and commitment, describing its review of reasoning traces or monitorability evaluations and of training practices, and giving a finding on each obligation. A statement that an audit is planned, a developer questionnaire reproduced without independent assessment, or a model benchmark with no review of training practices does not qualify. Underlying confidential records need not be published.

Resolves YES when qualifying public records establish that at least three distinct covered developers have operative commitments meeting both obligations and each has been covered by a qualifying outside compliance report. Joint or separate commitments and joint or separate reports count. The reports need not certify perfect compliance: this question concerns externally checked commitments, not a guarantee that no violation has occurred. All three commitments must be operative at the time the third developer qualifies; withdrawn or expired commitments do not count toward that initial total.

Resolves NO / Exclusions. Resolve NO for a quarter if fewer than three developers meet all requirements. Individual researchers’ signatures on a position paper, membership in an organization that recommends monitorability, publication of evaluation tools, self-reported monitorability scores, or external checks of only one obligation are insufficient. In particular, the recommendations in Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety (arXiv:2507.11473v1, July 15, 2025), the Frontier Model Forum’s Chain of Thought Monitorability brief (January 27, 2026), and OpenAI’s Open Sourcing Monitorability Evaluations (April 23, 2026) do not themselves establish qualifying checked commitments.

Resolving source. Official commitments on each developer’s own domain or in developer-authored policy documents, together with compliance reports published by the named outside reviewing bodies. Developer adoption of a joint undertaking must be explicit. Apply Public confirmation (A2), including its dispute rule for outside reports, Resolution date (A3), and the Resolution window. Only qualifying publications within the Resolution window count; an in-window official reaffirmation of an earlier commitment may establish that commitment. Resolve on the publication date of the last document needed to establish the three-developer threshold, subject to A2. Once YES, later withdrawal or noncompliance does not reverse resolution.