Independent embedded evaluators at frontier AI labs
Loading…
For each laboratory and date, will there be clear public evidence that the laboratory has put its September 2026 commitment to independent embedded evaluators into substantive practice?
Resolve YES when a new or materially expanded, continuing engagement is operating: independent evaluators have meaningful access to internal systems, staff and AI development or safety practices, and meaningful independence to assess and communicate their findings broadly in line with the laboratory’s commitment. The substance of the arrangement matters; there is no minimum number of days, requirement to publish the contract, or checklist of office equipment.
A future promise, an ordinary external model evaluation or a prior bounded exercise alone does not count. Credible confirmation from the laboratory, evaluator, published findings or reliable reporting is sufficient; both parties need not issue announcements. Resolve NO for a date if qualifying public evidence has not appeared by its end, even if later evidence reveals an earlier private start. A qualifying engagement subsequently ending does not undo earlier YES outcomes.
The two categories are Anthropic and OpenAI. The reference commitments are Dario Amodei’s September 2026 proposal and Sam Altman’s endorsement. Narrow confidentiality or security redactions are compatible with the commitment; routine company control over unfavorable conclusions is not.
Bars include events up to their close, 23:59:59 UTC: “By end of Sep 2026” includes everything up to 30 Sep 2026, 23:59:59 UTC.
The starting probabilities of this market were calculated from the work needed to turn the two companies’ commitments into an ongoing evaluator programme: choose a team, extend its access, agree reporting rights and begin work.
Both labs have already hosted evaluators inside their organisations. METR’s Anthropic exercise gave a researcher internal systems access and staff support for three weeks. Its OpenAI investigation involved several visits, interviews and access to internal records. These relationships make an expansion of existing arrangements a plausible route.
The main uncertainty is how much work remains on staffing, access and the right to communicate findings. The estimates use three possibilities:
| Route | Weight | Timing from the commitment |
|---|---|---|
| Extend an existing relationship; most arrangements are ready | 55% | Typically around six weeks |
| Substantial work on staffing, permissions or reporting terms | 35% | Typically around five months, with a wide range |
| The commitment stalls or settles into an arrangement that falls short | 10% | Remains unresolved |
These weights and times are judgments based on the work involved. Public accounts of earlier exercises give useful evidence about access and feasibility, but leave the time spent negotiating arrangements unclear. The model allows considerable variation around each timeline.
Anthropic receives a timing advantage because it developed the proposal and described concrete access and reporting arrangements. That suggests more preparation before the announcement. OpenAI endorsed it the same day and can build on its own evaluator relationships.
A study of voluntary environmental agreements found more substantive follow-through among firms involved in establishing the programme than among later joiners. The gap between joining dates was much larger than here. It supports giving preparation some weight; the size of the adjustment remains a judgment.
The central assumption is that OpenAI takes about 1.5 times as long as Anthropic. The uncertainty around that assumption includes a meaningful chance that OpenAI moves faster. Their probability of eventual implementation is the same. The chart also shows how the estimates change with easier or harder negotiations.

Calculation code · Sources and their use · Model assumptions and results
Liquidity over time
i Shows the market’s available liquidity over time. Liquidity can decay continuously and may change discretely if liquidity adjustments are applied.2 comments
Log in to comment.
METR published a summary on October 22 of a three-week exercise (July 21 – August 11) in which one METR and one Redwood Research employee stress-tested OpenAI's internal misalignment monitoring system. They worked on OpenAI-issued devices on site, with access to OpenAI's internal codebase, monitor prompts and configuration, and internal documentation, and interviewed about a dozen researchers and engineers across the monitoring, safety-systems and security teams. The result was a 61-page internal report.
METR describes it as a pilot for giving external researchers deep access to internal infrastructure. OpenAI could redact anything it judged confidential or sensitive from the public post, at its sole discretion, and METR had no unilateral right to describe those redactions. METR says that for "future versions of this exercise" it expects to follow the transparency and independence practices in its guidance on working with AI companies.
Findings include: an estimated 75–94% of internal agentic Responses API traffic from models at least as strong as GPT-5.5 was eligible for monitoring; RL training runs were not monitored at the start of the exercise; and two client-controllable headers could disable monitoring for new agents. OpenAI says every vulnerability affecting its production safeguards has since been patched. METR has not evaluated that claim. — METR, October 22
OpenAI's October 29 update says most vulnerabilities identified by METR and Redwood's July pilot have been patched, with one fix still in progress. OpenAI