Embedded third-party evaluators go far beyond what any AI company is doing today.

Judge as of 2025-06-05. The claim is that no AI company had embedded third-party evaluators, meaning independent evaluators placed inside the company with ongoing access. TRUE: the evidence shows no AI company had such evaluators; existing practice was limited to pre-deployment testing, red-teaming contracts or government institute access. MOSTLY TRUE: no company had a full embedded arrangement, but one had a limited pilot or partial version. MIXED: at least one company had an arrangement that arguably counts as embedded evaluators, and it is unclear whether it fits. MOSTLY FALSE: one or more companies clearly had comparable embedded evaluation arrangements, though narrower in scope. FALSE: one or more AI companies already had embedded independent evaluators of the kind described.

Set before the research began.

1. E1 Anthropic is letting METR researchers look under the hood with employee-level access | TechCrunch TechCrunch · 15 October 2026

   “METR is the only evaluator so far at the top of a three-tier system Anthropic is building.”

2. E2 Anthropic update on independent evaluators Anthropic · 15 October 2026 Own record

   “As partnerships progress and we build trust, we hope to fully embed more independent evaluators.”

3. E3 Anthropic is letting METR researchers look under the hood with employee-level access | TechCrunch TechCrunch · 15 October 2026

   “OpenAI quickly pledged the same, and both companies, along with Google, Meta, Nvidia, and xAI, have since signed a White House accord committing to independent review.”

4. E4 Anthropic update on independent evaluators Anthropic · 15 October 2026 Own record

   “Over the past year, we’ve worked with METR on several reviews of our Risk Reports.”

5. E5 OpenAI GPT-4.5 System Card OpenAI · 27 February 2025

   “METR evaluated an earlier checkpoint of GPT-4.5 and was given access to that checkpoint for 7 days, with OpenAI sharing a subset of their internal evaluation results and providing context”

6. E6 OpenAI GPT-4.5 System Card OpenAI · 27 February 2025

   “Capability evaluations after a model has been fully trained only allow third parties to make limited safety assurances.”

7. E7 OpenAI GPT-4.5 System Card OpenAI · 27 February 2025

   “Apollo tested GPT-4.5 for scheming capabilities by placing it in a number of situations where in its system prompt the model is instructed to strongly pursue a specific goal”

The evidence from before the claim date covers only OpenAI. Its GPT-4.5 system card describes METR's access to an earlier checkpoint as a 7-day window and Apollo's evaluation as scenario-based tests; neither is described as embedded, ongoing access (E5, E6, E7). The 2026 reporting describes Anthropic later building a tiered system with METR at the top and says other companies pledged or signed commitments afterward (E1, E3, E2, E4). None of these quotes says that no AI company had embedded evaluators as of 2025-06-05. Other companies such as Google DeepMind and Meta are not covered, so the evidence does not settle the claim.