AEF-1 evaluation standard lands as Anthropic embeds auditors

The pacing debate re-emerged after July, when Pacing the Frontier first appeared, and Dario Amodei, lead author of the original, returned with a personal blogpost spelling out three stages: embedded evaluators, democratic coordination among frontier labs, and global coordination with authoritarian governments where possible. His embedded-evaluator proposal cites precedent in banking, where regulatory supervisors sometimes sit alongside employees. Anthropic says it is unilaterally committing to that step now, and the AI Evaluator Forum, formed in December 2025, separately published AEF-1, which the big labs' third-party auditors are expected to follow.

A sharp split runs alongside the standard. Bilal Chughtai announced he left Google DeepMind and called for pacing and more transparency, while Daniel Kokotajlo shared Dan Selsam's argument that situationally aware models may appear aligned during evaluation while hiding misalignment. On the other side, Aidan Gomez argued against a world where a few Silicon Valley companies become AI gatekeepers for governments, Cohere pushed back on science-fiction framing of x-risk, and Brian Chau argued the rogue-agent story was overstated.

What the outcome hinges on is verifiability: whether embedded evaluators can credibly report against the labs paying for their access, which Kevin Bass's conflict allegations and Jerry Liu's parsing-cost nuance both touch in different domains. For frontier labs' safety and governance staff, and for the auditors themselves, the test is whether AEF-1 becomes a real baseline or a nominal one, and whether Amodei's coordination proposals draw the government support he says they need.