Independent frontier evaluation becomes credible only when evaluators can inspect systems early enough and deeply enough to test real operating conditions. This matters now because model evaluations increasingly depend on tools, access modes, system prompts, and deployment conditions that a public endpoint may not expose.
Amodei argues for pacing frontier development while preserving the potential social and scientific benefits of advanced AI. He also calls for independent evaluators to receive access closer to the level needed for continuing investigation rather than one-off release checks. The mechanism is a standing access regime with scoped credentials, confidentiality rules, reproducible tests, protected publication rights, and escalation when findings conflict with a release plan.
A workable institution must protect model secrets without turning the evaluator into a privileged contractor whose incentives depend on the provider. The evidence is Amodei's policy argument, not proof that a specific governance design has already achieved independence or prevented frontier risk. Watch for concrete access contracts, funding structures, disclosure rules, and examples where evaluators changed a release or deployment decision.