Anthropic and OpenAI extend desks, badges and laptops to third-party evaluators for ongoing model checks, raising regulatory capture questions ahead of IPOs.