Fallback to a trusted model

Reference classes
OpenAI current estimate (inferred) 2/10 (0.50)
Anthropic current estimate (inferred) 1/10 (0.50)
Resolution
Source: OpenAI & Anthropic system cards, Preparedness / RSP updates, or independent audits
Includes: Public primary-source confirmation that Fallback to a trusted model is operating across frontier deployments. Runtime/training safeguard on Internal deployments (robustness r = 60%).
Excludes: Aspirational roadmap commitments, offline toy-model pilots, or unverified third-party claims.