Self-Verifying Agent Architecture Dissociates Commitment from Binding Drift

This architecture uses a deterministic Executive to own all beliefs, while a language model only submits typed proposals. A pre-registered prediction must match observations via code to prevent drift in long-horizon agentic tasks.