A stateful MCP control layer that forces AI agents to define a verifiable win condition, map reality, name non-negotiable constraints, test a manipulable lever, record evidence, lock a decision, verify results, then rollback or double down. Hard gates reject skipped steps with errors.