Quantum calibration AI agents need every action tied to a measurable signal, a pass condition, and an escalation path. MIT's case shows the contract.
Anthropic's conflict tests show why multi-agent coordination needs mandate invariants, escalation rules, and independent acceptance checks.