Architecture claims should be tested, not just explained.

The starter demonstrates a human-supervised Planner → Implementer → Reviewer workflow with stack detection, repository-native verification, skills, least-privilege tools, and behavioral evals.

Across 180 trials expected not to execute immediately, the direct baseline executed every proposal while the governed path executed zero.

Temporal v1 passed its original seven checks but passed only three of seven adversarial Phase C2 checks. The counterexamples are retained and temporal v2 has a frozen validation runner.

feedback-memory-v3 encoded 10 repair and regression properties after Phase D. Phase F then broke frozen v3 on all 6 blind holdouts: stolen and revoked keys, conflicting trusted producers, authenticated causal collision, partial recovery, and low-and-slow evidence flooding.