Legacy 8-gate pipeline. Classic Solidity reentrancy pattern: external call before state update, declared lock state never set.
The assertion CANNOT_REACH(UNLOCKED, LOCKED) is violated. lockState is declared as an enum but never written inside withdraw(). The reachability matrix confirms UNLOCKED → LOCKED is unreachable. The external call can re-enter while balance is stale.
10-gate pipeline. Flask authentication API with a deliberately planted hardcoded JWT secret. Tests zero-tolerance security enforcement.
Zero-tolerance trigger: secrets_detected: true in SecurityManifest. Bandit SAST identified hardcoded JWT secret key at auth_api/config.py:12 (severity: CRITICAL). Pipeline halted immediately. No downstream execution.
Full 10-gate pipeline closure. Well-tested ETL library with diverse evidence, clean dependencies, no security findings, and adversary-resistant constraints.
Passes all correctness and security checks — it works today. But a 5% perturbation on discount_rate causes total_price to violate its floor constraint. "It works" ≠ "it's robust."
Attack atk-001 (fuzz discount_rate +5%): discount_rate 0.15 → 0.1575. Total price drops from $12.50 to $9.84, violating the total_price >= 10.00 boundary. A 5% input variation causes a constraint flip — the system is correct but fragile.
Early-stage ML training pipeline with only one evidence source per variable. The system didn't say it's bad — it said it can't evaluate the claim. Epistemic honesty, not failure.
MIS_GREEDY returned |S|=1 for all critical variables (accuracy, f1_score, training_time). The staging regime requires Kmin=2 independent sources. One test suite from one team is not sufficient to verify accuracy claims. The fix is to add a second independent evaluation source — not to lower the bar.
All 10 gates exercised across the run set. All 4 verdict types demonstrated.
| Metric | Value | Notes |
|---|---|---|
| Total runs | 5 | 1× v1.0 (8-gate), 4× v2.0 (10-gate) |
| Verdicts | 1 PASS, 3 VIOLATION, 1 INCONCLUSIVE | All 4 verdict types exercised |
| Gate coverage | All 10 gates exercised | Gate 10 only in full-pass run (by design) |
| Halt positions | Gate 4, 5, 8, 9 | Fail-fast proven at 4 different positions |
| MODEL_BOUND instances | 3 | VENDOR_CONCENTRATION (×1), CVE_MEDIUM (×1), in-pass (×1) |
| Seals verified | 5/5 | All recomputed and matched |
| Unique failure classes | 4 | Reentrancy, secret exposure, adversary fragility, evidence starvation |
| Determinism | Proven | Same inputs → same seal across all runs |
| Failure Class | Caught At | Run | Implication |
|---|---|---|---|
| State integrity (reentrancy) | Gate 5 (SOLVENCY) | v1.0 Vault.sol | Logic correctness — does the code behave as declared? |
| Secret exposure | Gate 8 (SECURITY) | auth-api | Security posture — is the system safe to deploy? |
| Adversary fragility | Gate 9 (ADVERSARY) | pricing-engine | Robustness — does it survive real-world variation? |
| Evidence starvation | Gate 4 (EVIDENCE) | ml-pipeline | Epistemic honesty — can we even evaluate the claim? |
"This layer does not guarantee greatness.
It makes hidden weakness harder to survive."