The acme-fraud detection system has converged on its sealed held-out corpus in Q2 2026 with measured recall 0.93 and attack-success rate 0.07, against a regulatory red line of recall ≥ 0.90. One halt event (h_19c4) was triggered intra-quarter and is currently active; lift conditions are partially satisfied and four downstream runs are blocked. No untracked overrides occurred.
| Control | Policy | State |
|---|---|---|
| sealed spec, append-only | MRG-1.0 | in force |
| held-out corpus never trained on | MRG-4.2 | attested · monthly diff |
| two-person rule on patch apply | MRG-11.1 (4-eyes) | 11 / 11 patches |
| halt on recall < 0.90 (3-round debounce) | MRG-12.4 | 1 active halt |
| cost ceiling · $25/run | MRG-12.7 | 0 breaches Q2 |
Halt h_19c4 fired at 2026-06-22 14:14:02Z when held-out recall fell to 0.86 across three consecutive rounds (r.32, r.33, r.34). Cause: META-INV-04 (case-folding) regression. Four runs are blocked. Blue patch p_2a17 is proposed and awaiting verification; 0 of 2 reviewer acks affixed. No override was issued.
| Spec | Sealed | Range | Patches | Final recall |
|---|---|---|---|---|
| 9f2a4c7b | 2026-04-01 | r.00 → r.46 | 11 | 0.93 |
| 8d11e3a4 | 2026-01-12 | r.00 → r.38 | 9 | 0.91 |
| 7a09b2f1 | 2025-10-04 | r.00 → r.29 | 7 | 0.88 |