mmune-bench
Open benchmark for runtime failure modes
Tool-agnostic scores against the same injection ground truth. Undetected modes publish as explicit zeros — never omitted rows. Marketing copy matches the table.
mmune baseline
Local reference-estate run-all → export-mmune (empty detection_events; optional mmune overlay image not wired for this publish) → score. Honest zeros: no live mmune detections were collected. Modes with mmune_detects_today=true (02, 09) remain 0/1 until a release re-run with the overlay. Modes 07/10 are opt-in (MMUNE_AUTO_DUPLICATE_SCAN), not default-deploy detection.
tool `mmune` `0.4.0+git.b1373fd` · hash `sha256:6e945f86caa8f39edf82c0b53661d528c04e12ed58a26ff7c1714be44425bd11`
| Mode | Labels | TP | FP | Recall | Precision |
|---|---|---|---|---|---|
| 01Semantic Value Mutation | 1 | 0 | 0 | 0.0% | — |
| 02Silent Field Drop | 1 | 0 | 0 | 0.0% | — |
| 03Data Type Mismatch | 1 | 0 | 0 | 0.0% | — |
| 04Ingress Path Topology | 1 | 0 | 0 | 0.0% | — |
| 05Payload Structure Corruption | 1 | 0 | 0 | 0.0% | — |
| 06Cross-System State Desync | 1 | 0 | 0 | 0.0% | — |
| 07Historical Data Injection | 1 | 0 | 0 | 0.0% | — |
| 08Silent Process Death | 1 | 0 | 0 | 0.0% | — |
| 09Ingestion Throttle Failure | 1 | 0 | 0 | 0.0% | — |
| 10Ingress Message Duplication | 1 | 0 | 0 | 0.0% | — |
| overall | 10 | 0 | 0 | 0.0% | — |
Snapshot imported at build time from website/data/mmune-bench-baseline.json— no live fetch from GitHub. Refresh on each mmune release tag (charter §7). Submit third-party scores via PR to the public repo. Modes 07 and 10 (duplicate dimension) are marked opt-in (MMUNE_AUTO_DUPLICATE_SCAN) in the public manifest — not default-deploy detection.
Also see the architecture overview and the technical whitepaper §04.