Physics-grounded deterministic verification layer for AI agents, robots, vehicles, drones — and the SDKs that drive them. Every claim ships with an on-disk artifact; every aggregate recomputes from open metric files.
Four layers, closing state 2026-10-09 — full ledger with verbatim metrics, audit logs, closing marks and SHA-256 manifest: phm-evidence-ledger
dpc-bench — paired clean/corrupted benchmark for physical-consistency defect detection (7 task families, ~10.2K records, three-caliber labels L1/L2/L3). Six synthetic families each modeled on a real, publicly documented defect (AgiBot-World #164, PaXini data-kit #3, GalaxeaLeRobotToolkit #5, NGSIM unit semantics, DJI log frame losses, UZH-FPV kinematic violations); one real-data family derived from NGSIM I-80 (USDOT public domain). Zero-false-positive-on-clean is reported separately. Open board — submissions welcome.
| Layer | Count | Unit | Headline result |
|---|---|---|---|
| Physics-engine gate (Master 2B) | 1,900,000,000 +100M baseline | program executions | FP=0 / FN=0 observed, 10 classes, REACHED_2B mark |
| Verifier validation scenes | 1,000,000,000 | scene executions | per-class n/ok recorded verbatim |
| Robotics stream gate | 1,000,086,001 | data frames | collision class: 43,440,642/43,482,000 actively BLOCKed |
| SCL fuzzing | 1,000,000,000 | program executions | no crash observed; 3,226,364,783 rule hits; 2.41 ms/program |
Caliber: three layers count executions, one counts frames — mixed by design, stated per line. Zero-valued results mean "not observed in this test set", not absolute guarantees.
- Dify Official Marketplace — PHM Physics Check merged into the official plugin registry (langgenius/dify-plugins#3445, 2026-10-22): deterministic physics checking for agent platforms. Zero network calls, verdicts recomputable by anyone.
- microsoft/PyRIT — under review (PR #3507, 2026-11): deterministic physics-consistency scorer built on this package. Six first-round correctness issues fixed same-day; PyRIT's creator then pushed four follow-up commits, verified 356 scoped tests, and approved ("no blocking concerns"). Merge pending separate maintainer decision — not yet merged.
- MCP official Registry —
io.github.PHM-Audit/phm-lite-check@1.0.1listed at registry.modelcontextprotocol.io (verified 2026-11-06); listing = metadata publication, not endorsement. - MedCheck (in development) — deterministic medical-advice boundary checker (dosage bounds, symptom-conclusion consistency, supervision-evasion detection) with three-state verdicts. Case replay on public litigation records: 9/9 internal, 6/8 cross-case, control-group zero false positives (v0.2.1, SHA-256-anchored artifact).
- Independent physical audits (open data) — all-clean passes over 2,241 Sentry objects, 381 active-spacecraft TLEs, and lunar-sample metadata (263 samples / 3,300 records); 5 Earth mini-moon objects flagged for NaN-semantics defects; C-MAPSS README-vs-paper metric mismatch documented (0.001% vs 73.6%).
Deterministic, operator-declared, pre-execution adjudication with an audit ledger — gate-first (benign FP=0), evidence-first (every number recomputable), closed engines with open metrics.
Scripts are closed-source core assets; metrics, logs, and gates are open.