Documentation
Model card
Witness classifies photographs of returned rotating-equipment parts into ISO failure modes. It is a two-stage system, and neither stage is a single trained checkpoint. This card documents both.
SHEET
MC-01
Two-stage defect classifier
REV B
SCALE NTS
Model details
- System
- Witness two-stage defect classifier
- Stage 1
- PatchCore-style memory-bank anomaly detector over classical patch features (96px greyscale, 8×8 grid, 5-dim descriptors, greedy-coreset memory bank, max-patch nearest-neighbour distance). Runs as a stateless Python function.
- Stage 2
- ISO forced-choice classifier. Default: deterministic feature heuristic. Optional: a vision-language model via OpenRouter (configurable), constrained to return a code from the part family's ISO catalog.
- Standards
- ISO 15243 (rolling bearings), ISO 10825 (gear teeth).
- Not used
- EfficientAD (excluded on MVTec patent grounds); deep timm backbones (do not fit serverless size limits).
Intended use
- Primary
- Triage and standards-linked documentation of returned-part damage, per tenant, behind human review.
- Users
- Reliability and warranty inspectors who confirm or correct every auto-suggestion.
- Out of scope
- Autonomous accept/reject of parts without inspector review; any safety-of-life decision made on the model output alone.
Training data
- Reference sets
- Each tenant enrolls its own photographs of known-good parts, per part family. Stage-1's memory bank is built only from that set, nothing else.
- Synthetic data
- None. No synthetic defects are generated or used.
- Bundled datasets
- The dataset registry admits only redistribution-clean licences; MVTec, Kolektor and similar are excluded or cite-only.
Metrics
- Stage-1 threshold
- Calibrated leave-one-image-out from the reference set's own scores, times a tunable margin. Reported per enrolment (bank size, reference-score spread).
- Stage-2 confidence
- Gated at 0.7: at or above auto-files, below routes to the review queue. The gate, not the model source, decides.
- Caveat
- These are operational calibration figures, not benchmark accuracy on a held-out defect set. This is a per-tenant system without one.
Ethical considerations & limitations
- Human authority
- The inspector's determination overrides the model and is what downstream analytics and warranty reports use.
- Provenance
- Model output is retained immutably alongside the human decision, so a disputed classification is auditable.
- Small reference sets
- Below ~20 reference images the normal-set is unstable; results are marked indicative and the human review load rises.
- Distribution shift
- Lighting or camera changes from the enrolled set inflate anomaly scores; retraining the reference set is the remedy.