Four working concepts
Findings are expressed in terms the sterile processing floor already uses.
- Missing
- Extra / unexpected
- Substitution candidate
- Uncertain — needs review
Vision AI
Assistance on the tray. Authority stays human.
Vision is one capability of Steritrack, not the product. It proposes findings on a set photo; a qualified operator confirms, corrects or rejects them — and that human decision is the record of truth.
Human ground truth
Model output and human verification are stored as distinct records, never merged.
- Operator confirmation required
- Corrections captured explicitly
- No automatic release
Dataset and annotation
Capture campaigns, annotation queues and review flows are implemented in the platform.
- Capture campaigns
- Annotation queue
- Ground-truth review
Engine contract
A versioned inference contract keeps the platform independent from any single model.
- Model-agnostic interface
- Run and prediction history
- Contract conformance tests
Explainability
Each finding is traceable to the image, run and model version that produced it.
- Per-run predictions
- Image-level evidence
- Reviewable history
Honest maturity
Detection quality depends on datasets built with each site during evaluation.
- No clinical accuracy claim
- Site-specific calibration
- Assistive by design
Steritrack makes no claim of trained clinical production accuracy. Vision output is assistive and always subject to human verification.