← Back to the chapter · Previous detail
Technical detail · IoT Machine Learning Fundamentals
A minimal inference contract
An inference contract makes four sets of evidence reviewable.
| Contract Item | Question | Evidence | Failure Mode |
|---|---|---|---|
| Input window | Which samples, timestamps, units, and context fields must be present? | Feature code, schema check, missing-data rule, and replay test. | The model scores partial or shifted data as if it were complete. |
| Model artifact | Which model version, threshold, scaler, and feature definition are deployed together? | Versioned artifact bundle, release note, checksum, and rollback path. | A new model runs with an old scaler or threshold. |
| Output handling | What should the system do on low confidence, missing inputs, stale readings, or unsupported context? | Fallback rule, manual-review path, alert suppression rule, and operator-facing status. | The system emits confident alerts when it should degrade or ask for review. |
| Feedback loop | How are outcomes, false alarms, misses, drift, and operator decisions fed back into review? | Monitoring dashboard, audit log, label queue, review cadence, and retraining gate. | The model keeps running after the deployment has moved outside its reviewed evidence. |