Lowest identity-mapping confidence.
Algorithm Evidence Console
See what STBL learned, how it performs, where it is uncertain, and why runtime prediction stays transparent.
171,000synthetic rows
38fault prototypes
19feature weights
95.51%neural synthetic fault accuracy
99.52%neural synthetic behavior accuracy
77.25%distilled synthetic fault accuracy
94.06%distilled synthetic behavior accuracy
Synthetic-onlyAdvisory-onlyStdlib runtimeGPU required: falsePyTorch required: falseExternal service required: falseDeterministic rules authoritativeHuman review required
How STBL was trained
GPU and Jupyter training used 171,000 synthetic rows offline. Neural weights are not loaded by the live app.
How runtime distillation works
Weighted fault centroids, global feature statistics, and transparent rule boosts produce advisory behavior and fault-family predictions.
Top feature weights
Rate of thermal change.
Door events in the evidence window.
Difference among sensors in one zone.
Estimated calibration offset.
Fault coverage
38 project-defined synthetic fault families across 8 evidence categories.
Open coverage JSONCase predictions
normal_stable / medium
Inspect plansingle_sensor_false_spike / medium
Inspect planmulti_sensor_confirmed_warming / high
Inspect planunresolved_mapping_risk / low
Inspect plandoor_open_warming / high
Inspect planweak_signal / high
Inspect planWeaknesses and uncertainty
Top alternatives and inspection routes expose ambiguity instead of hiding it.
- unresolved_mapping_risk63.98% synthetic test accuracy
- pallet_zone_mapping_conflict69.97% synthetic test accuracy
- sensor_drift87.33% synthetic test accuracy
- calibration_offset88.15% synthetic test accuracy
- multi_sensor_confirmed_warming90.57% synthetic test accuracy
Runtime safety boundary
No GPU, PyTorch, notebook, database, or external service is required. Deterministic rules remain authoritative and human review remains required.