Evaluate reported machine conditions
Compare adapter predictions with independently reviewed labels. Every comparison matches a named source, asset and capture timestamp. Files are processed locally in this browser.
No evaluation loaded.
Fault recall is the share of reviewed faults detected. Fault precision is the share of fault alerts that match reviewed faults. Missing predictions reduce coverage and can count as missed faults.
Fault detection
Source coverage
Overall scores can hide weak cameras. Each row keeps its own sample count and fault examples; no fault examples means recall is not measurable.
| Source | Scored | Accuracy | Coverage | Faults detected | Missed faults | False alarms | Missing / unmatched |
|---|
Confusion matrix
Rows are reviewed states; columns are predictions. Unknown predictions remain visible as abstentions.
Input format and evidence requirements
Both JSON files use schema_version: 1 and provenance: "site" or "synthetic". The reference contains samples; the prediction file contains predictions. Each row needs source_id, asset_id, captured_at (RFC3339 with timezone), and state.
States: running, idle, setup, blocked, starved, faulted, maintenance, off, unknown. Duplicate identities are rejected. Review uncertain footage as unknown; do not manufacture labels to fill gaps. Include ordinary operation as well as failures to measure false alarms. Download the example for the exact structure.
A submitted site label is an assertion, not a certification. A high score on a selected sample does not establish performance across an entire factory.