Choice
Accuracy, balanced accuracy, macro/micro F1, per-class precision and recall, confusion matrix, ECE, and coverage-risk.
Benchmark methodology
Every metric should answer a decision the team can act on.
The methodology separates data quality, task metrics, calibration, threshold coverage, row errors, runtime, usage, and report limitations.
The report keeps method and evidence together.
Noul
Accuracy, precision, recall, F1, Brier, ECE, ROC-AUC/PR-AUC when valid, two-sided certainty, and threshold sweep.
Truthful status
Missing usage remains UNKNOWN. Incomplete work remains PARTIAL. Insufficient data never becomes a production threshold.
Browse the full method