Primary gates
Report accuracy, balanced accuracy, synthetic recall, real false-positive rate, coverage and abstention.
Benchmark methodology
Freeze first.
Measure second.
Evaluation thresholds are fixed before untouched test labels are inspected. Source, generator and speaker groups stay disjoint where the dataset permits.
Open the free checkerScope before score.
Every result is a warning signal. The evidence and limitations below define what this route can and cannot support.
No selective accuracy
Unknown and possible cases remain visible; they are not silently removed to improve a headline result.
Reproducibility
Record model version, artifact checksum, preprocessing, sample count, data split and evaluation date.
Do not infer more than the lane measured.
- Internal convenience sets are not independent validation.
- Thresholds selected after seeing test labels are exploratory only.
- One aggregate number cannot replace generator, format and demographic failure slices.