Score
Calibration of raw evaluation scores into the score a session reports.
Raw scores from the segment scorer are algorithm output: they cluster high and thin out at the bottom, which reads unfairly to a learner. Calibration is the perceptual curve that turns algorithm output into the score users see — and per ADR-023 ("the 0-100 / 0-1 score from a documented curve"), the curve is VoxaTrace's to own and document, while banding the calibrated score into stars/colors/verdicts stays the client's.
One currency (ADR-026)
The session applies the configured calibration exactly once, at the scoring seam. Everything called "score" — per-segment results, the session score, and the SessionConfig.scoreThreshold the auto-advance policy compares against — is in calibrated units. Measurements (pitch-accuracy ratio, cents, durations) stay raw: they are facts, not scores.
Presets
All presets share one sigmoid template, `0.5·(tanh(3.4·r − b₁) + tanh(3.7·r
b₂))`, differing only in the shifts b₁/b₂ — which guarantees the family never crosses: for every raw input, LENIENT ≥ BALANCED ≥ STRICT.
| raw | LENIENT | BALANCED | STRICT |
|---|---|---|---|
| 0.3 | 0.46 | 0.27 | 0.14 |
| 0.5 | 0.78 | 0.62 | 0.41 |
| 0.7 | 0.93 | 0.86 | 0.74 |
| 1.0 | 0.99 | 0.98 | 0.96 |
BALANCED is bit-identical to the curve Riyaz has shipped for years (including its quirk that a perfect take reads ~0.98, kept so client score history stays comparable). IDENTITY is the default: no consumer's numbers change without opting in.
Custom curves are validated at construction (ADR-022): monotonic non-decreasing over 0..1 and finite everywhere; a broken curve fails at create, not via strange scores mid-session.