kambabench/docs/EVALUATION.md

# Evaluation protocol

**Status: not yet written.** Roadmap item v0.3 ("Freeze scoring protocol").

This document will specify, once the evaluation set is frozen:

- exactly how to run a system against `eval_unscripted.jsonl` and produce a report,
- the scoring implementation used (`score_asr.py`, `kb_io.py`) and its pinned dependency versions,
- how raw vs. normalized CER/WER are computed and reported side by side,
- how dialect-level, domain-level, speaker-level, and read-vs-spontaneous breakdowns are produced,
- the confidence-interval / significance-testing method used when comparing two systems, and
- the exact command(s) needed to reproduce any published number end to end.