kambabench/docs/PROVENANCE.md

# Cross-corpus provenance and leakage audit

**Status: not yet written — awaiting a real run of `dedup_check.py` and `domain_audit.py`**
against the candidate source corpora (needs an environment with real network access to
huggingface.co; see those scripts' docstrings).

This document will hold, once that audit runs:

- the full cross-corpus collision report (text-hash, filename, and targeted audio-hash
  collisions between DDD-Kenya, Thiomi, FLEURS `kam_ke`, and Omnilingual),
- for every collision found, the resolution (confirmed duplicate / independently-recorded
  same sentence / false positive) and which clips were excluded as a result,
- the domain/theme/dialect/speaker distribution audit results for each source corpus, and
- per-clip `provenance` and `provenance_verified` values that downstream `eval_unscripted.jsonl`
  entries should carry, per SCHEMA.md.