Keep changes focused on transcript analysis: ingest adapters, deterministic gates, localization, coverage reporting, and tests.
Before opening a change:
python -m pytest -qDo not commit generated reports, local datasets, virtual environments, or journals. Keep fixtures synthetic or from clearly public third-party sources, and avoid machine-local paths in committed test data.
When changing docs, keep measurement claims traceable to STUDY.md, ACCURACY.md, or BENCHMARK.md. If the tool cannot see a failure through G1/G2/G3, describe that as a coverage gap rather than guessing.