What
The golden harness locks the deterministic half of the pipeline (transcription hygiene,
sentence merging, length budgets, SRT timestamps, anchored placement) against recorded
output. Add a case built from a language with different sentence structure.
Why it matters
The current cases are Latin-script and short. Languages with different word order or much
longer words stress the length-budgeting and merging logic in ways the current fixtures do
not reach.
Where in the code
tests/golden/ — each case is a pair:
<name>.input.json # segments in
<name>.expected.json # recorded output
Re-record after adding an input:
UPDATE_GOLDEN=1 pytest tests/test_golden.py
Then read the diff before committing it — that diff is the whole point of the harness.
How to verify
pytest tests/test_golden.py
No whisper, ffmpeg, network or API key is needed, so this runs anywhere.
Effort
~1 hour.
What
The golden harness locks the deterministic half of the pipeline (transcription hygiene,
sentence merging, length budgets, SRT timestamps, anchored placement) against recorded
output. Add a case built from a language with different sentence structure.
Why it matters
The current cases are Latin-script and short. Languages with different word order or much
longer words stress the length-budgeting and merging logic in ways the current fixtures do
not reach.
Where in the code
tests/golden/— each case is a pair:Re-record after adding an input:
Then read the diff before committing it — that diff is the whole point of the harness.
How to verify
No whisper, ffmpeg, network or API key is needed, so this runs anywhere.
Effort
~1 hour.