saintpick.

A clean caption file can still be wrong

A reproducible, synthetic experiment showing where a small caption file checker stops helping. No commercial AI caption editor was tested.

The six fixtures

FileFlags
cleanNone
reversednonpositive_duration
overlapoverlap
emptyempty_text
long_linelong_line
semantic_errorNone

What the checker does

Checks nonpositive duration, overlapping cues, empty text and lines over 42 characters. The line-length threshold is our editorial heuristic, not a platform specification.

The semantic trap

Intended source: “The meeting starts at ten.” Caption fixture: “The meeting starts at nine.” The file passes these structural checks. A human still needs to compare meaning with the actual source and watch the final playback.

Reproduce the result

Download the Python script and run it in a new local folder using Python 3. It writes six .srt fixtures and results.json. Only Python standard-library modules are used. Read the recorded results.

python3 run.py

Limitations

A clean result is a prompt for the next review, not publication approval.