From the 0.23.0 validation run on the Cueva source (1145.2s, enhanced microphone track).
What happened
driftSuspect fired on 16 of the agent's 18 proposals. In most of those, the agent's independent say --transcribe --words measurement was right and the SRT-derived removedText was wrong, sometimes badly: once it displayed content the agent intended to KEEP as though it were being removed, once it showed zero overlap with the actual cut.
The warning is doing its job. The problem is what the warning implies about the field it guards: on this source, removedText was close to useless as a sanity check, and the agent re-transcribed every single cut to trust it. That roughly doubled the run's tool calls.
Why this matters more than it looks
removedText is the field a human or an agent reads to decide whether a proposal is safe to accept. If it is unreliable on a whole class of source, then the cheap review path is unavailable exactly where review matters, and the expensive path (re-transcribe per proposal) becomes mandatory without anything saying so.
Related mechanism, already filed: #52 (the session transcript is of the source), and the Whisper gapless-cue behaviour measured in #60, where a source cue absorbs the pause after it, 547ms per cue on this material against 302ms in the human reference. A removedText built from cue boundaries inherits that drift directly.
Directions
- When
driftSuspect is true, either withhold removedText or mark it as derived-from-drifted-cues rather than printing it as if it were what the cut removes.
- Consider deriving
removedText from word-level timings when the session has them, rather than from cue spans.
- If neither is cheap, at minimum say in the proposal output that the quoted text is not trustworthy on this source and name the command that settles it, the way other disagreement readings in this codebase already do.
From the 0.23.0 validation run on the Cueva source (1145.2s, enhanced microphone track).
What happened
driftSuspectfired on 16 of the agent's 18 proposals. In most of those, the agent's independentsay --transcribe --wordsmeasurement was right and the SRT-derivedremovedTextwas wrong, sometimes badly: once it displayed content the agent intended to KEEP as though it were being removed, once it showed zero overlap with the actual cut.The warning is doing its job. The problem is what the warning implies about the field it guards: on this source,
removedTextwas close to useless as a sanity check, and the agent re-transcribed every single cut to trust it. That roughly doubled the run's tool calls.Why this matters more than it looks
removedTextis the field a human or an agent reads to decide whether a proposal is safe to accept. If it is unreliable on a whole class of source, then the cheap review path is unavailable exactly where review matters, and the expensive path (re-transcribe per proposal) becomes mandatory without anything saying so.Related mechanism, already filed: #52 (the session transcript is of the source), and the Whisper gapless-cue behaviour measured in #60, where a source cue absorbs the pause after it, 547ms per cue on this material against 302ms in the human reference. A
removedTextbuilt from cue boundaries inherits that drift directly.Directions
driftSuspectis true, either withholdremovedTextor mark it as derived-from-drifted-cues rather than printing it as if it were what the cut removes.removedTextfrom word-level timings when the session has them, rather than from cue spans.