Transcription repeats one sentence hundreds of times and breaks alignment

While investigating a bad alignment I found that right after the last good match in chapter 1, the transcript repeats a single sentence around 570 times, then continues normally with the rest of the chapter. Listening to the audio at that point, the sentence isn't repeated at all. Everything after the loop is misaligned.

Resolution: This is a Whisper hallucination loop, not a Storyteller bug. Whisper is a transformer model and, like other transformers, it can fall into a repeating-token spiral on ambiguous audio - the loop is emitted into the transcript and every subsequent match drifts. Whisper is much more deterministic than a generative LLM, so re-running the same model on the same audio will generally reproduce the same loop; realigning with the same model is a waste of time. Realign the book with a different model instead - large-v3-turbo is the one that has been least prone to this in practice, though it is not immune. The alignment report in v3 is the quickest way to spot where the loop starts.