transcript-cleanup
Transcript cleanup
Turn a garbled transcript into two files: the RAW exactly as received and a CLEANED sibling a reader can trust. The work is recovery, not writing: every sentence in CLEANED is one the speaker said, every uncertain word is flagged with its raw wording, and the file ends with the questions only the human can answer — the person who was in the room resolves more garbles in a minute than any search.
The stakes on each side: an unflagged guess reads as fact and poisons everything built on the file (a knowledge library, a quote, a decision), while a file that flags every second word is unreadable and hides the flags that matter. Flag what changes the meaning; smooth what does not.
Run in four phases
The phases are a generate–evaluate–repair loop with a research step in front. Keep them separate: the model that drafts is not the model that checks it.