On 17 October 2026 we had four transcripts that had failed one of our own automated checks. Our rule says such a transcript is not served to new requests. That leaves a question every transcription service eventually faces and few write about: what do you do with a transcript you have decided not to trust? You cannot just run it again and publish the new one, because the new one might be worse. This post describes what we built, the four first cases, and what happened to each.
Why you cannot simply regenerate on demand
Two things go wrong if every visitor can force a fresh transcript. It costs real money every time, so a few clicks could exhaust a monthly budget. And a fresh transcript is not automatically a better one: the same model, asked again, can produce a different result that is equally broken, or broken in a new way. So a new transcript has to be earned, not assumed.
What we built
- A regeneration is a candidate, not a replacement. It is stored next to the current version. The current version, and every earlier one, stays exactly as it was, so there is always something to go back to.
- Only we can start one, with a stated reason, and the system allows one at a time per video and three attempts per transcript version. A request refused for budget does not use up an attempt.
- The candidate goes through the same review as any transcript. It is promoted only if that review has finished and is not worse than what it would replace. A candidate that fails a serious check can never be promoted. A candidate that is merely flagged more than the current version needs a person to say so, in writing, with a reason that is stored.
- Promotion is reversible. Every change of the preferred version is logged, and going back is one step.
The four cases
1. A twelve-minute drone video: fixed
The held transcript had failed the coherence check. The regenerated candidate passed the serious check and was flagged for two things, coverage and coherence, as warnings. That is better than a failure, so it was promoted. The result is not a clean bill of health: the new transcript is still marked “flagged for review”, and anyone using it should check it. Cost: about 27 cents of AI work.
2. A ten-minute video about everyday gear: not fixed
The candidate failed the same coherence check as the original. The system refused to promote it, as designed, and we discarded it. The original stays held. For some videos, running the same model again does not help, and the right outcome is to keep saying “we do not trust this transcript”. Cost: about 20 cents.
3. A 27-minute video: the job died
This was the long-video failure described in why AI transcription fails: the job started and never reported back, because of a platform limit on outside calls that we had not yet found. It was recorded as failed, the 32 cents it had already spent was kept in the books, and the rest of its reserved budget was released.
4. A 104-minute panel discussion: refused before any work
The monthly budget did not have room for a video that long. The request was refused up front, nothing was spent, and it did not use one of the three attempts.
What the four cases teach
- One of four was actually repaired. A held transcript can be rescued, but it often is not.
- The review is what makes regeneration safe. Without it, we would have promoted case 2's equally broken candidate.
- Limits are not a nuisance, they are the feature: case 4 was refused by the budget limit before any money was spent. Case 3 shows the other side. A limit we had not known about cost 32 cents and produced nothing. That is why, on 18 October, we paused videos over 15 minutes until a long video had been tested end to end. On 20 October, after long videos were changed to run as many small jobs, test videos of 15, 58 and 104 minutes all completed, and videos up to two hours are accepted again. The full story is in why AI transcription fails.
- “Flagged” is a working state, not a failure. Case 1 now serves a transcript that comes with an honest warning.
We also tested the reversal on a short, clean video: the candidate was promoted and then rolled back, and the log recorded both steps. See the dated entries on the what's new page, how the checks decide what is held in how we evaluate transcription quality, and a real result on the worked example.
Need a transcript to try this on?
Paste a public YouTube link and get a full transcription with timestamps. Free within fair-use limits, with a free account.
Try AIOS Scribe