"A clean-looking transcript can be perfectly, fluently wrong."
What the Quality Check Actually Measures
When Recall accepts a transcript, it runs a structural quality assessment. It's genuinely useful — it catches the tell-tale shapes of a broken transcription:
Repeated segments — the same line 40 times, the classic signature of a model looping on silence.
Long repeated spans — a phrase that runs on far past where any human would stop.
Timestamp regression — timestamps that go backwards, meaning the projection is confused.
Known stock phrases — the boilerplate hallucinations a model emits when it hears nothing ("thanks for watching," subtitle-credit filler).
These are real signals. A release flagged severe can even be blocked from generating a summary, because summarizing garbage produces confident garbage. So the quality check earns its place.
But It Measures Shape, Not Truth
Here's the line you must not cross: structural quality measures the shape of the text, never its correctness. The two come apart in both directions:
A transcript can be structurally spotless — no repeats, clean timestamps, no stock phrases — and still be wrong. The model can mishear a name perfectly fluently. Nothing in the shape reveals that.
A transcript can be structurally messy — a few repeats, one odd span — and still be completely usable for finding the moment you're searching for.
So the quality score is advisory. It informs a human reviewer and can gate an expensive derived step. It is never a proof of correctness, and — this is the invariant the next track is built on — it never auto-promotes a transcript to 'good enough to be knowledge.' A clean structural score and 'trustworthy as a source' are two different judgments, and the machine only gets to make the first one.
Why This Restraint Matters
It would be so tempting to let a green quality score mean 'approved.' It's automatic, it's cheap, it feels objective. That's exactly the trap. The moment a proxy metric becomes the approval gate, you've quietly redefined 'correct' as 'passes the structural check' — and now every fluent hallucination sails straight through. Recall keeps the metric in its lane: a helpful advisor that a human can overrule in either direction, never a judge. Hold that distinction; Track 7 turns it into the hard wall between 'complete' and 'approved.'
Code
Quality gates an expensive step; it never approves truth·python
assessment = {
'flags': ['repeated_segment', 'stock_phrase'],
'severity': 'severe', # advisory only
}
# LEGITIMATE use: gate an expensive derived step.
if assessment['severity'] == 'severe':
skip_summary_generation(release) # don't summarize garbage
# FORBIDDEN use: treat a clean score as approval.
if assessment['severity'] == 'clean':
approve_for_corpus(release) # NO. shape != truth.
# a fluent mishearing is 'clean'.
# Structural quality informs a human; it never auto-approves.
Find an automated quality score in your own world — a lint pass, a coverage percentage, a health check, a spam score. Ask two questions: does a passing score prove the thing is actually correct, and is that score used anywhere as an automatic approval? If a green score silently approves something, find the case where the shape is clean but the substance is wrong, and add a human judgment step for the decision that actually matters.
Hint
The dangerous pattern is 'score is green, therefore ship/approve/trust.' Separate the two roles: let the metric block obvious garbage and inform a reviewer, but route the real approval — 'is this actually good enough for its purpose?' — through a judgment the metric only advises. Clean shape and trustworthy substance are different questions.
Progress
Progress is local-only — sign in to sync across devices.