Transcript Quality Gating

Context

Summry is adding a layered constraint gate so the product can become more reliable on knowledge-dense YouTube videos. Gate 1 created durable gating fields. Gate 2 added a metadata-only prior. Gate 3 moves the decision closer to the actual source material: the transcript.

This matters because metadata can only guess whether a video is a good fit. The transcript determines whether Summry has enough spoken evidence to summarize responsibly.

Problem

Before this gate, the worker moved from transcript fetching directly into chunking and summarization. That meant low-quality transcripts could consume embedding and LLM work before the system had recorded whether the source material was strong, usable, weak, or empty.

Weak transcripts are not all the same. A visual-heavy tutorial may still be worth summarizing with caution, while an empty transcript should not be summarized at all.

Implementation

This gate adds transcript-quality analysis after captions or transcription are available and before chunking starts.

The analyzer checks:

• usable segment count
• average segment length
• repeated segments
• visual-reference markers
• filler density
• sentence completion
• generated-caption and language-confidence signals

The worker persists the resulting transcript-quality label. Strong and usable transcripts continue normally. Weak but non-empty transcripts continue with a caution-style eligibility reason. Empty transcripts stop before chunking and summarization, then use the existing failed-processing path with a clear reason.

Tradeoffs

The benefit is that Summry now records source-quality information before spending more processing work. This makes later prompt routing, UI warnings, and debugging easier.

The tradeoff is that the first scoring rules are heuristic. They are intentionally simple and testable, but they will need calibration against real videos. Gate 4 should use transcript quality as one input rather than treating this score as the full eligibility decision.

Lessons Learned

Transcript quality belongs before summarization, not after. Once the worker has captions or a transcription result, it has enough information to decide whether to proceed normally, proceed cautiously, or stop early.

Keeping the check before chunking also keeps the failure boundary understandable: if there is no usable spoken text, Summry should say so instead of producing a hollow summary.

Next Steps

Gate 4 should combine metadata prior, transcript quality, and transcript sample heuristics into final knowledge-density eligibility. That is where unsupported videos should stop more intentionally and where caution should become a user-visible product state.