Context-Aware Unsupported Video Gating

Problem

The first version of metadata gating treated unsupported keywords as global signals. That made the implementation simple, but it also meant a valid technical review could be rejected just because its metadata mentioned a broad word like "gaming."

For Summry, that is the wrong failure mode. The product should be conservative about unsupported decisions because unsupported videos stop before summarization. A false unsupported result blocks a potentially useful knowledge-dense video.

Context

Summry's current product boundary is intentionally narrower than all YouTube videos. It works best on spoken, knowledge-dense videos such as tutorials, explainers, lectures, interviews, analytical commentary, and technical reviews.

Some metadata signals are high confidence. A music category, lyric video title, reaction title, challenge clip, shorts marker, or gaming category can reasonably suggest the video is outside the current summarization scope.

Other signals need context. A technical benchmark can mention gaming performance without being a gameplay video. That difference matters because metadata is only a prior, not the final understanding of the video.

Implementation

I split unsupported metadata detection into higher-confidence signals and contextual gaming signals.

The gate now checks hard unsupported keywords against stronger metadata fields: title, category, and tags. It also treats explicit music and gaming YouTube categories as unsupported evidence.

Gaming became contextual. A video is unsupported when the YouTube category is gaming or when the title strongly indicates gameplay-style content, such as gameplay, gaming highlights, or let's play phrasing. A Science & Technology benchmark that merely mentions gaming in tags or description can continue through the technical review path.

The tests now cover:

• gameplay metadata staying unsupported
• a technical benchmark mentioning gaming staying supported as a review
• lyric/music metadata staying unsupported
• reaction/challenge metadata staying unsupported

Tradeoffs

This moves the unsupported gate toward higher precision. That reduces false rejections, which is valuable because unsupported is terminal for processing.

The tradeoff is that some vague entertainment videos may proceed farther into transcript and knowledge-density gating instead of being stopped by metadata alone. That is acceptable for this stage because later gates have more evidence than metadata does.

Lessons Learned

Unsupported gates should require stronger evidence than category selection. A category mistake can still be corrected by transcript analysis, but an unsupported mistake stops the product experience.

The word "gaming" is not a category by itself. In technical reviews, it can be an evaluation dimension. In gameplay videos, it is the primary content type. The gate needs enough context to tell those apart.