hypit boundaries: cut candidates and the 0.1 threshold

hypit boundaries returns candidates[] {at, score}: possible cut times scored from a 12 fps, 32x32 mean-absolute-difference pass, with defaults rate 12 and threshold 0.1. It is a mechanical series to read alongside frames, not a verdict on where the edits are.

It is unbilled and, like the rest of the lane, flag-gated: it is listed where the matching environment flag allows them (development auto-on, production opt-in), so check tools_list or the OpenAPI before you build on one.

How is this different from reference ingest shots?

Reference ingest's shots[] is the finished answer: frame-exact cuts from ffmpeg scdet and PySceneDetect voting together, tiling [0, duration] with no gap. hypit boundaries is raw material: scored candidates for an agent that is reading the clip in stages, with tiles and cuts to confirm them.

What do rate and threshold change?

rate is how many frames per second are compared (default 12). threshold is the score a difference must reach to appear (default 0.1). Lower the threshold and you see more candidates, including fades and fast motion that are not cuts; raise it and a soft dissolve may vanish. The docs give no tuned values beyond the defaults, so start with them and verify each candidate on a tile.

curl -X POST https://api.sume.com/v1/hypit-understand/boundaries \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: hypit-bd-001" \
  -d '{"understanding_id":"<probe job id>","rate":12,"threshold":0.1}'

When would I pick boundaries over ingest?

When you are already inside a hypit read of the clip and want the cut series next to your transcript and tiles, or when you want to see near-threshold candidates that a voting detector would drop. When you just need a clean shot list to drive a remake, reference ingest gets there in one call and tiles the whole clip with no gap.

Neither tells you why an edit exists. The cut series is timing; the reason a cut lands on a particular word is your reading of the tile and the transcript together.

How do I confirm a candidate?

Run tile with at[] set to the candidate times, or every_frame for a window of 4 seconds or less, which returns every native frame. Then use cut to burn source time into an excerpt (label_time) if you want to review it as video.

There is no stage that depends on the transcript here: boundaries, overview tiles and dense tiles can run in parallel containers right after the probe.