| name | segment-evidence |
| description | Fuse video shot segments with storyboard frames, OCR, transitions, and timeline artifacts into one per-segment evidence map for grounded video review. |
Segment Evidence
Use this after video-shots when a reviewer needs one ordered map of what
evidence exists for each video part.
This skill owns:
- reading
video.shots.json
- attaching overlapping storyboard frames
- attaching OCR text evidence and OCR quality counts
- attaching overlapping transition and timeline evidence
- writing
segment.evidence.json
Inputs
Use the same bundle locators as video-artifact-intake:
{
"outputDir": "./output/example-run",
"maxTextItemsPerSegment": 8
}
maxTextItemsPerSegment defaults to 8. outputPath is optional if the
artifact should be written somewhere other than segment.evidence.json.
Outputs
The result contains manifestPath and manifest. The manifest has one segment
per shot, with:
- segment start/end times
- representative shot frame path when available
- overlapping storyboard frames
- overlapping timeline items
- overlapping transitions
- text evidence from OCR and timeline artifacts
evidenceStatus: usable, weak, or empty
Sequencing Guidance
Run video-shots first. Run segment-storyboard plus storyboard-ocr when
global storyboard sampling leaves too many empty segments. Run
storyboard-transitions and timeline-producing tools first if you want richer
segment evidence. The tool can still run with only video.shots.json, but most
segments will be weak or empty.
Interpretation Notes
segment.evidence.json is an evidence router. It says which artifacts support
each segment. It does not prove full semantic understanding by itself.
Treat empty as an abstention signal. Treat weak as a request to inspect the
representative frame or source video before making claims.
Repo-side runner:
node --import tsx scripts/harness/segment-evidence.ts
Installed-pack runner:
node ./.video-evaluator/agent/run-tool.mjs segment-evidence