Artifact

detector_report.json explains model mode, warnings, and review boundaries.

The detector report helps teams understand what happened during processing and when to review or escalate.

Example shape

{
  "detector_mode": "cpu_first",
  "sampled_seconds": 20,
  "warnings": ["low_light"],
  "fallback_recommended": false
}

How to use it

Render detector warnings near answers and events. If low light, motion, or occlusion appears, avoid overconfident claims.

Review posture

Do not treat an artifact field as a final decision. Pair important artifact claims with evidence references, confidence, detector warnings, and human review boundaries when the workflow is high-risk.

A good product surface links the artifact to the user interface. When a reviewer clicks an event, the UI should show the timestamp, supporting evidence, confidence, and any detector warning. When an answer is generated, the same evidence should be visible. This makes debugging possible and prevents the documentation from becoming a theoretical schema list that developers cannot operationalize.