Camera Harness

The user sees one scene. The implementation handles multiple time windows, coordinate systems, model boundaries, and lifecycles. Without a shared immutable identity, visual continuity becomes a presentation property rather than an evidence guarantee.

Model execution alone is not trustworthy product behavior. The scheduler, freshness policy, memory bounds, and final commit check are what keep slow inference from becoming stale UI.

Current evidence supports short-lived track IDs, but controlled occlusion, overlap, camera motion, disappearance, and re-entry performance remain unproven.

Frames are captured after the transcript. No immutable frozen frame, exact crop, region ID, visible track ID, or Microscope best-frame record is attached to the answer. The answer can therefore describe a different evidence window from the label the user sees.

Readiness evidence must be commit-pinned to the active production graph.

Several tests proved isolated contracts rather than the active product graph. Test count alone could not establish that a feature remained integrated.

The next valid benchmark is a commit-pinned stage budget covering capture, queueing, detector time, semantic time, dropped work, memory, and energy per completed task.

type VisualEvidence = {
  frameId: string;
  capturedAt: number;
  source: { width: number; height: number };
  transform: { mirrored: boolean; visibleWidth: number; visibleHeight: number };
  trackId?: string;
  region?: { x: number; y: number; width: number; height: number };
  cropId?: string;
  inferenceId?: string;
  model?: { name: string; version: string };
  processingBoundary: "browser" | "local-service" | "hosted-service";
};