A static ID-photo corpus can compare per-frame visual decisions

A static ID-photo corpus can compare per-frame visual decisions, but cannot establish live auto-capture timing or jitter. Pair identical model-visible frames across candidates, freeze human-adjudicated labels separately from execution metadata, then use real phone frame sequences for time-to-green, premature capture, and transport latency. A short categorical prompt also needs an explicit reason precedence: visible glare that obscures text may be labeled merely "text unreadable" without it. This method applies to camera guidance systems with a separate downstream OCR step; static-image results and tiny synthetic timings are not release evidence.

Applies when: comparing realtime camera guidance and stateless image classifiers before a production capture decision. Keywords: ID camera capture evaluation realtime model latency glare framing corpus labels