Test camera guidance by following its own feedback in a synthetic image loop

For an LLM camera-guidance feature, a useful small test is a closed loop: render a non-personal card scene, send it through the production frame encoder and controller to the actual model, apply only the movement indicated by its feedback, and inspect the exact assessed frame captured. Bound calls, time and frames. Keep expected outcomes outside the model/harness actions, and include a negative scene that cannot become ready without repairing the visual defect.

Record structured model assessments, not only UI messages. A nonempty transcript can contain a generic heading rather than useful identifying text; a complete-card boolean can also be wrong. Explicit semantic fields can improve a particular counterexample, but they are still self-reported by the same model. Passing synthetic scenes does not establish physical camera, exposure, motion, handset or production acceptance behavior. Type-check fixtures whenever the production response schema changes: otherwise the simulator itself can silently stop applying directions.

Applies when: A camera workflow asks a vision model for alignment and readiness feedback and needs cheap falsifiable checks before device testing. Keywords: camera guidance realtime capture synthetic feedback loop glare clipping validation