Bind the chosen OCR value to the image source being approved

A parallel image-context check and text-only identifier selector can both pass while referring to different objects. The context model may correctly see a readable target identifier, while the selector returns a background number, equipment model, or specification. Conjoining their pass flags does not establish that the selected value belongs to the approved source. Preserve spatial/source evidence or evaluate the selected value against the image, and test false accepts as well as retries. A visual camera guide alone provides no binding when inference receives the full frame. Extra validation adds latency and needs independent evaluation; blindly cropping may remove needed context.

Applies when: OCR candidates and image-context approval are evaluated independently. Keywords: OCR Gemini Qwen candidate selection source binding geometry false positives