OCR robustness benchmarks and verification acceptance can reward opposite behavior

OCR benchmark strength is a useful prior for transcription, but it is not a prior for accepting an identity or equipment image. A recognizer can correctly read a screen reproduction, an unrelated record, or an invalid printed identifier while a verification workflow should reject it or abstain. Evaluate transcription, field association, source validity, and acceptance separately. Fine-tuning should preserve these separate targets rather than distill an incumbent's single pass/fail output. This applies to image-to-structured-data verification; it does not imply that a specialist OCR model can replace visual checks.

Evidence/example — 2026-09-20: Real5-OmniDocBench deliberately includes screen photography, warping, illumination and skew to test document parsing (https://huggingface.co/datasets/PaddlePaddle/Real5-OmniDocBench ; https://arxiv.org/abs/2603.04205). Good transcription on screen photography is desirable in that benchmark but cannot establish original-document acceptance in a verification application. The inference is about metric alignment, not a measured fraud-detection rate. OpenAI's vision fine-tuning documentation also excludes images containing faces from training, making training-data eligibility a separate feasibility gate for full portrait-bearing IDs (https://developers.openai.com/api/docs/guides/vision-fine-tuning). No private images, customer data, or internal benchmark statistics are included.