OCR shortlist priors should use acquisition-matched benchmark slices

OCR shortlist priors should use acquisition-matched benchmark slices. A model competitive on clean PDF transcription can be a poor first choice for photographed documents. Establish the required capability and image acquisition conditions before selecting the leaderboard; keep transcription, document structure, schema extraction, and capture-policy evaluation separate. This applies to candidate screening, not claims of production superiority. Missing leaderboard rows mean unavailable evidence, not low quality.

Evidence/example — 2026-09-22 public-source review: MDPBench's matched photographed-document scores are PaddleOCR-VL-1.6 76.3, Qwen3-VL-8B-Instruct 65.0, GLM-OCR 63.7, Qwen3.5-9B 62.7, and LightOnOCR-2-1B 58.5 (https://github.com/Yuliang-Liu/MultimodalOCR/blob/main/MDPBench/README.md). Paddle's 83.0 is the private-split score, not overall; overall is 78.9. These parsing scores are not exact identifier accuracy. A different acquisition/task benchmark can therefore justify a different first experiment without establishing that one model is universally better. For schema/null/record correctness, consult ExtractBench's separate contract and deterministic metrics (https://github.com/run-llama/ExtractBench); it is authored by a participating vendor and concerns enterprise documents, not roadside scene text. Use resulting hypotheses to choose a small paired domain test rather than averaging incomparable benchmark scores.