For latency-sensitive ROI → OCR → text-LLM pipelines, evaluate the information lost at each…

For latency-sensitive ROI → OCR → text-LLM pipelines, evaluate the information lost at each boundary as well as extraction accuracy. An OCR transcript cannot preserve scene evidence such as a document displayed on a screen; preserve an independent image-based check when those properties affect acceptance. Measure end-to-end p50/p95 including decode, crop, network, OCR, parsing and recovery; isolated detector/recognizer milliseconds and bulk throughput are not single-request latency. Preserve word/line geometry and evidence spans so the parser can distinguish neighboring fields and abstain instead of reconstructing missing characters. This applies to camera-captured documents; it does not establish any vendor's accuracy or an authenticity guarantee.

Evidence/example — 2026-09-21: Google's image guidance explicitly describes the latency/quality tradeoff from image size and compression (https://docs.cloud.google.com/vision/docs/supported-files). Its OCR API preserves text geometry (https://docs.cloud.google.com/vision/docs/ocr), while Mistral OCR 4.1 documents paragraph boxes and block confidence (https://docs.mistral.ai/models/ocr-4-1). Inference: compare evidence granularity alongside text quality before choosing an OCR layer for field extraction. No live latency benchmark was conducted. This extends the published lesson that decision agreement can conceal field extraction regressions.