Separate a VLM provider exit from OCR specialization, and preserve visual policy evidence

When replacing a multimodal extractor for both reliability and latency, treat provider replacement and pipeline specialization as separate experiments. First test a compatible alternate image reader against the existing input/output and abstention contract; independently test compact OCR and ROI stages by document type. This applies when the old VLM performs both transcription and visual-policy checks. It does not imply that a new provider is accuracy-equivalent or that OCR alone can establish original-document, wrong-subject, or reproduction judgments.

A crop can preserve text while removing evidence needed for visual policy. Keep access to the original image until those checks have their own validated implementation. A text-only LLM can map observed spans to fields but cannot reconstruct unreadable identifiers. Evaluate false accepts together with coverage/retakes, and measure the full network-plus-inference path: individually fast stages do not guarantee a fast serial pipeline.

Evidence/example — 2026-09-21 Public documentation review: Claude's vision guide warns against cropping out key visual context and notes that resizing/compression can harm small-text readability (https://platform.claude.com/docs/en/build-with-claude/vision). Mistral OCR 4's release documentation explains that schema extraction adds a downstream language model, and explicitly excludes real-time/latency-sensitive processing from its intended use (https://mistral.ai/news/ocr-4/). Modal documents that payloads above 2 MiB use object storage in us-east even with routing-region selection (https://modal.com/docs/guide/region-selection). These establish architecture constraints, not an application benchmark or a model winner. The recommended two-track transition is an inference from those constraints; which candidate wins remains workload-specific.