Select OCR experiments by the decision they can change

Select OCR experiments by the decision they can change. Use separate priors for task quality, integration reliability, and deployment economics; do not turn adoption into an accuracy score. A shortlist can test three distinct hypotheses: an inexpensive existing API path, an independent quality reference, and an open customizable model. Add a sibling only for a named unresolved error or economic tradeoff. For a localized text task, a small recognizer plus explicit context validation may answer a more useful question than another general VLM.

Applicability and limits: this is a screening method, not a model ranking. Similar vendors, predecessor scores, and family size do not guarantee correlated task outcomes. Source OCR, crop preservation, exact fields, and acceptance behavior need separate measurements against adjudicated answers. Use matched live baselines when claiming current replacement gains; historical output agreement remains compatibility evidence.

Evidence/example — 2026-09-22: public source review found Qwen3-VL-8B-Instruct offers standard serving paths (https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct), while GLM-OCR exposes a specialized 0.9B document/extraction path with constrained prompt modes (https://huggingface.co/zai-org/GLM-OCR). These are reasons to consider different experiment roles, not proof of equal policy coverage. Hugging Face explicitly counts GET/HEAD requests to query files, with separate work needed to deduplicate users or exclude CI (https://huggingface.co/docs/hub/models-download-stats). Therefore download counts support an ecosystem prior only. No task-specific accuracy or serving advantage was measured by this source review.