Hallucination-aware evaluation and confidence-guided selective routing for hybrid VLM-OCR: an empirical study on scene text and receipt documents

Vision-language model (VLM)-based OCR can improve transcription on difficult document images, but strong reading performance may coexist with visually unsupported completion. We present SHC, a hallucination-aware evaluation protocol and curated set that separates visible from full transcripts and jointly reports clipped macro character error rate (CER), an operational critical-span flag rate, and empty-output rate. SHC v1.0 contains 5550 samples from six public corpora, including 3897 Test samples. We compare seven primary systems and three diagnostic OCR-specialized baselines and study prompt sensitivity. Qwen3-VL-8B attains the lowest full-transcript CER on SHC (31.85%) but the highest operational flag rate (14.34%), whereas Chandra attains the lowest flag rate (5.39%) by abstaining on 71.8% of samples. We also evaluate ConfRoute, which retains PaddleOCR output above a confidence threshold and otherwise calls Qwen. The frozen historical threshold τ =0.90 is reported without retroactive retuning: on a 116-document internal Test set, ConfRoute yields 26.43% CER versus 26.44% for always-Qwen while reducing estimated latency from 5.009 to 1.099 s/image. On a frozen 900-image CC-OCR evaluation, ConfRoute reduces mean latency by 32.3% and VLM use to 68.33%, with an absolute Eval-Trans decrease of 0.0637 relative to Qwen. A post hoc canonical-source audit identified reuse across some SHC roles and overlap between the independently constructed Baseline and SHC resources; strict source-disjoint validation is therefore not claimed. The results support evaluating OCR accuracy, visual support, and abstention separately and treating selective routing as a transparent, domain-sensitive cost–accuracy control mechanism rather than direct hallucination mitigation.