Straight-through processing (STP) demands a calibrated probability for extracted key‑value fields from financial documents without human review, together with a bounded guarantee on the residual error of the auto‑approved tier. Modern vision‑language models (VLMs) can extract key‑values out of the box, yet their verbalized confidence signals are unreliable and weakly correlated with field correctness. This paper introduces a decomposed confidence layer that evaluates three interpretable channels—perception, layout, and validation—and combines them with a final conformal risk control, enabling reliable STP of financial documents. The approach is validated on three public datasets covering real invoices, synthetic invoices, and ad‑buy forms, using two VLM families (Qwen3.6-27B and Gemini-3.1-Flash-Lite). The decomposed score consistently improves the separation of correct from incorrect extractions, raising AUROC from 0.54‑0.74 for VLM verbalized signals to 0.90‑0.99, with contributions from all three channels. Crucially for industrial deployment, the native VLM confidence could clear only 0.1%‑7.0% of fields under risk control at a target error, whereas the proposed method enables usable STP.
Review: The approach decouples multimodal signals and applies risk control, achieving high‑confidence automated processing of financial documents and showing strong potential for real‑world deployment.