Route Evidence Before Revising: Channel-Selective Verification for Low-Call Multimodal Repair

Post-hoc multimodal correction is a one-edit decision: boxes, counts, geometry, and temporal tracks may support different answers, so a verifier must repair errors without overwriting correct drafts. Selective Channel Verification (SCV) manages this trade-off with a utility–risk router that invokes a grounding, counting, geometry, or temporal critic on only its aligned cache record. On identical GQA draft/cache states pooled over five seed-paired runs, SCV repairs 3,411 initially wrong drafts and damages 1,578 correct ones, compared with 3,204 repairs and 1,685 damaged drafts for a union-trained, capacity-matched verifier. Spatial accuracy is 58.2$\pm$0.3 versus 56.6$\pm$0.3, a 1.6-point difference (95% CI: 1.1, 2.1). A second intervention sends all four records to the unchanged channel-trained critics; spatial accuracy falls from 58.4 to 55.4 and over-correction rises from 3.8% to 4.4%. NLVR2 shows the same spatial and draft-preservation pattern, and replacing the backbone with Qwen2.5-VL-7B-Instruct yields 62.0$\pm$0.3 versus 60.8$\pm$0.3 spatial accuracy. An outcome-blind proposal-disagreement score places the largest difference—2.5$\pm$0.6 points—among examples with competing channel-local proposals. At approximately 1.4 verifier calls per question, SCV improves the repair–damage operating point and establishes evidence access as an architectural choice in multimodal verification.