Typed Evidence-Conflict Reranking at Video-Language Handoffs

A reranker's full-set improvement is the prevalence-weighted sum of its gains on different query populations, but standard video-retrieval reporting leaves that accounting implicit at the evidence handoff. We introduce the Fixed-Pool Answer-Conflict Audit, which identifies semantically close candidates that disagree with answer evidence inside immutable within-source top-10 pools, and an additive decomposition that separates audited and complementary gains. The operating point is fixed before final-evaluation scoring; blinded adjudication measures proposal precision at 83.0% on TVQA and 78.5% on Ego4D-NLQ, while paired reranker gaps remain positive over two orthogonal five-point threshold sweeps. Typed Evidence Conflict Reranker (TECR) aligns Entity, Region, Utterance, Action, and Temporal evidence, then factors five localized conflicts into applicability, answer relevance, and reliability. Against a BERT-large cross-encoder, TECR improves full-set nDCG@10 by 0.8 and 1.2 points but audited nDCG@10 by 4.2 and 4.7 points; populations comprising 12.7% and 13.4% of the queries contribute 66.6% and 52.5% of those mean gains. Cross-Encoder + OpSup, Type-Agnostic Conflict, and TECR each exhibit a 3.2$\times$–5.75$\times$ audited-to-full conditional gain ratio under the common cross-encoder reference. TECR further retains 1.9- and 1.7-point audited advantages over a supervision- and capacity-matched type-agnostic control, and Charades-STA extends both the gain decomposition and typed-structure advantage to a third corpus. These results turn the delivered evidence state into a measurable model-comparison object and show that typed alignment improves the handoffs carrying a disproportionate share of reranking utility.