Evidence Contracts: Verifiable Control of Evidence Reliance in Multimodal Models

Multimodal models can produce correct predictions while relying on evidence they are meant to ignore, or failing to rely on evidence they are asked to use. Existing evidence-control methods can alter routing patterns or reliance scores, but these changes do not establish the stronger property of behavioral compliance: whether a requested evidence-use relationship holds under admissible intervention mechanisms excluded from adaptation and model selection. We introduce Evidence Contracts, a specify–enforce–verify framework that compiles natural-language evidence-use instructions into explicit, behaviorally testable requirements. Evidence Contracts support compositional IGNORE, REQUIRE, PREFER, and PRESERVE clauses, respectively specifying invariance to designated evidence, task-consistent dependence on required evidence, relative dependence between evidence sources, and preservation of task performance. We ground each clause in multimodal inputs, enforce the resulting contract through contract-constrained adaptation, and verify compliance using validated intervention families held out from adaptation, model selection, and calibration. Across seven multimodal benchmarks, Evidence Contracts achieve 68.4% held-out contract satisfaction rate (CSR) on the common applicable benchmark–backbone settings, compared with 36.5% for GUIDE, while retaining 98.6% of original task utility. Compliance also transfers to unseen contract targets, operator–target pairings, and clause compositions, reaching 63.9% average CSR. These results show that evidence use can be specified, enforced, and independently verified as a behavioral commitment within declared, validated intervention distributions, rather than treated only as a steering preference.