Graduated Dissent: Budgeted Disagreement Resolution for Multi-Model Inference

Recent empirical work demonstrates that large language models cannot reliably self-correct reasoning without external feedback, and large-scale evaluation across hundreds of models reveals substantial error correlation even between models with distinct architectures and providers. When generator and evaluator share failure modes, self-evaluation may provide weak evidence of correctness, and repeated self-critique may yield diminishing returns. External evaluation can address this, but external evaluation is expensive. A natural question arises: given a fixed verification budget, how should a system allocate costly decorrelated evaluation across queries? We propose graduated dissent, an inference architecture that treats this as a resource allocation problem. Multiple proposers generate candidate analyses in separated contexts. A comparator estimates whether divergence between proposals is superficial, within an expected domain noise floor, or structurally meaningful. Only high-signal disagreements trigger expensive procedures: steelman exchange, adversarial cross-examination, or external verification via formal proof checkers, executable tests, or numerical invariants. The proposed approach is budgeted inference: escalation occurs when the expected information gain from decorrelated evaluation exceeds its cost, given domain-calibrated priors on disagreement signal content. The protocol combines three mechanisms: context separation between generation and evaluation to reduce inheritance of error-producing reasoning traces, graduated triage to concentrate verification compute where decorrelation has the highest expected value, and a steelman exchange that encourages genuine engagement with opposing reasoning structures. We define domain-calibrated threshold structures and propose pre-specified benchmark families targeting technical reasoning reliability. This paper is a protocol proposal with pre-registered evaluation design; empirical results against the specified benchmarks will be incorporated in a subsequent version. The contribution complements existing evaluation approaches by providing an inference framework that may improve what reaches human judgment.