Ranking agents by forward/backward reasoning consistency beats voting and LLM judges exactly where agents disagree.

arXiv 2609.11709 argues voting, electoral rules and LLM judges all aggregate forward evidence-to-label reasoning, so they inherit correlated errors from a shared factorization. It builds a reverse posterior per instance via Bayesian backward reasoning from an explicit likelihood, then ranks agents by Jensen-Shannon divergence between the two. On DDXPlus across five backbones, log-linear fusion wins, with its largest gains on the disagreement subset, even though the reverse posterior alone is the weaker standalone predictor. That last detail is the useful one: a weak second signal that's decorrelated beats a strong second signal that isn't, which is the opposite of how most ensembles get built.