When a language model ignores retrieved evidence and answers from memory instead, this is not a knowledge failure, the correct information is present. It is a routing failure; the model commits to its parametric answer despite the evidence signal. We show this distinction is mechanistically meaningful, localizable, and correctable at inference time without training data or weight modification.

A single prefill-margin computation detects routing failures (AUC ≥ 0.93 across five model dataset pairs) and gates a format-appropriate logit-bias correction. On MedQA multi-option conflicts, this achieves 77.5% restoration at zero harm, exceeding the strongest prompting baseline by 67.5 pp and matching domain-matched ITI without training data. On open-ended PopQA (3,141 items, three models), balanced alias promotion achieves 80-88% restoration at near-zero harm, Pareto-dominating prompt-only baselines in five of six cross-model conditions. The central mechanistic finding is that effective late-layer intervention sites are structurally distinct from the causal commitment locus (a five-head cluster, joint flip rate 88.6%, p < 10^-9), suggesting that steering can succeed through residual-stream write access without exactly identifying the full decision circuit.