LLMs deployed as judges in high-stakes classification are commonly assumed to let their generated rationale shape the verdict; we show this assumption fails in disinformation detection. We introduce the Multi-Bias Injection Protocol (MBIP), which perturbs the reasoning stage along four cognitive axes (confirmation, anchoring, framing, authority) while holding the input fixed, and quantifies induced rationale variation with technique-count metrics anchored on the number of persuasion techniques the model reports. Across five disinformation datasets, MBIP produces substantial rationale variation but near-zero verdict flipping; an explicit rationale-deference prompt raises the flip rate only modestly, to 0.118–0.300 across three open-weight models, with API-only gpt-4o-mini in the same range. We call this verdict-level invariance Decision Rigidity. At the representational level, our residual-stream probe localizes a candidate commitment band late in the network, where biased and neutral final-class trajectories converge. The band’s location is architectural, replicated by a neutral-vs-neutral null; a within-band alignment-rate elevation and a random-lexicon control gap co-occur with the behavioral invariance and attenuate in Gemma-2-9B. All representational evidence is correlational. Generalization to other high-prior, evidence-dependent classifications and attenuation on reasoning-intensive tasks remain falsifiable hypotheses.