SICD: Measuring Semantic Surrender and Epistemic Resistance Under Biomedical Interference

Large language models (LLMs) are increasingly evaluated by final answers, but high-stakes biomedical failures often emerge inside the reasoning trajectory that connects evidence to a conclusion. We introduce Semantic Interference and Cognitive Dissonance (SICD), a controlled stress test that measures whether biomedical chain-of-thought reasoning remains anchored to the correct clinical domain when prompts inject contradictory semantic pressure. SICD pairs high-acuity clinical cases with four interference levels and scores each reasoning chain using UMLS-derived signals, centered on the Split Density Ratio (SDR): the fraction of target-domain concepts among all target and interference concepts. Across matched 10-case runs, GPT-4o-mini exhibits semantic surrender: SDR falls strongly as interference increases ($\rho=-0.657$, $p<0.0001$) while oscillation remains zero, indicating fluent adoption of the adversarial frame. Claude Haiku 4.5 instead exhibits epistemic resistance: at full dissonance it often rejects the false premise, defends the target diagnosis, and shows SDR correction at the highest interference level. These results suggest that adversarial biomedical drift is not only a question of hallucination or incoherence; it is a question of whether a model preserves semantic allegiance to the clinical evidence under pressure.