PhysioSkeptic: Skeptical Reasoning Grounded in Signal Quality for Multimodal Rhythm Analysis

Large language models can reason over physiological summaries, but standard waveform-to-text interfaces often discard beat-level reliability, limiting performance under noise and cross-modal conflict. We introduce PhysioSkeptic for retrospective ECG–PPG rhythm classification. Its PhysioPatch encoder aligns waveform patches to cardiac landmarks and emits heartbeat-level tokens and a Patch Report containing rhythm cues, confidence estimates, and signal quality. This interface exposes beat-level quality as auditable evidence for conflict checks and quality-conditioned belief revision. On MIMIC-III-Ext-PPG, PhysioSkeptic reaches 0.886 Macro-F1, 6.2 percentage points above generic Debate and 4.1 points above an encoder-matched control; it has the lowest observed ECE among GPT-5.2 methods and reduces anchor-trap retention from 56.6% to 16.8%. Removing SQI anchoring produces the largest observed reduction among reasoning controls. In a zero-shot SQI proxy-transfer audit on MC-MED, learned SQI remains informative without retraining, while an auxiliary weak-label AF stress test improves AUROC from 0.913 to 0.934. A distilled Llama-1B student reaches 0.847 Macro-F1 with local inference. These results show that exposing beat-level signal quality can improve the auditability of multimodal rhythm reasoning under degraded measurements.