We present a lightweight, deterministic post-processing approach for clinical text classification based on entanglement between clinically meaningful concepts. Our system was developed for the SMM4H 2026 shared task on insomnia detection and related information extraction from clinical notes. For Subtask 1, we introduce an entanglement-based rescue layer that models dependencies between sleep disturbance, daytime impairment, and sleep-targeted medication evidence. Applied as a false-negative correction on top of an LLM baseline, this approach improves recall while preserving precision. On the official test set, the rescue layer increases F1 by 25% without degrading precision (1.00). Local experiments show larger gains on weaker runs, suggesting a stabilizing effect on variable LLM outputs. For Subtask 2, we implement an LLM-based system for rule-based evidence and span extraction. Results highlight the effectiveness of modeling clinically grounded dependencies and suggest directions for improving evidence extraction and span matching.