Hidden State Forensics for CMC Document Integrity: Engineering Validation of Autonomous Dissonance Perception

We present an engineering validation of hidden state-based anomaly detection for pharmaceutical CMC (Chemistry, Manufacturing, and Controls) documents. Building on the Autonomous Dissonance Perception (ADP) framework, which predicts that language models exhibit detectable internal state perturbations when processing logically contradictory inputs, we conduct four rounds of experiments using Qwen2.5-7B-Instruct on consumer hardware (Mac Mini M4, 24GB). Round 1 (25+25 short prompts) establishes that CMC-specific logical traps produce distinct hidden state signatures (AUC=0.800 after length control). Round 2 (10+10 full documents) reveals signal dilution in whole- document analysis (AUC=0.630, not significant). Round 3 (102+10 section-level chunks) recovers the signal through document chunking (AUC=0.972, p