Semantic Opacity: Content-Blind LLM Document Processing with Formal Information-Theoretic Guarantees

Large Language Models (LLMs) are increasingly deployed for document formatting, yet their attention-based architectures inherently comprehend input content, creating fundamental barriers for privacy-sensitive applications. We introduce semantic opacity — an architectural approach that provides formal, information-theoretic guarantees that an LLM never receives semantic content during document processing. Our system transforms raw text into categorical structural fingerprints encoding only quantised length class, word-count class, and structural type signals before any LLM invocation. We prove that total information leakage is bounded above by 7.5 bits per text block — far below the threshold required for content reconstruction. Empirically, the system achieves 95.6% formatting accuracy across a combined corpus of 131 documents (551 classification decisions, 95% CI [93.6%, 97.2%]) spanning synthetic, hand-annotated natural-language, and out-of-distribution evaluations, while maintaining perfect opacity: zero content words leaked across all scanned content. A content-aware baseline achieves only 96.9% — demonstrating that content access provides a mere 1.3 percentage-point marginal benefit. We present the complete, self-contained, and deterministically reproducible implementation alongside cross-corpus generalization experiments and real feature ablations. Our central result — that formatting intelligence is separable from content access — establishes a novel paradigm with immediate applications in legal, medical, and privacy-preserving document processing.