I Made LLMs Read a 500-Page Specification With 100% Accuracy — Without Fine-Tuning

LLMs fail on large normative documents not because they can't reason, but because they can't navigate. I built a compiler that produces 14 structured indices encoding a domain expert's mental map — chain addresses, ontological routing (WHAT/WHY/HOW/WHEN/WHERE), tier-weighted reading plans, and normative priority scoring. The same models that failed 28% of queries with full-context access achieved 100% accuracy with 7× fewer tokens. Tested across Claude, GPT-4o, and Gemini. All evaluation artifacts and approximate source code are public.