Deep Document Processor

Transform large documents into token-efficient context artifacts optimized for downstream agent reasoning. Goal: maximum decision-relevant signal per token, not human-readable summaries.

Naive approaches fail predictably:

• Full inclusion: blows budget, buries signal
• Naive truncation: loses structure, misses late insights
• Single-pass summary: misses cross-references, flattens hierarchy

The Protocol: Four Passes

Pass 1 — Structural Survey (≤30 seconds)

Extract the document's skeleton before reading content. Creates a mental map guiding attention in later passes.