Data Cost Protocol

Compact structured data delivery for AI agents. Positional arrays, schema registry, and adaptive density — because JSON was designed for humans, not for agents.

dcp-docs.pages.dev.

Implementation:

github.com/hiatamaworkshop/dcp-wrap

github.com/hiatamaworkshop/dcp-rag

Documentation

• Specification — Core protocol, $S header, benchmarks
• Schema-Driven Encoder — System→AI encoding, schema generation, output controller
• Shadow Index — 5-level density spectrum (L0–L4), adaptive selection
• Agent Profile — Per-agent error rate observation, density adjustment
• Native Operations — DCP as a processing format, primitive operations for brain AIs
• Validation — Passive education, cost gradient incentives

Research

• Format Comparison — NL vs JSON vs DCP accuracy
• Lightweight LLM & Density — Sub-4B model compatibility, shadow level testing

Key Ideas

• Strip what machines don't need — No keys, no labels, no repetition. Schema once, data by position.
• Schema is the single source of truth — Add a JSON file, get an encoder. No code changes.
• DCP ≈ JSON in accuracy, ~50% fewer tokens — Empirically verified across multiple model sizes.
• Adaptive density — System observes agent capability and adjusts schema verbosity automatically.
• L0 (fields only) is optimal for lightweight models — Protocol markers are noise at ≤4B.