GCF: A Token-Optimized Wire Format for Structured LLM Interactions

GCF is an AI-native wire format for structured data in LLM systems. 2,500+ evaluations across 11 models and 4 providers. 43 billion+ lossless round-trips across 5 formats. Deployed in 12 production systems.

Comprehension: 100% accuracy on every frontier model on standard workloads. 90.7% on structurally complex data where JSON drops to 54.1% and TOON to 68.8%. No model has ever been trained on GCF.

Token efficiency: 50-92% fewer tokens than JSON (50-69% per call, up to 92% with session deduplication). JSON overhead is 81% structural noise at scale (52.4% repeated field names, 28.6% structural characters); GCF overhead is 0.2%.

Tokenizer safety: GCF is, to our knowledge, the only wire format whose grammar was reverse-engineered from attention-level experimentation and BPE failure mode analysis. GCF's grammar characters were selected from the set empirically verified to have near-zero merge rates across 43 tokenizers from 20 providers. Pipe delimiter: 0.47% merge rate, 24 mergeable words.

JSON: 8.17%, 1,939 mergeable words. TOON tab: 32.91%, 1,238 mergeable words. Controlled experiments across two
architectures (GPT-NeoX, Llama), two scales (410M, 1.3B), and three domains (structured data, code, molecular notation) prove that preventing delimiter merges produces 3-738x better structured data comprehension, with zero natural language cost. Every attention head in a standard BPE model is structurally constrained by delimiter merging; the model has 4x more structural attention capacity than the tokenizer allows it to use.

Six implementations (Go, TypeScript, Python, Rust, Swift, Kotlin), 157 conformance fixtures, zero dependencies. 

Specification v3.2 Stable. Lossless across 17 serializationwe shoul formats.

See companion papers:

- Blackwell, D. (2026) Tokenizer-Attention Coupling — 10.5281/zenodo.20925910
- Blackwell, D. (2026) Stranded Attention — 10.5281/zenodo.21158886
- Blackwell, D. (2026) Developmental Atlas of Attention Head Specialization — 10.5281/zenodo.21205389