LICENSING

A compiled binary containing the full Railgun codec, an evaluation harness that generates seven synthetic weight distributions (matching real model profiles), runs tier benchmarks across all routing modes, tests algebraic integrity, and produces both JSON results and an interactive HTML report with charts. No GPU required—runs on CPU.

The core codec is protected as a trade secret. The compiled binary allows full performance verification without exposing the underlying algebraic structure. This protection model is standard for high-value quantization IP.

The evaluation package uses synthetic weight tensors that match real model distributions (normal, heavy-tail, bimodal, sparse, etc.). For production license holders, custom integration with your specific model weights is included in the support package.

At the Full reconstruction tier, Railgun recovers the exact fp16 representation of every weight. This is not an approximation—it's a mathematical identity. The intermediate tiers (R-only, R+G, R+G+B) provide progressive quality at lower VRAM cost, useful for scenarios where approximate inference is acceptable.

Evaluation is free under NDA. Production licensing is negotiated based on deployment scale, model sizes, and support requirements. Both flat-fee and revenue-share structures are available. Every deal starts with a conversation—no published price list, because every deployment is different.

The codec is designed for TMU (Texture Mapping Unit) acceleration on modern GPUs, achieving up to 5× throughput per watt compared to ALU-based quantization. The evaluation package runs on CPU for accessibility. GPU-accelerated builds are available under production license.