Value System Kernel

Most safety guardrail logic sits in Python middleware: it parses context, calls a classifier, and returns a verdict while your GPU idles. value-system-kernel challenges that architecture by running injection and jailbreak detection directly in a CUDA kernel using pure bitwise operations and FMA intrinsics — no semantic parsing, no Python overhead. The workflow loads danger reference vectors into GPU memory, then geometrically verifies incoming input vectors against them at inference time, with the kernel designed to keep latency flat by avoiding the warp divergence that conditional branching introduces on GPU hardware.

The differentiating claim, as the repo describes it, is the shift from ‘Semantic Software Guard’ to a ‘Pure Bitwise Aesthetic’ scan space anchored in physical memory addresses. By operating at the IEEE 754 bit level rather than token or embedding level, V2 attempts to make the safety check a hardware-cost operation rather than a model-cost one — the vendor frames this as eliminating what they call the ‘Alignment Tax’ of traditional approaches.

This fits a narrow use case: CUDA safety researchers, kernel engineers prototyping hardware-level guardrails, or teams building custom inference pipelines who want to validate whether geometry-based filtering is architecturally viable before committing to a full implementation. It does not fit teams looking for a tested, documented guardrail layer they can attach to an existing LLM API or Python inference stack. The repo carries an explicit ‘blueprint concept’ label — the vendor states it is engineered to establish architectural direction, not to ship in production as-is.

Both V1 and V2 kernel files are present alongside test headers, giving researchers a starting point for experimentation. No license is stated on the page, which creates adoption friction for any commercial or redistributed use. API access is absent; self-hosting on CUDA hardware is the only supported path.