ThoxOS GPU lane
See which kernel the ThoxOS GPU layer binds on each THOX host, compare the Tile and CUDA C++ launch paths measured on the same RTX 4060 Ti workstation on the same day, check which launches the Rust compiler rejects, and rerun the RMSNorm reference the kernels are tested against. Every number comes from the report's evidence set, labelled by how it was obtained.
Applies the GPU layer's fallback order for FP32 RMSNorm: a qualified Tile kernel first, then the edge-ops CUDA C++ kernel, then thox-rmsnorm on the CPU. Tile binds only where it can run and the registry qualifies it for that compute capability. Edit the fields to model another host.
Tile RMSNorm against the production edge-ops CUDA C++ kernel on the THOX RTX 4060 Ti workstation (sm_89), same protocol, host and day. Both sides time output allocation, launch, synchronization and copy back to the host. Read the caveats before the ratio.
Build a call to kernel::add(output, x, y) and see whether rustc accepts it. The checker encodes Proposition 1 of the report; its verdicts are tested against the seven compile cases measured on cutile 0.4.0.
The FP32 contract from thox-tile-ops, recomputed in your browser with the same generator: y = x · 1/√(mean(x²) + ε) · w. Known-answer vectors come from the Rust reference; the check passes only on bit-exact agreement.