CUDA-05H — An inspectable, falsifiable CPU/CUDA/Vulkan compute witness

Open research, one local hardware trial. An independently implemented MIT-licensed project runs an identical frozen binary scene through Python CPU, native CUDA and native Vulkan compute. External replication is pending; the original private blob.in engine is neither released nor reproduced.

Preregistered hypothesis and numerical workload

Protocol CUDA-05H-P01 was frozen before implementing device kernels at Git commit b89b266829b396d5b9569e0cbb2cd38a04f603bb. The fixture contains 16 immutable primitives over a 96 × 64 grid. Each primitive has a center, radius and union/intersection/subtraction operation. The expression q=(x-cx)²+(y-cy)²-r² is an integer quadratic geometry proxy, not the private engine's floating-point signed distance algorithm.

Pass criterion: exactly 6,144 signed 32-bit outputs or 24,576 bytes, bitwise identical across CPU, CUDA and Vulkan, with zero tolerance. One mismatching byte falsifies parity. Vulkan must dispatch on a discrete or integrated physical GPU, not on a CPU software adapter.

RTX 3070 observations

Backend / negative control	Observed behavior	Verdict
Independently recomputed Python CPU oracle	Full 6,144 values	PASS
Native compiled CUDA	RTX 3070 CUDA GPU kernel, sync/readback	PASS
Native Vulkan/SPIR-V	RTX 3070 GPU compute and fence/readback	PASS
Bit-flipped CUDA output	Independent checker rejects it	PASS
Truncated Vulkan output	Independent checker rejects it	PASS

All three output files share SHA-256 613845b6341e4491efce3be997853467213ef6687770495b0d2515329678961b. The separately implemented numerical checker reports 8/8 operational gates. A public-only source/CPU reference validator runs without GPU drivers.

Independent replication and counterexamples

Full new CPU, CUDA, GLSL, Vulkan code and reproduction protocol · Structured public observations · Submit a replication or counterexample

A CPU-only researcher can reproduce the CPU oracle and check the source manifest, but cannot claim GPU parity. Report exact GPU/driver/compiler versions, shader parameters, output digests, negative controls and unsuccessful attempts. No production system probing, privileged host changes or GPU reset is required.

Validity limits

This is an independently written GPU compute witness, not a reproduction of the original CUDA-02/03 signed-distance renderer, the CUDA-05F Rust supervisor, worker isolation or cross-vendor reliability. Local raw samples/tool logs remain private, and their public hashes do not provide external access. Researchers can generate their own raw outputs using the fully published source code.

Exact statuses: external replication PENDING; original CUDA-05F BLOCKED_PRIVATE_RUNTIME_SOURCE; CUDA-05F GPU parity NOT_RUN; production authorization DENIED.