A CUDA validation lab built around replayable failures and explicit evidence boundaries.
Vanta compares GPU results with CPU references, checks metamorphic relations, runs Compute Sanitizer, serializes counterexamples, and replays them against a chosen binary.
- Correct reduction, vector-add, and tiled-matrix-multiply campaigns
- Clean memcheck runs for the correct implementations
- Detection and replay of deliberate tail and tile defects
- 16 of 18 mutants killed in the first mutation campaign
- One equivalent mutant recorded as equivalent
- One instrumented deadlock recorded as infrastructure failure, not detection
Correctness, sanitizer output, replay classification, and performance remain separate evidence. A clean sanitizer run does not prove numerical correctness or race freedom.
The v3 mutation matrix is a protocol until a GPU campaign produces committed evidence.
Run on a GPU, inspect the campaigns, replay failures, and read the tolerance model.
Development history and evidence index.
See GUIDE.md for build presets and dependencies.
Functional CI and performance evidence are separate. See GUIDE.md and LAB_RULES.md / EVIDENCE.md in stra-ta/.github for manifest provenance and the one-command suite (./scripts/verify.sh / ./scripts/confidence.sh or tools/verify.sh).
CI is functional only. Performance evidence requires a committed manifest with machine metadata (commit, compiler, kernel, CPU, arch, build type, seed, argv) and a link from the claim to that artifact. See stra-ta/.github for lab-wide caveats.