Candidate verification
Recorded September 20, 2026 — P8
These results describe the original snapshot, not the current source or release. Local artifact paths are provenance records, not public downloads.
Archive index · Current support
This unpublished 0.1.0 candidate contained 97 operations and 404 signatures. Full-platform acceptance was not achieved.
Recorded results
Apple M1 Pro, macOS arm64, 16 GiB; Node 24.19.0, pnpm 12.4.2, Rust 1.96.0. Node 22.13.0, 24.19.0 and 26.8.0 consumed the same archive without rebuilding native.
| Check | Historical result |
|---|---|
| Build, generation, format, lint/clippy and types | Passed; existing reductions benchmark max-depth warning only |
| TS tests | 1,877 passed on each of the three Node versions |
| Rust | 11 passed |
| Registry/evidence/candidate policy | Seven passed, including incomplete/stale/dirty/changed-evidence rejection |
| Signature smoke | 1,616 signature/mode combinations, 11 lengths, ordinary/resident; AVX2 not-run |
| MPFR accuracy and extension rehearsal | Passed |
| Tarball consumers | All three Node versions, declaration modes and missing-native fallback passed |
| Chromium / WebKit | 14 each, 28 total, including Vite and Worker lifecycle/response time |
| Firefox | 14 launch failures: Could not find profile folder, exitCode=1; no page assertions executed |
| Pipelines / Workers | 952 / 108 timing cases, plus ten response groups and 80 memory observations |
| Consistency | sourceUnchanged, assetsUnchanged and archiveUnchanged were true |
The overall candidate was failed because Firefox could not start. Four other native targets remained pending. Browser launch failures did not exercise page assertions.
Size and performance
The first archive was 398,539 compressed bytes and 1,004,007 unpacked bytes, 27 files: one local native addon (453,904 bytes), scalar Wasm (141,709), SIMD Wasm (147,764), JS/types and licenses. It excluded source/tests/AsyncTask and was not the final five-platform package size. Its SHA-256 was d7045570e9af3329baab797b419bdf60b5db6adf7359709e170e96e62139bc04.
First-candidate f32, 1,048,576-element mul/add/clamp medians in ms:
| Backend | Separate, reused output | Fused, reused output | Resident fusion | Resident end-to-end |
|---|---|---|---|---|
| native scalar | 4.847 | 0.797 | 0.772 | 1.594 |
| native NEON | 1.311 | 0.091 | 0.090 | 0.629 |
| Wasm scalar | 9.263 | 3.919 | 3.817 | 4.204 |
| Wasm SIMD128 | 1.742 | 0.307 | 0.122 | 0.367 |
The fair JS reused-output loop was 1.580 ms. Resident timings excluded transfer; end-to-end included the entire arena lifecycle. Three-node Worker total times were native scalar/NEON 1.724/0.452 ms and Wasm scalar/SIMD128 4.321/0.792 ms, slower than their synchronous equivalents. Large 64-node work improved main-thread response intervals without promising shorter completion times.
Evidence and remaining scope
The first candidate was artifacts/candidates/p8-2026-09-20/; the final recheck was artifacts/candidates/p8-final-2026-09-20/. Each retains its report, source archive and hashes. These records describe the old dirty workspace and old HEAD, not a later committed source tree. Preserve them as historical evidence.
The other four native targets, AVX2 execution, Firefox and kernel comparison remained open. Minimum-Node local evidence existed by this phase. Stable API/ maintenance commitments, versioning and publication were a separate process. Use the current candidate guide and roadmap.