OpenFluke
w2a
w2a is our honesty check. It's the testing harness that measures Welvet and publishes the results, so you never have to just trust a marketing claim. If something only half-works, w2a says so.
Why a separate proof tool?
It's easy for software to claim it's fast and correct. w2a exists to actually check. It runs Welvet across every combination of settings, times how it performs, and stamps each result as fully working, partly working, or not yet done. Keeping this outside the engine means the tests can be brutally honest, and the engine itself stays clean and small.
Measured, not guessed
Timed runs across every numeric type, quant, and backend: real numbers, not vibes.
Honesty stamps
Each result is labelled: done, gap, or fail. Gaps are declared — not counted as complete.
Public logs
Suite reports ship as downloadable release files, versioned to match the engine's scorecard.
PASS — good enough to ship
One end-to-end board: 0 fails, every suite case green. Gaps are honesty stamps (paths not claimed done) — mostly gdn / convt — not silent skips. Ready for Caber deploy and GitHub release.
| Layer | Cells | OK | GAP | Cases |
|---|---|---|---|---|
| step | 43,903 | 43,903 | 0 | 13 |
| tween | 32,457 | 32,457 | 0 | 10 |
| dna · evolution | ~16k ea | all OK | 0 | 6 ea |
| dense · mha · cnn* · rnn/lstm · norms · … | ~5.7k ea | ~all OK | 0–few | 14–25 ea |
| parallel | 6,415 | 6,369 | 46 | 20 |
| convt1–3 | 5,763 ea | 5,649 ea | 114 ea | 14 ea |
| gdn | 555 | 93 | 462 | 17 |
| Total | 228,054 | 227,194 | 860 | 456 |
Run it yourself
cd welvet/w2a go run . # interactive suites go test ./tests/dense -v
Suite logs (suite.txt / suite.pdf) ship as GitHub Release assets, versioned to the
Welvet scorecard, not committed into git. The engine packages never contain tests; the proofs live here.