A slow kernel
should never reach main.
Performance regressions ship silently: a dependency bump degrades a hot kernel, nobody notices, and production latency spikes weeks later. Kernalin puts a statistically honest performance gate on every pull request — and fails the build before the regression merges.
Why regressions slip through today
Noise fakes a pass
A naive “within 5%” check lets a real slowdown hide inside run-to-run jitter — and a lucky fast run hides a genuine regression.
No baseline you can trust
Comparing across mismatched GPUs, drivers, or workloads produces confident nonsense. Most CI perf checks don’t even try to control for it.
Correctness and speed are checked separately
A change can pass unit tests and still ship a fast-but-wrong kernel — or pass a perf check while silently corrupting output.
One gate: correct, faster, and proven — or it doesn’t merge.
GitHub Actions integration
Trigger the full verification + regression gate on every pull request; upload results automatically; fail the build on regression thresholds.
Tiered BLOCK / WARN / INFO / PASS
Every metric diff maps to one enforceable outcome. A credible regression on a gated metric is a hard BLOCK — it cannot merge.
Distributional, noise-floor-aware
A Bayesian ROPE verdict over a bootstrap posterior, judged against each GPU’s own jitter — so the gate can’t be fooled by noise in either direction.
Auditable metric-diff comment
Every verdict is backed by a stored, reproducible diff posted right on the PR. No black-box pass/fail.
It can’t get lucky, drift, or wave a regression through.
99.4% of the posterior mass sits below the ROPE — convincingly beyond what this GPU's jitter could fake. This one earns the number.
Make your benchmark impossible to lie to.
Put a verification-first performance gate on every pull request. Join the early-access program.