Local-first verification · runs in your browser
Don't trust the number. Check it.
Four precision tools that compute in your own tab and prove their own work — nothing you load ever leaves your browser.
Four tools
Pick the one that fits what you're holding.
You bring: two .npy dumps
Two arrays that should match and don't. The first differing element, both bit patterns, the ulp distance.
Runs in your tab · unlimited
You bring: nothing
A real divergence caught on real model weights: the honest kernel vindicated, a fabricated one exposed, a signature failing on one tampered byte.
Runs in your tab · fixed scenarios
You bring: royalty statement exports
DistroKid, TuneCore, ASCAP, the MLC and more — merged into one Excel earnings tape, tied out to the cent against each statement's own totals.
Runs in your tab · 120 free pages / month
You bring: a system to document
Register a system, attach documents, seal a pack, share a read-only link the other side verifies itself. One domain is measured; seven are labelled sample.
Account needed · what it assesses →
Scope, plainly
What runs today — and what doesn't.
The four tools above are live and run in your browser. Everything else is scope, not product: only RMSNorm — the operator behind the array and referee tools — exists today. Custom work (parity reports, first-divergence post-mortems, tolerance audits, fleet-agreement reports) is scoped with you; nothing sits on a shelf.
- “Which side is correct?” — not a function of two output arrays.
- Whether a gap is legal reordering rather than a bug — that needs the operation, which two dumps don’t carry.
- A pass/fail CI gate on a divergence — same reason.
- A signature read as a verdict — a valid signature proves the bytes are unaltered; it is not a pass, and a failed check seals just as validly.
Something diverged that these tools do not cover?
We take on custom work and consultation on numerical divergence.