AI does not replace a specialist — it multiplies one. Deigma turns that into something you can show: a passport built from your real work, not from what you write about yourself.
GitHub for outcomes; a local CLI and editor extension for how you work with the model. No forms, no tests, nothing to fill in.
We read up to six months of finished work — pull requests, and the runs of commits of anyone who works on main. For an active developer the first claim is proven before anything else is installed.
A claim appears when its own interval is tight enough. Nothing unproven is published anywhere — not on the profile, not on the badge.
A résumé is a claim. A certificate tests theory. An interview is forty-five minutes of an artificial task — and most companies still ban AI during it, while the actual job is done with AI every day. Real skill lives in months of work, and nobody measures those months.
So companies buy tools, call themselves AI-first, and have no way to tell who inside them actually got stronger. That gap is what this is for.
We are asking you to run our software next to your own. That deserves a straight answer before anything else on this page.
Pull request lifecycle and timestamps, CI conclusions, review states, salted hashes, counts and timings of your sessions. That is the whole list — and it is generated from the schemas the software is built against, so the traffic audit page cannot drift from what actually ships.
The proxy is fail-open by design: if anything inside it breaks, your call passes straight through and the measurement is dropped. Your tools working matters more than our data.
Leaving takes one click. Deleting the account destroys the key your rows are encrypted with — backups included — and you get the audit event that recorded it.
Each claim is measured on its own evidence and published only when it is measured well enough. Every value carries the range it was measured within — because a number without it would be a guess wearing a suit.
An example passport — the anatomy is real, the person is not.
The set differs by how you work. Someone who delegates to an agent proves different claims than someone who accepts suggestions in an editor — and the passport says which way of working it measured, instead of penalising you for the one you chose.
Not the value, not a hint of it. A separate suite tries to pull an unproven number out through parameters, caches, link previews and address guessing — including reconstructing it from the width of an interval.
Private by default. A private profile answers exactly like a handle that never existed — same body, same timing. A one-time link opens once, for one recruiter, and expires.
No webhooks or exports triggered by anyone's score. No ranked lists through the API. A passport is evidence for a conversation, not a filter that runs without you.
Claims about how you work are published only when real work stands behind them. An empty profile with rich telemetry is unreachable by design.
A latent-ability model from the same family used to calibrate GRE, PISA and medical licensing exams. Task difficulty is inferred from data, so a run of easy work cannot buy a score.
Every snapshot is signed. Anyone can check the signature offline against the published key — and it attests the numbers without exposing anybody's work.
The engine is validated computationally; external validation comes from a blind pilot that has not run yet. We say so here rather than after you find out.