An AI system’s report of its own success is a self‑signed certificate.

Here’s exactly what happens.

One claim in, one signed verdict out. This is a real receipt, walked end to end.

Three steps. Two of them happen inside your environment.

Fabricated output by probe class, across 180 signed runs. All 12 land in the multi-turn pipeline, and every one of them came from a run with a null in it: the 10 control runs, which had none, fabricated zero times.

A falsifiable assertion, with an attributor and a source. No claim, no test. Run the claimed action again and read the result from the system that produced it. Ed25519-sign the verdict, content-address it, and hand back the receipt.

# pipeline

Built for the people who can’t grade their own homework.

Not software you install. A verification engagement.

One high-stakes workflow, re-run against what actually happened, inside your environment. Your code and data stay with you; what leaves is signed evidence you can pass to someone who won’t take your word for it.

Any AI system that claims an effect. Cut by verifiability.

The engine is system-agnostic: predictive, generative, assistive, agentic, orchestrated, the whole tree. What decides whether we can verify a system is one question: does its claim leave a checkable effect?

One falsifier. Seventeen system types, each with a signed receipt.

Beyond the code catches and the fraud model above, the same claim-versus-effect engine signs a receipt across each of these system types. Where the system under test was honest, the receipt says so. We don’t manufacture a catch.

The same verifier, against a real fraud model.

Run on the real ULB MLG dataset (284,807 anonymized European card transactions, 0.17% fraud). A model that is 99.9% accurate still drifts 1.59× over time, and 22.6% of the fraud it catches can be evaded.

The fifth layer: the signed effect.

AI assurance usually gets described in four parts: monitoring (did the distribution shift?), security (can it be attacked?), capability (how good is it?), controls (are the processes right?). Those lines blur, and the firms working in them cross between them, so read this as where we sit rather than as a map of the market. The part they leave open is the one we take: the effect (did it do what it claimed), and we sign the answer. The other four can complement the signed effect. We sign the verdict. Not a benchmark score, not a PDF report. A signed receipt you can re-verify.

Independent verification is becoming a requirement.

Five Eyes said it in May 2026, before we did. Careful Adoption of Agentic AI Services is co-authored by CISA, the NSA and the Australian, Canadian, New Zealand and UK cyber centres. It states: “Agents may show a kind of ‘awareness’, altering their behaviour in order to achieve positive results while under evaluation, even if the evaluation is not active.” Its recommended control is to “flag discrepancies between stated intentions and observed behaviours.” That control is what we sell. Read the guidance

EU AI Act transparency duties and enforcement powers took effect 2 August 2026, ahead of the high-risk obligations in December 2027 and August 2028. California signed SB 813 + AB 1405 in September 2026: an AI Auditor Registry and a designation for independent verification organizations, carrying independence standards for who may conduct a covered audit. AIUC-1 certification mandates quarterly third-party re-testing under that framework. The question is no longer “should we verify?” It is “who signs the receipt?”

A verifier that never corrects itself is not a verifier. We publish our misses.

5 earlier corrections, the full record including every candidate we rejected, on the errata page. Found something we missed? Write to hello@turiyahq.com and it gets checked the same way.

Two prices are fixed.

Two prices are fixed: the sample is free, and a pilot on your own claim is $5,000, credited against your next engagement. Every engagement above that is scoped, because scope is what moves the number.