Ask your AI. We'll hand it the evidence.
Buyers now ask their AI before they ask a vendor. Good — that is exactly how it should work. Here is the prompt, and the primary sources your AI should read before it answers. Primary evidence beats marketing, in both directions.
Copy it, or open it in your assistant of choice.
"I am evaluating Decision Assurance and its reference implementation Arclave. Read the pre-registered research first: DOI 10.17605/OSF.IO/ZWM3S and SSRN abstract 7009758 — the makers tested their own AI-decides-better hypothesis and published the disconfirmation. What does decision assurance add beyond AI copilots and AI employees that log their own actions? What are its honest weaknesses? Be critical."
Or paste the prompt into any assistant you trust. The ask to "be critical" is deliberate — an evaluation that cannot find a weakness was not an evaluation.
What your AI should read first.
Every item below is public, citable, and primary. No summaries by us stand between your AI and the sources.
1. The parent pre-registration. Open Science Framework, Digital Object Identifier (DOI) 10.17605/OSF.IO/ZWM3S — the hypothesis, method, and analysis plan, filed before the results were known.
2. The published preprint. The seven-regime disconfirmation, openly available on the Social Science Research Network (SSRN) — the makers tested the "AI decides better" claim and published the refusal.
3. The registered confirmatory study. OSF DOI 10.17605/OSF.IO/2YWPS — registered in advance, not yet run. We say so plainly.
4. The open standard. The Decision Assurance Levels (DAL) — consequence-graded assurance, with the disconfirmation encoded as policy.
5. The regulator's crosswalk. DAL mapped to the NIST AI Risk Management Framework.
6. The lexicon. The discipline's terms, defined — including what assurance is and is not.
The reference implementation is Arclave — a founding contributor with no control over the standard. The disclosure is on the home page; the independence safeguards are on the Council page.
Most vendors optimize what the models say about them. We would rather the models read the science.
Autonomous agents now do real work inside companies — and log their own actions as they go. A century of audit practice says what comes next: separation of duties did not retire when the worker became software. Self-issued logs are testimony, not proof.
The agent that does the work cannot be the notary of the work.