1. Define what must be true
Share the AI result to check. We establish the object, site, version, intended use, references and required qualifications, then agree the quote.
Field audits of AI resultsYour AI said it.
Your AI said it.
An expert checks it on site.
An AI result, a real site or object, and a qualified expert: we establish the facts, correct errors and retain traceable evidence.
Your audit in three steps.
2. Check against reality
A qualified expert compares the AI’s factual claims with the real site or object, identifies errors and omissions, proposes corrections and collects evidence.
3. Report and make it verifiable
You receive the findings, severity, corrections and limits. A reference and public QR verification record identify the report using the information agreed with you.
A functional benchmark grounded in reality.
The benchmark provides a reference: what the AI claims is compared with what the expert can observe and document. A common framework applies across sectors, supported by appropriate professional references.
Five axes guide the audit: factual accuracy, completeness, error severity, evidence traceability and regulatory alignment within the examined scope. A counter-verification can then examine a corrected result in comparable conditions.
The Hu-score: four levels of reliability.
The score concerns the original AI result checked, its version and intended use, on the audit date.
- 4/4 — Reliability supported within the verified scope. No identified discrepancy and sufficient evidence.
- 3/4 — Reliability with minor corrections. Limited discrepancies without changing the professional conclusion.
- 2/4 — Insufficient reliability. A major error or omission requires substantial correction.
- 1/4 — Critical error. An error or omission may lead to a serious consequence in the intended use.
If a determining fact remains unverified, no overall Hu-score is awarded. Demonstrated alerts remain visible. Corrections do not retrospectively change the initial score: checking them produces a separate finding.
The Hu-score describes checked facts; it does not guarantee every future output of an application. It is not a certification or a general validation of safety or compliance.
Useful evidence, controlled disclosure.
The minimal public record, after its content has been authorised, may display the report reference and version, publishable scope, audit and issue dates, the responsible expert’s professional name and role, protocol version, report status, Hu-score when one could be assigned, PDF hash and linked version references. Client documents, evidence photographs, private contact details and sensitive information remain in the private file for authorised recipients. The QR code does not grant access to that file.
In healthcare, checks are factual and documentary, performed by an appropriately qualified professional. Huxpert never validates a medical decision.
Recognised reference frameworks.
The approach draws on NIST AI RMF 1.0 and its generative AI profile: testing, evaluation, verification and validation, human oversight and attention to confabulation. The EU AI Act provides reference points for accuracy, traceability and human oversight. ISO/IEC 42001 contributes a complementary AI management perspective.
These references do not mean that Huxpert is certified or endorsed by their authors. Applicable requirements depend on the result and context examined.
Further reading.
Your AI said it. Let’s check it on site.
The five-question pre-diagnostic prepares the audit. It does not award a reliability score. A sample PDF report is available on request.