Government will buy AI agents and regulate them. There is no accredited way to test one.
A car has an MOT; a medical device has a notified body. An AI agent that reads patient records or decides on a benefit claim has only its vendor's description. We are building the missing test — independent, re-runnable, admissible — in Scotland, with a consortium of UK universities.
A model vendor can copy a check. It cannot make the check independent. So the method — how a run is scored, what the error bounds are, which behaviours count as malicious — is being placed with universities, not with us. Edinburgh Napier holds the TRL 4 record and, after a ministerial meeting in August, the mandate to coordinate. Xplore supplies the instrument and the traces.
- 1Procurement. A test any bidder's agent can be put through before award — the same test, re-run on every version.
- 2Oversight. A record of what each agent saw and did, kept apart from what it claimed, that an auditor can read without a language model.
- 3Compliance. EU AI Act record-keeping, ISO/IEC 42001, NIST AI RMF — met by evidence, not by a supplier's attestation.
Xplore is a member of Scotland's Critical Technologies Supercluster and supported by Interface Scotland. UK patent priority 2025.
Xplore Intelligence Ltd · Edinburgh · hello@xploreintelligence.co.uk · Stand: Innovation Alley, Tue 15 – Wed 16 Sep