AI you can stand behind.
Every claim an agent writes is checked against the record behind it, and every action it takes against who authorised it, without a model in the loop. The first is ProveML and Vera; the second is PAC and Elixir. A person releases it to your customers. The checks are what let it run further.
eight years on digital trust, agents its newest test
What runs.
ProveML
Checks what an AI writes against your data and your thresholds, claim by claim. Open source.
PAC
Sets what an agent may decide on its own, earned from a track record rather than configured.
Elixir
Governs a fleet of client systems under those rules today, and acts only as far as its own record allows.
Where it comes from.
Eight years as Web3 lead at Howest’s Cyber3Lab: research and implementations on EBSI and verifiable credentials;Hyperledger Fabric since 2021 and Besu, certified on both; the2025 LF Decentralized Trust Community Award. The question then is the question now: how a claim can be believed by someone who was not there when it was made. ProveML’s trust adapter reads a record as a signed credential, so it is one answer, not two.
What we build against.
Confidently wrong
The answer that reads perfectly and is not true. Every claim is bound to its measurement and checked.
Doing more than asked
Credentials used for whatever the agent decides the goal is. Access is scoped to the task and expires with it.
Failing quietly
A stopped job looks like a calm week. Systems have to prove they are alive.
Drifting from intent
Between what you asked and what it did there has to be a record.
The digital identity world has reached the first half of this. At Global Digital Collaboration 2026 in Geneva, one session in seven was about agents: who they are, who authorised them, what they may pay. What they then claim is the next question, the one we work on with the Advanced AI Society’s Proof-of-Control work and the Personal Data Portability Protocol.