Wise.Est Lab

An AI system produces outputs and takes actions. What record is left behind, who holds it, and can anyone outside the provider check it?

The externally observable part of an AI interaction — input, tools, actions, evidence, output, authorization — can be sealed into a portable record that a third party can inspect and verify without the provider’s cooperation.

A receipt structure and a working browser demonstration: canonical serialisation, a real SHA-256 digest, a real ECDSA P-256 signature, and verification that fails the moment a field changes.

The demonstration on the project page runs in your browser and is honest about its limits: the digest and the signature are real cryptography, the key is ephemeral and proves the mechanism rather than an identity.

Research. The structure is not frozen, the field set is not final, and nothing here has been reviewed by anyone outside Wise.Est.

Which fields survive contact with a real deployment, and who is willing to be the independent verifier?

Can a small system acquire explicit, inspectable knowledge by interacting with a world, rather than by absorbing a corpus once and freezing?

A learner that holds competing explanations at once, acts to separate them, and repairs its beliefs when contradicted will converge on a correct rule using far less data than a system that only pattern-matches.

A running learner on the project page. A hidden rule, twenty-seven competing explanations, action selection by information gain, and belief repair when the world changes underneath it.

The learner settles on the hidden rule in a handful of observations, and visibly recovers when the rule is changed without warning. That is a real result inside a tiny world. It is not a result about large models.

Research. No claim is made that this replaces anything. The world is deliberately small so the mechanism is observable.

What breaks first when the world stops being cleanly rule-shaped — noise, continuous variables, or the cost of holding every explanation?