Public evidence for trustworthy AI.

Initiated by Digital One ↗, the enterprise-AI company behind the Digital One AI Suite ↗, the Foundation will operate independently: work is contributed by practitioners, published under open licences, and governed by a board on which no single interest holds a majority.

Formation by notarial deed in Amsterdam is underway. The statutes, Dutch Chamber of Commerce (KVK) registration and governance details will be published on this page — the first entries in the Foundation's own public record.

For formation, governance or partnership enquiries:

A dedicated foundation address will follow once registration completes.

The first working notes have been contributed to the Foundation’s public record. The Project Handshake is a series on tamper-evident record-keeping — the constructions, the reasoning behind each choice, and the defects found along the way, published under an open licence so that they can be examined by people who owe us nothing.

Six papers: the threat model that places the operator inside the adversary; a Merkle construction for records that survive redaction, and the encoding defect that made an earlier version of it silently wrong; the derivation of a signed preimage from the replay attacks it has to survive; an honest account of what verification cannot see; how governance, liability evidence and agent authority compose as three separate layers over one shared specification; and what a revocation cannot recall, once custody of an action has passed.

A second series has been contributed to the record. The Text Layer is working notes on what a language model is actually handed when it is given a document — what conversion to plain text keeps, what it loses, how a saving should be measured before it is claimed, and what still lies beyond the text layer.

Four papers: the method and the measurement; the number that had to be corrected before it could be published; the open research on the pictures, drawings and plans that no text layer describes; and where converted text should live, and what repetition makes compressible.

A third series joins the record. The Delegation Record is working notes on what a switch in the request path can attest from its own position — which tenant, which route, which host it dialled — and what it merely receives from an upstream and writes down because it arrived.

Eight short papers, in the order an operator meets the question: a step is handed to a model from another provider so that the harness’s default model is not paid for on work that does not need it. The saving is the reason to route, and the evidence for a saving is which model did the work and what it cost — which is precisely what a component reading plaintext in the request path cannot establish.

A fourth series joins the record. The Measured Endpoint is working notes on scoring live model endpoints: who judges free-form output when the judge is itself a model, and what a probe fleet’s signed results do and do not make evidence.

Four papers, each part of the method set beside its limit. A panel drawn from several providers dilutes a judge’s preference without removing it; a signature binds a result to a key, not to the world. Neither paper publishes a rubric, a prompt set or a panel, so neither lets a reader re-run a score — and both say so before anyone asks.