Announcing Aver, the first AI safety research project at Nicholic

Aver: two ledgers that are never allowed to merge is published on the research page. It is the first note there.

Aver is an assistant that shows where an answer came from and what is disputed about it. Underneath it sits Touchstone, a credibility engine holding two ledgers that stay separate: a falsifiable record of a source's accuracy, built only from things that can be checked, and a map of how public communities receive that source. The second is displayed to the reader and never acted upon. It does not gate retrieval, filtering or ranking, and the note sets out why that separation is the whole design rather than a detail of it.

The safety argument is the part worth reading. An assistant that is fluent and confidently wrong about a dose or a deadline can injure someone, which is why the note describes tiering by the consequence of being wrong rather than by who might object to the answer. A system that quietly withholds sources under pressure fails in the opposite direction, invisibly, and looks like working software the entire time. Both are safety failures, and the note argues that removing the reason to manipulate a credibility score is a better design than building one and then defending it.

Nothing in it has been run. The design is written and a first draft of the code is written; no question has been put to it and no answer has come back. That is stated in the note's second paragraph rather than left for a reader to discover, and every safety property it describes is written as intent rather than as a claim — because a control that is not switched on yet is not one this company will describe as real.

That is also the reason for publishing at this stage. A commitment written down before there is anything to lose by writing it can be checked later against whatever actually gets built.