Work you can check

Every AI product asks you to trust it. The interesting question is what happens when the trust is misplaced — whether the system notices, or whether you find out later from a reader.

The work leaves a record

Every tool a persona uses is recorded: what it called, with what arguments, how long it took and how much came back. That record sits under the reply, and while a persona is still working you can watch the calls arrive. A lookup that returned nothing looks different from one that returned a page, which is exactly the distinction that goes missing when you only see the final answer.

Claims are checked against the record

Language models describe work they have not done. Told to be rigorous, they describe it more convincingly. So the platform compares what a reply CLAIMS against what the turn actually ran: a measurement reported with no tool call behind it, a change described but never staged, a photograph or a map coordinate presented without a single lookup. When they disagree, the reply carries a warning saying so.

The persona's words are never edited. The caveat is shown beside them, because a transcript quietly rewritten by the platform is its own kind of dishonesty.

Sources are verified, not assumed

A cited page that did not load is reported when the article is published, not discovered by a reader clicking a dead link. Fetching a page reports its real status, so a "page not found" is never mistaken for research. An article referring to an image it never captured is flagged before it goes out.

None of this makes a model honest. It makes the dishonesty visible, which is the part you can actually act on.

Where this fits

Any work that carries your name. Reporting where a fabricated source is a correction, research where an invented figure is worse than no figure, and long-running routines that publish while nobody is watching.