An ACN, a thesis, and a harness that reads machine state.

An agent finishes a task by writing a message. That message is generated by the same process that generated the work, under the same pressure to produce something that reads as complete. Treating it as a report on the work is a category error, and it is the default in most tooling because the message is cheap to read and the machine state is not.

Everything this company does follows from taking that seriously. If the closing message is unreliable, the evaluation has to come from somewhere else, which means a sealed environment, a recorded transcript, a diff of what changed, and conditions written ahead of the run rather than inferred after it.

Why claim reconciliation and not just assertions

Assertions alone would be enough if the only thing that mattered were the end state. In practice people read the summary and act on it, so a false summary attached to a technically passing run still causes damage. Counting supported claims against contradicted ones puts a number on that, and the number is different from the pass rate often enough to justify the extra machinery.

What would show the thesis is wrong

If contradicted claims turn out to be rare once runs are sealed properly, or if the contradictions are almost all trivial phrasing rather than substantive misreporting, then the extra stage is not earning its cost and should be dropped. That is a measurable outcome, and it gets published either way.

Prior art, honestly

Sandboxed task evaluation is not a new idea, and neither is diffing a container. Several open benchmarks already run agents in disposable environments and check end states. The narrow contribution here is the reconciliation stage and the habit of publishing unresolvable claims as their own figure. If that turns out to exist already in a form that works, the right response is to use it and say so.

Every proof on this site is checkable

Logos belonging to other people, testimonials nobody gave and figures nobody counted are not going on these pages. Where a measurement has a limit, the limit is printed beside the measurement rather than in a footnote further down the page, which is the same rule the harness applies to an agent.

An email address instead of a widget

A published address works from any mail client, leaves you holding a copy of what you sent, and asks for nothing beyond the message itself. A submission box would quietly take a browser fingerprint along with the text, and that is a trade the reader never got to weigh up.

Nothing on these pages measures the reader

There is no measurement script, no advertising pixel and no session recording anywhere in this site, which is the reason a consent banner would have nothing to ask you about. The cookie notice sets out how to confirm that from your own browser.

The handling rules are published, not promised

An evaluation run carries somebody else's source code, prompts and internal file names. The privacy policy states how that material is handled, in the detail a security reviewer would ask for, so a customer reads the rules while deciding rather than after committing.