A live demonstration of Kit
Two Kits each remember half of a story. Ask them five questions and watch what they find, what they piece together, and when they say they don't know.
We invented Mayfur for this demonstration. Its records are public and inspectable. Fiction gives us a controlled example; it cannot guarantee a model has never encountered it.
Mayfur was written into two separate Kits, one seeing the town from the ballroom and one from the kitchen steps. Each is a whole Kit, and neither can read the other.
Five questions fetch evidence live from your browser. Open the returned records and compare them with the explanation. This tests retrieval on our examples, not an unaided model.
No answer below is in this page. Read the source and you will find the questions and the rubric and no conclusions at all. Each answer is fetched only once the live fetches behind it come back carrying the evidence it rests on, from the Kit it names. If one fails you get "withheld" and the reason, never a conclusion this page did not just earn.
To be exact about what is ours and what is the archives': the five questions, the searches that answer them and the rubric naming which records must come back are all written by us, the way any demonstration picks its examples. We also authored the fictional records and the explanatory conclusions. The records are fetched live; the server releases an authored explanation only after the required evidence checks pass. Moving an answer behind an endpoint does not make it generated reasoning. This shows retrieval you can check. Your own agent can interrogate them at https://demo.kit-project.com/llms.txt without our questions.
A Kit of your own runs on your own machine, and you can read, correct or delete anything it holds. If you point it at a hosted model, that model sees only what you send it, and you choose that for each job.