Machines that think with receipts.
Every answer arrives with the grounds it rests on. What cannot be shown is refused, by name.
Open the appNo model gives the answer.
The native engine runs no language model. In the app, a model may propose a reading of a question you type, and nothing it says becomes an answer. Answers are computed from your rows and refused where they cannot be shown.
Not yet is an answer.
Where the record is incomplete, it says so and names what is missing. A blank is never read as no.
Checked by a stranger, where marked.
The part that checks an answer did not compute it and does not trust the part that did. In the app, whole-source counts and the active-account definition pass that check. Every other figure is computed from your rows and arrives with a receipt of what it used, not a certificate.
Written in your sentences.
Policy goes in as plain language and comes back as something a person can read, line by line, before it runs.
Check ten lines yourself.
Every verdict names the file and the line it rests on. Pick any ten, open the file, compare. Nobody has to trust us and nobody has to read the code.