We think intelligence is less about capability than the governance of belief — holding a belief, revising it correctly when the evidence changes, and proving the change. Today's models mostly regenerate rather than revise: they re-assert beliefs they were told to drop. We don't claim to have solved this; we build the parts that are deterministic, and we measure the rest.
We started from the smallest corner that's provable today and shipped it as two pieces of one program:
- sagrada-linter — see it bite: catch the AI coding rules you already changed before they break your build.
- er1-spec — see the receipt it emits: a proof a stranger can recompute offline, in any language, without trusting us.
Everything we publish leads with something you can run and ends on a problem we haven't solved yet. — Mars Ausili