AI you can put under contract.

Noesis is a neurosymbolic runtime that answers only what it can prove from a governed memory, cites every premise, quarantines what it merely believes — and says so, out loud, when it doesn't know.

Most AI systems hand you an answer and leave you to guess how much to trust it. Noesis inverts that: every answer arrives wearing a contract that tells you exactly what kind of claim it is and what backs it. You don't calibrate trust by feel — you read it off the answer.


The four modes every answer wears

Each response is one of four kinds, and the kind is part of the answer, not a footnote:

The line between Speculative and Proven is the whole product. A guess and a verified fact never sit on the same footing, and a guess can't quietly become a proof.


Abstention is a feature, not a failure

The failure that matters in a clinic, a courtroom, or a compliance workflow is confident, ungrounded, and wrong. So Noesis is built to do the opposite of fill the gap:

"I don't have grounded facts about this."

When a chain can't be grounded, Noesis abstains — loudly — instead of returning a fluent guess that reads like an answer. We show the refusal screen on purpose, because a system that knows the edge of what it knows is worth more than one that never admits an edge exists. Absence stops masquerading as safety.


It asks before it remembers

A memory is only as trustworthy as what's allowed to change it. In Noesis, what the system treats as true doesn't get rewritten by a stray sentence: by design, a change to governed memory is a confirmed, auditable step, not a side effect. It asks before it remembers. The aim is that there is no silent corruption of what the system believes — and that you can always ask why a fact is there and get a source, not a vibe.


What we haven't finished

Honesty is the brand, so here is the edge of it. The modes, the citations, and the loud abstention are how the system behaves today. The hardest guarantees — sealing proof authority so that only a verifier can ever mint a "Proven," and making every memory write fail closed when a warrant is missing — are engineering we are still hardening, for us as much as for anyone building governed memory. We are not done, and we would rather say that than pretend a guess can never slip through anywhere in a whole system today. That is the direction we build toward, and we will publish the proof when it is real — not the week a trend peaks.


Proven first in medicine

We hold Noesis to the hardest bar we could find: medicine, where a confident wrong answer is worse than no answer. On the provable path it does not fabricate — it proves, cites, or abstains — and we publish the scoped, gated evidence for that rather than asserting it.

The core is domain-agnostic. Today it proves itself in medicine; the architecture doesn't care which body of knowledge it reasons over. Load a different one and the same guarantees apply — the modes, the citations, the abstention, the governed memory travel with it.

(See the evidence on the homepage and the research page for the scoped, dated claims behind this — every number links to a re-runnable artifact.)


Runs on hardware you own

For the domains Noesis is built for, the important answers often can't leave the building. So the direction we're building toward is an appliance that runs on hardware you own — your data, your premises, your audit trail, on your side of the wall. We'll publish the specifics when the hardware envelope is real and measured, not before.


Join the pilot waitlist

Noesis is in early pilots with teams whose answers have to be right and have to be checkable. If that's you, tell us the decision you can't afford to get wrong.

Join the pilot waitlist →

Bring your hardest cases. The offer is that you check the guarantee yourself.