Most AI systems are one undifferentiated model: a guess and a verified fact come out the same way, in the same voice. Noesis is built on the opposite principle — a strict separation where learned models may propose, but only a verified derivation carries authority. Two laws hold the whole thing together.
The two laws
1. Authority follows the validated artifact, never the substrate. A raw model output is telemetry — a candidate, not a conclusion. Authority climbs a sealed axis: Candidate < Evidence < VerifiedProof. A guess can never wear the authority of a proof, because only a verifier can mint one.
2. Nothing crosses a plane except through its typed seam. Information moves between parts of the system only through a checked interface. A step may summarize evidence, but it can never quietly erase it — so any answer traces back to what grounded it.
These two sentences are the safety story. Everything else is machinery that enforces them.
Four planes over a governed memory
Noesis separates concerns into four planes, each with a different job and a different level of authority:
- Learned models propose. The neural layer reads language, perceives, and suggests — and runs with low authority. Nothing it says is treated as true on its own.
- Verifiers alone mint proofs. Only the checking layer can promote a claim to "proven," and only when the derivation actually holds.
- Memory is governed. What the system treats as true is a versioned, audited store — not a scratchpad. Every fact carries its status, its source, and its history; a change is a confirmed step, not a side effect.
- Action is bounded and audited. Anything the system does is constrained and logged, so its effects are as inspectable as its answers.
A causal engine that knows when it can't answer
Real decisions are causal — will this intervention help; would this have happened otherwise. Noesis reasons across the three levels of causal questions (association, intervention, counterfactual), and — before it answers — checks whether a question is even answerable from what it knows.
When a safety-critical query is unidentifiable — the available knowledge can't support a causal answer — the engine refuses, rather than returning a number that looks exactly like a real one. Across our tested refusal battery it did so with zero false positives — every unidentifiable query correctly declined. That's a scoped, re-runnable result (we publish the evaluation), not an unscoped promise that it can never err.
Refusing the questions it can't ground is not a limitation we tolerate. It's the feature.
Domains as signed knowledge products
The core is domain-agnostic — it holds no medicine, no law, of its own. A domain loads as a signed knowledge product that carries its own proof of fitness: the facts, the rules, and the evidence they were checked against travel together, so the core never has to trust domain content it can't verify. Today that domain is medicine; the architecture doesn't care which body of knowledge it reasons over.
The honest part
Whether this exact combination — governed memory, a sealed proof axis, refusal by construction — is uniquely necessary, or measurably better than the alternatives, is not something outside researchers have independently studied yet. We believe the architecture is right, and we publish the scoped evidence and the open problems so the claim can be checked rather than taken on faith.
Where the ideas came from
Short pieces on the decisions behind the architecture — written as we made them, kept as we made them.
- Why we built neurosymbolic beings — the case for not choosing between symbols and neural nets
- One hundred columns — how we think about AI cognition, and why the count matters
- Teaching AI to be good before teaching it to be smart — the protected core, and why safety is a capability rather than a restriction
- The hallucination problem is a design problem — not a model problem, which changes what you can do about it
- Crystallization: teaching AI to remember — turning what a system read into what it knows
- What video games taught us about AI memory — an unlikely source for a memory architecture
- Why Santiago — what The Old Man and the Sea has to do with building a being