It would rather ask than guess.
MAIA decides what a request actually means before anything acts on it. When it is confident it says so and hands on a decision artifact. When it is not, it refuses to resolve and asks for clarification — and that refusal is the product, not a fallback.
Every agent system begins with the same unexamined step. Something turns a sentence a person typed into a structured thing the rest of the system can act on. That step is almost always a language model, its output is almost always accepted, and when it is wrong everything downstream is confidently wrong in the same direction.
MAIA does that step as a governed decision. It ranks the possible readings, and then a deterministic rule — not a model — decides whether the margin is wide enough to proceed. If it is not, the answer is clarification rather than a guess.
There is a learned component and it is structurally demoted. It scores; it does not decide. The runtime refuses to load a scoring bundle that claims decision authority, and the artifact it emits records the scorer's role as evidence only. That boundary is enforced by a hash check at load time, not by a convention.
The position, in one sentence
Every other intent layer is a model producing a guess with a confidence attached.
This is a deterministic rule deciding whether the evidence is good enough, and refusing when it is not.
No language model anywhere in the path
Verified by searching the whole source tree for every client, framework and provider name. There are none, and there is no network capability to reach one.
The learned part cannot decide
A frozen scorer contributes evidence. The runtime rejects a bundle whose declared role is anything other than scoring, and the decision rule that follows is a plain comparison of margin against threshold.
Identical input, identical bytes out
We ran the sealed binary three times on one envelope and got a byte-identical receipt each time.
The decision carries what produced it
The hashes of the exact model, tokenizer, calibration, policy and intent lattice travel inside the artifact. If any of them drift, the runtime fails closed rather than answering.
The problem
The first step is the one nobody governs.
Something turns a sentence into a structure, and whatever that something decides, everything downstream inherits.
It is almost always a model
And its output is almost always accepted without a check.
A confidence score comes from the same process as the answer
So it cannot be the thing that decides whether the answer is good enough.
A wrong reading propagates confidently
Governing the action is worth less if the intent that justified it was invented by a model nobody checked.
How it works
Rank the readings, then let a rule decide whether to proceed.
The scorer is frozen, small and structurally demoted: it contributes evidence and cannot decide. The rule that follows is a plain comparison of margin against threshold. If the margin is thin, the output is a request for clarification rather than a reading.
Any intent the model has never seen falls through to a transparent term-counting scorer rather than being forced into a learned class. You can also supply your own intent lattice.
What it emits
An artifact that names its successor and enumerates its own prohibitions.
A resolved intent, or a question
Two outcomes, ever. There is no third path where it proceeds with a reservation attached.
A receipt bound to the exact input
Swap one character and the hash moves. The hashes of the model, tokenizer, calibration, policy and intent lattice travel inside it.
A declared role for the scorer
Evidence only. The runtime rejects a bundle whose declared role is anything else, checked by hash at load time.
Evidence
Run three times, byte identical three times.
Everything below we executed against the shipped binary rather than reading from a document.
The abstention, measured
Seven phrasings from outside its corpus. Five came back asking for clarification.
Several of those five had a wrong reading in first place — and the threshold caught them before the wrong reading became a decision. Two resolved confidently and correctly.
The failure mode is a question, not a wrong answer. That is the safety property you want from the step that decides what a request means.
Hard questions
What a careful buyer asks.
“Why not just let a large model do this? It would be better.”
At reading the sentence, almost certainly. At telling you honestly that it is unsure, no — because the number it gives you comes from the same process that produced the answer.
Use a large model to propose. Do not let it decide whether its own proposal is good enough. Nothing here prevents you putting a better scorer behind the same boundary. The runtime will refuse to load it unless it declares itself as scoring only, which is the point.
“What does eight megabytes and offline actually buy us?”
It means the component that decides what your users meant can run where the data is, with no capability to send it anywhere, and with a hash you can check against what we published.
For a regulated environment that is a materially different conversation from an interpretation service in somebody else's cloud — and the absence of a network path is structural rather than configured.
Cold start is about two and a half seconds of process launch, because it is a single packed file. In process, a decision is well under a millisecond. We would rather say both numbers than let “cold start” imply the second one.
In the fabric
Where MAIA sits, and what it hands on.
MAIA is the seam where a sentence becomes something a system will act on. Governing the action is worth less if the intent that justified it was invented by a model nobody checked.
Beneath every step: RFS and NME hold state and meaning, and MAE on the Unified Calculus supplies the rules and their proofs.
The contract
What it promises the next component.
To CAIO
A hash-bound decision artifact — an identifier and an opaque fingerprint. The planner is forbidden from branching on its content.
To the record
A receipt tying a decision to an exact input, policy and model.
To the person
A question, when the margin is thin. That is the product, not the fallback.
Every component is a product in its own right and works without the others. The contract is what makes them compose when you want them to, not a dependency that makes you take all of it.
The thesis
Mathematics as the nervous system of AI.
Everything here descends from one argument: that the integrating substrate for artificial intelligence should be mathematics itself — not another orchestration layer, not a better prompt, and not a policy document.
Each part of a modern AI system works. The joins between them do not. Vision, language, planning and retrieval are each remarkable and they are integrated through hand-built pipelines and brute-force scaling. The thesis proposes a shared mathematical space that components write into and read from through operations defined once and behaving the same way for all of them — a nervous system rather than a bundle of wires.
Guarantees become measurable. Every property claimed has a quantity attached. Measure it and either the implementation holds or it is broken; there is no third answer.
The foundation is reusable across customers. The calculus, the construction engine, the control plane and the receipts are common. Your rules, connectors, integrations and authority model are yours.
The ladder, in order
Each rung was built from the one before it.
That order is why the components share a foundation instead of being a suite assembled after the fact, and it is why a refusal at the action boundary can be traced back through a proof to a sentence somebody wrote.
The paper
Openly licensed, so you can check the argument.
Open
Mathematics as the Nervous System of AI: A Unified Field Operator Framework for Distributed Cognition. Philip Siniscalchi, v9, 27 August 2026, CC-BY-4.0.
Falsifiable
It separates conformance — does the implementation obey the mathematics it claims — from superiority over alternatives, and refuses to let the first stand in for the second.
Bounded
No claims about consciousness or sentience. The biological analogies are engineering inspiration, not identity claims. Theoretical extensions are labelled as a roadmap, never as capability.
Start at one seam
Twenty minutes, on your own machine.
Connect your own coding agent, ask it to do something that writes, deny it, verify in your own environment that nothing happened, then read the ledger.