Aria

The model is disposable.

Aria is not a model. She is a control plane, the layer that decides which model does what, verifies the result against a different one, and refuses to let any of them take an action that cannot be undone.

ArchitectureCognitive control plane
ModelsAny. None named in the codebase.
VerificationCross-family, deterministic-first
StateExternal, versioned, replayable

The architecture

Five laws. Everything else is an implementation detail.

01

The model is disposable

Any model can be swapped for any other without destroying system state. Nothing in our codebase names a model. Every component requests a capability, and one configuration file decides who fills it at runtime. A build test fails if anybody hardcodes a name.

02

State is external

Conversation history is a rendering of state, never the state itself. The model proposes a transition; the controller commits it. This is what makes a run reproducible, and it is why task drift cannot occur, the objective is not stored anywhere that can drift.

03

Verification is ordered

Physics, then execution, then deterministic rules, then database constraints, then an independent model, and only then a model checking itself. Arithmetic is verified by arithmetic. Code is verified by running it. A model judging a model is the last resort, not the default.

04

Confidence cannot authorize

An irreversible action requires authorization, not certainty. Ninety-seven percent confidence does not send an email, delete a file, or spend money. Irreversibility is a property of the action, not of the belief.

05

Compute follows risk

A trivial question gets one model. A contested one gets five, plus retrieval, plus a deterministic validator, plus an adversarial verifier. The advantage is not more agents. It is spending compute where it changes the answer.

The control plane

A task enters once and is answered by however many models the risk justifies.

ROUTER CLASSIFIES RISK ROUTINE · 1 AGENT CONTESTED · 4 AGENTS AGENT retrieval AGENTretrieval AGENTreasoning AGENTplanning AGENTperception VERIFIER family I VERIFIER family I VERIFIER family II DIFFERENT LINEAGE SYNTHESISER ANSWER + CITATION AUTHORISATION, NOT CONFIDENCE IRREVERSIBLE ACTION ROLLBACK

Hover or focus any box to read what it owns, and what it explicitly does not.

The control plane owns

  • state
  • identity
  • provenance
  • authority
  • scheduling
  • retrieval
  • verification
  • rollback
  • risk classification
  • routing
  • observability

The models own

  • inference
  • generation
  • planning proposals
  • specialized reasoning
  • perception
  • hypothesis formation
Where this stops

The split above is a design commitment, not a proof. Nothing in the architecture stops a model from being asked for something in the left column, what stops it is that the controller does not offer the interface, and an interface can be added by anyone with commit access.

So the guarantee is exactly as strong as the build test that enforces it. That test is public, it runs on every commit, and if it is ever deleted this claim goes with it.

Correlated error

Two models that agree may be agreeing for the same wrong reason.

Models trained on overlapping data share failure modes. When a verifier comes from the same family as the generator, it agrees with a wrong answer far more often than chance, and it returns that agreement as a verification signal.

A swarm that is confidently and unanimously wrong is worse than a single model being wrong, because the error now arrives with a confidence stamp on it.

So our router will not place two models from the same family in consecutive links of a verification chain. We measure the correlation. We publish it.

4.0% Verified, and wrong
1.0× Against independent verifiers

Each model alone is wrong one time in five. At ρ = 0 they fail independently and the pair is wrong together 4% of the time. At ρ = 0.98 they are very nearly the same model, and almost every one of the generator's errors survives its own verification.

Where this stops

The figure above is a model, not a measurement. It shows what correlation does to a verification chain; it does not tell you what ρ actually is between any two production models.

That number is the one we are running the benchmark to get, and it is the row marked Verifier correlation, cross-family in the table. Until it is filled in, nothing on this page claims to know it.

The long memory

She does not summarize you.

Most systems compress old conversations into summaries, which loses the detail, or stuff the whole history into context, which costs a fortune.

Aria indexes every message by time, by entity, and by meaning. When you refer to something from two years ago she resolves it to the specific messages and injects only those, a few hundred tokens instead of a few million.

You get the exact words back, with the date.

Where this stops

Anything you marked private is session-scoped and never enters the index. She cannot recall a thing you asked her not to keep, because it was never written down.

Constraints

What she will not do.

  • No irreversible action without your explicit confirmation.
  • No claim without a source that can be checked.
  • No inference from how fast you type or how long you pause.
  • No storing what you asked her not to store.
  • No answer that sounds confident when it is not.
  • No rewriting of her own safety constraints: those are outside her reach, permanently, by construction.

The benchmark

We publish the methodology with the result.

Every row links to how it was produced. When a number changes, the date changes with it, and the previous value stays in the history rather than being overwritten.

The benchmark has not been run.
MeasureResultMethodUpdated
Cost vs. monolithic frontier call
Uplift, 2-year-old base model
Verifier correlation (ρ), cross-family
Verifier correlation (ρ), same-family
Reservoir hit rate
Cost per active student / month

Not yet measured. We are running the first full benchmark. We will publish the methodology whether or not we like the result.

The Constitution

Ninety-seven thousand words about what she is not allowed to do.

Aria's governing documents specify her behaviour under crisis, her handling of minors, what she may infer and what she may not, and what happens to anything a student tells her in confidence.

They include the parts we got wrong in earlier versions and how we corrected them.