NAO

Public research layer · Experimental autonomous-agent research

Building autonomous intelligence for uncertain environments.

An experimental autonomous-agent architecture designed to perceive, reason, doubt, remember, anticipate and evaluate its own decisions.

Gold is the first proving ground. Autonomous intelligence is the real project.

Explore NAOView interface

Official public home: nao.co

Decision trace · illustrativeNon-live
PERCEIVEREASONDOUBTREMEMBERANTICIPATEEVALUATEt

Schematic representation of the agent loop. Not performance data.

Built in public · Observable intelligence · Governed autonomy

Cognitive model

How NAO thinks

NAO separates what is observed, what is interpreted, and what remains uncertain — then places each decision in time.

  1. 01

    I SEE

    Observed facts

    What the system can directly establish from its environment at a given moment.

  2. 02

    I THINK

    Interpretation

    The reading NAO builds from available facts, without confusing interpretation with certainty.

  3. 03

    I DON’T KNOW

    Uncertainty / limits

    What remains incomplete, contradictory or insufficiently demonstrated stays explicitly unresolved.

  4. 04

    PAST

    Recent context

  5. 05

    PRESENT

    Current state

  6. 06

    FUTURE +30m

    Possible scenario

    Never presented as certainty.

Perceive before acting. Preserve uncertainty. Separate facts, interpretation and decision.

Governed autonomy

Acting cautiously under uncertainty

For NAO, more autonomy must never mean less control. Observation, interpretation, authorization and action remain distinct.

  1. 01

    OBSERVE

    Establish what is present.

  2. 02

    INTERPRET

    Build a reading without confusing it with a fact.

  3. 03

    AUTHORIZE

    Check that action is permitted and sufficiently coherent.

  4. 04

    ACT

    Execute only when the required conditions are met.

BeliefProposalAuthorizationAction

Wait is a valid decision

When information remains incomplete, contradictory or too uncertain, NAO can remain in a waiting state instead of forcing an action.

No action can be the most coherent response.

  • Uncertainty is preserved, not hidden.

  • An interpretation is not an authorization.

  • Autonomy develops progressively and under control.

More intelligence should not imply less governance.

Memory & self-evaluation

NAO does not only decide. It compares what it expected with what actually happened.

A useful autonomous agent needs continuity. NAO is being developed to preserve prior hypotheses, observed outcomes, contradictions and unresolved questions so that later evaluation does not rewrite the past.

Observed outcome≠Retroactive justification

  1. 01Remember

    Preserve prior state and context.

  2. 02Compare

    Confront expectation with observed outcome.

  3. 03Evaluate

    Identify error, contradiction or remaining uncertainty.

  4. 04Revise

    Update the working interpretation without rewriting history.

In development

Persistent memory and descriptive replay exist in the broader NAO research stack; deeper calibration and governed cognitive adaptation remain under development.

Learning should improve the quality of reasoning before it changes the authority to act.

Architecture

One cognitive core. Multiple domains.

NAO separates general cognition from domain-specific knowledge. The core handles uncertainty, memory, self-evaluation and governed decision-making; specialized adapters connect that core to a particular environment.

NAO CoreDomain adapterEnvironment

The application uses NAO. It does not define NAO.

Transferability remains a research direction until demonstrated across additional real domains.

Read the architecture note →

NAO CoreGeneric cognition
  • EHUD++
  • Epistemic state
  • Memory
  • Self model
  • Contradictions
  • ACT / WAIT / BLOCK
  • Governance
  • Outcome / revision

EHUD++ is NAO’s coherence discipline: it weighs factual coherence, systemic coherence, robustness and adaptability against uncertainty, contradiction and risk.

X(t) = {facts, hypotheses, uncertainty, risk, coherence, contradictions, actions}

Domain adapters

  • Gold— First operational proving ground

    Observe market context, regimes and uncertainty in a live, consequential environment.

  • Bank— Future domain direction

    Compare offers against real user constraints, not price alone. Best offer ≠ cheapest offer.

  • Travel— Future domain direction

    Compare flights, hotels, timing, baggage and transfers as one coherent journey.

  • Shopping— Future domain direction

    Compare products across price, characteristics, availability, warranty and user preferences.

  • Personal— Future domain direction

    Tasks, projects, preferences and time constraints under explicit authorization.

Future domains are research directions. They are not available products and not demonstrated transferability.

Agentic loop

NAO Core + Domain Adapters

A useful agent is a coherent loop, not a stack of features.

NAO is designed around a bounded cycle: perceive what is present, represent what is known and unknown, reason under uncertainty, check authority, specialize through a domain adapter, observe the outcome and revise without rewriting history.

  1. Perceive

    Receive observations without confusing them with conclusions.

  2. Represent

    Separate facts, hypotheses, unknowns, contradictions and current state.

  3. Reason

    Compare competing interpretations under uncertainty.

  4. Govern

    Keep belief, proposal, authorization and action distinct.

  5. Specialize

    Connect the Core to a specific environment through a Domain Adapter.

  6. Observe outcome

    Compare what was expected with what actually followed.

  7. Revise

    Update the working model while preserving the original evidence and decision trace.

Efficiency comes from coherence, not feature count.

WAIT and BLOCK are valid outcomes when the loop is incomplete or authority is insufficient.

First proving ground

Gold is where NAO learns to operate under uncertainty.

The gold market provides a real, dynamic and noisy environment where timing, incomplete information, changing regimes and the cost of mistakes matter. It is the first operational proving ground for NAO’s broader autonomous-agent research: the first application of NAO Core in a complex, dynamic environment — not NAO’s final identity.

Research environment — not a performance claim

Operating environment · illustrativeNon-live
SIG/NOISEREGIME?

State markers and regime cues only. Not market data.

01Dynamic environment
A real market with moving, noisy conditions.
02Incomplete information
Signals are partial and arrive with delay.
03Regime changes
The rules of the environment shift over time.
04Cost of mistakes
Acting wrongly has concrete, measurable consequences.

Current status

Built in public. Developed in controlled stages.

Demonstrated today
Observable cognitive states, temporal framing, governed decision logic and responsive research interfaces.
In development
Deeper self-evaluation, memory, anticipation and progressively richer autonomous behaviour.
Long-term research direction
Transferable agent architecture for other complex and uncertain environments.

NAO interface

Private access — in preparation

A dedicated interface will make it possible to follow NAO’s observations, cognitive state and qualified projections.

Interface coming soon