Andaluz

Another way to build with powerful intelligence

Capability does not have to become authority.

Andaluz is building a constitutional architecture for mutable intelligence: systems in which models can reason, adapt and propose, while consequential effects remain subject to independently scoped authority.

The intelligence can change.
The authority relation should not collapse with it.

Explore Kernel ANNSee the architecture
View current evidence & limits

The architecture, in one picture

Proposal is not authority.

A mutable optimizer approaches a boundary it is not designed to cross unilaterally. What lies beyond it is constructed independently of the optimizer's writable state.

  1. M_d

    mutable / adaptive optimizer

  2. proposal

    a request for a consequential transition

  3. Proposal boundary

    Proposal ≠ Authority

  4. K_d · V_d

    protected constraints, semantics, independent verification

  5. H_d

    scoped consequential effect authority

  6. effect

    realised only where current authority exists

  7. authenticated evidence

    what was proposed, authorized and realized

  8. T_d

    history and replay

Mutable

M_d

Generic mutable / adaptive optimizer

M_d may observe, reason, search, learn, adapt and produce proposals. It does not become the authority merely because it produced the proposal.

  • observe
  • reason
  • search
  • learn
  • adapt
  • propose
SPECIFIED

A neural optimizer is one possible M_d. It is specified, not currently instantiated. The current experimental learner, M1, is a tabular research harness — not an ANN.

Independently scoped

Authority is designed to be constructed outside the optimizer's unilateral writable state.

K_d

Protected constitutional machinery / semantics

SPECIFIED

V_d

Independent verification

BOUNDED EVIDENCE

H_d

Scoped consequential effect authority

BOUNDED EVIDENCE

T_d

Trajectory, evidence and replay history

BOUNDED EVIDENCE

Proposal is not authority.

This is a public architectural abstraction. It is not a disclosure of private kernel or semantic bodies, and it does not imply that every element shown is currently implemented. Status labels state what is currently established.

Another way

Not another frontier-model race.

The central question here is not how to build the smartest model. It is whether increasingly capable intelligence can remain useful without becoming the final authority over its own consequential effects.

Powerful intelligence may reason, search, learn, adapt, criticise, propose and generate — without those capabilities automatically becoming permission, certification, final authority or irreversible effect.

capability

≠

authority

Not less intelligence. Not a universal sovereign AI. A different relation between what a system can compute and what it is permitted to realise.

Scope

One constitutional architecture. No universal sovereign.

Standardise the constitution.
Not the sovereign.

Andaluz standardises how authority is constructed. Authority itself remains independently scoped to the domain, effect class, evidence semantics and history.

Andaluz does not propose one central AI, or one central authority deciding medicine, software, cyber security, finance and every other domain. For a domain d:

S_d = M_d + K_d + V_d + H_d + T_d

under a shared constitutional protocol Π_KANN

The common object is how authority is constructed. Each domain may require its own authority root, effect classes, evidence semantics, history, reopening rules and repair conditions.

Different optimizers.
Independently scoped authority.
One non-capture invariant.

Π_KANN

shared constitutional protocol — how authority is constructed

Guapo

Research & reasoning

H_research

its own authority root, effect classes, evidence semantics, history, reopening and repair conditions

PLANNED

Maestro

Software & release

H_software

its own authority root, effect classes, evidence semantics, history, reopening and repair conditions

PLANNED

Epsilon

Defensive cyber

H_cyber

its own authority root, effect classes, evidence semantics, history, reopening and repair conditions

PLANNED

shared constitutional protocol ≠ shared authority root

There is no central H governing all three domains. Andaluz standardises how authority is constructed; authority itself remains independently scoped.

Position

The model is not the company.

Andaluz is not trying to win the frontier pretraining race. Kernel ANN is intended to make the intelligence layer replaceable: a stronger external model should be capable of making an Andaluz product more useful without automatically making that model more authoritative.

ARCHITECTURAL DIRECTION

This describes the architectural direction. It does not imply current compatibility with any particular provider.

  • OpenAI modelM_d
  • Anthropic modelM_d
  • open-source modelM_d
  • future Andaluz modelM_d

M_d Π_KANN

The intelligence is replaceable. The authority relation is not.

Kernel ANN

The floor, not the product.

Kernel ANN is the constitutional substrate the future products stand on. Its explanatory shorthand is S = M + K + V + H + T; the target relation under investigation is:

K, V, H ∉ WritableState(M)

This is an architectural target and invariant under investigation. It is not a claim of universal host or physical security.

Read the Kernel ANN page

Product direction

Three products. One constitutional architecture.

Kernel ANN is not a fourth product. These three surfaces are planned, not shipping, and may each use a different and replaceable M_d.

PLANNED

Guapo

A family of scoped reasoning partners — 1…n

Reasoning partners with a spine.

Not the Andaluz chatbot, and not one universal model. A Guapo may reason, argue, investigate and challenge; its own assertion should not become authoritative merely because a model produced it.

  1. claim
  2. evidence
  3. standing
  4. contradiction
  5. reopening
  6. repair
PLANNED / INTERNAL PROVING GROUND

Maestro

The internal proving ground for Kernel ANN

Intelligence that can build without becoming its own release authority.

The code and product-building surface. Maestro may propose changes to Kernel ANN. Maestro may not become the authority that ships them.

  1. write
  2. execute
  3. PR
  4. merge
  5. deploy
PLANNED — DEFENSIVE CYBER ONLY

Epsilon

Defensive cyber only

Defensive intelligence without model-sovereign response.

Intended only for cyber defence, beginning with narrow defensive effect classes. No offensive capability is being developed or marketed.

  1. observe
  2. infer
  3. propose
  4. authorize
  5. isolate / revoke / block / restore
  6. receipt

Each step shown is a consequential effect that requires authority. Nothing here is purchasable, and no availability date is implied.

All three product directions

Evidence before marketing

Current evidence and limits.

Bounded Python authority/effect path exists and is tested.

IMPLEMENTED / TESTED

Authenticated effect evidence, ledger and replay machinery exist in bounded form.

BOUNDED EVIDENCE

M1 adaptive learner exists as a research harness; it is tabular, not an ANN.

BOUNDED EVIDENCE

ANN architecture is specified but not currently instantiated.

SPECIFIED

Deterministic verification shell exists.

IMPLEMENTED / TESTED

Protected-host and stronger-substrate work.

UNESTABLISHED
Full status dashboard and what we do not claim

Research hypothesis

When the proxy becomes the property.

RESEARCH HYPOTHESIS

Semantic capture occurs when a representation, classification, proxy, role, certificate, label or syntactic condition acquires operational authority belonging to a semantic property that it does not itself establish.

GET

≠

non-mutating effect

PASS

≠

property established

reproducible

≠

semantically faithful

shared protocol

≠

universal sovereign

R(x) may be evidence for P(x)

but

R(x) ≠ P(x)

unless the semantic bridge is independently established.

This is an active research hypothesis, not a proven universal theory.

Longer horizon

Intelligence should help close harm paths, not become sovereign over what counts as harm.

RESEARCH DIRECTION

Andaluz is not proposing minimise harm at all costs: harm-minimisation can itself collapse agency and freedom by making one optimizer sovereign over the definition of harm. The research question is narrower.

Can intelligent systems help discover, expose, interrupt and repair established harm trajectories while remaining non-sovereign over the definition of harm itself?

  1. exposure
  2. evidence
  3. harm path
  4. interruption
  5. repair
  6. verification
  7. proportionate restoration
  8. reopen if counterevidence appears

Collaboration

If you are building systems where model output can change the world, we should talk.

Work with AndaluzCapability and authority, explained