Andaluz

Evidence before marketing

Current research status.

This page states what currently exists, what is only specified, what is under test and what remains unestablished. Earlier theoretical and formal work establishes the research lineage; experimental claims for Kernel ANN depend only on the evidence produced by the Kernel ANN experimental programme.

Dated status

Research status — 26 September 2026

Kernel ANN is an experimental research architecture investigating whether useful adaptive optimisation can remain separate from the authority required to realise protected effects.

The latest milestone is an executed, integrated prototype connecting source verification, precisely specified structural computation, journal admission and current authorisation to a bounded effect lifecycle using memory writer and observer adapters.

The prototype records operation intent, consumption of one-use permission, the writer's reported outcome, observation and retained history separately. In the tested duplicate and recovery paths, the same operation does not obtain a new writer attempt.

This is a reviewed experimental candidate, not an adopted protected deployment. The latest P3-06 candidate received a separate AI-assisted review of source and retained execution evidence. That review did not rerun the recorded suite. Independent human or institutional validation of the current composition has not been established.

What “adapter” means here

The lifecycle code runs, but the consequential target and its observation are represented through controlled memory-based components rather than the planned native protected-file operation. The milestone is executed code, not merely a specification — and it is not deployed native enforcement.

Implementation status

Component progress, by stage.

Implementation, execution, review, adoption and deployment qualification are different stages. Engineering progress is not the protected-environment guarantee.

  • P3-01

    Offline wire / identity contracts

    Accepted offline contracts. Not deployment qualification.

    ACCEPTED — OFFLINE CONTRACT
  • P3-02

    Substrate / build preparation

    Partial preparation, including recorded compiler-fixture execution. A complete builder and protected-target qualification remain outstanding.

    PARTIAL
  • P3-03

    Journal / recovery correction

    Implemented, exercised and used downstream. The candidate remains unadopted; physical durability is unqualified.

    IMPLEMENTED / EXERCISED — NOT ADOPTED
  • P3-04

    Source verification and fixed structural computation

    Implemented and used downstream. Checks a precisely defined structural property; not general semantic truth.

    IMPLEMENTED / EXERCISED
  • P3-05

    Binding / current-authorisation correction

    Implemented, exercised and used by P3-06. Actual deployment roles and credentials remain unqualified.

    IMPLEMENTED / EXERCISED — NOT ADOPTED
  • P3-06

    Integrated adapter lifecycle

    Executed integrated adapter lifecycle, separately AI-reviewed. Candidate not adopted.

    EXECUTED / AI-REVIEWED — NOT ADOPTED
  • P3-07 / P3-08

    Actual enrollment and complete protected composition

    Not established.

    UNESTABLISHED
  • Later

    Deployment assessment and separately reproduced release

    Future work.

    FUTURE WORK

Current executed adapter composition

EXECUTED / AI-REVIEWED — NOT ADOPTED
  1. Test proposal
  2. Source verification
  3. Structural computation
  4. Journal admission
  5. Current-authorisation checks
  6. One-use adapter operation
  7. Observation and retained history

Explanatory view, not a complete state machine. Recovery, refusal and uncertainty are not permission to continue.

Native protected-file effect boundaryPLANNED / NOT YET QUALIFIED

The current composition stops before the planned native protected-file effect boundary. File writes used to store test journals do not themselves demonstrate that protected effect primitive.

Separate evidence strands

Three experiments, not one system.

  • Earlier M1 experiment

    HISTORICAL

    DemonstratedDemonstrated bounded adaptive tabular learning under changing authority conditions.

    BoundaryM1 is not an ANN. It is not the proposer integrated into the current P3-06 demonstration.

  • Historical Windows pipe-effect demonstration

    HISTORICAL

    DemonstratedDemonstrated a bounded real-process path involving a fixed accepted source, structural checking, current permission, an actual pipe event and supervisor capture.

    BoundaryGenuine historical native-effect evidence, but not the newer P3-06 native lifecycle or protected-deployment qualification.

  • Current P3-06 composition

    EXECUTED / AI-REVIEWED — NOT ADOPTED

    DemonstratedIntegrates fresh-source checking, structural computation, admission, current authorisation, memory-adapter effect, observation and journal/recovery history.

    BoundaryIts proposer is a deterministic test fixture. An adaptive learner or neural model has not been demonstrated inside this current composition.

These are separate experiments. Their strongest features have not yet been demonstrated together in one integrated system.

Limitations

What the current evidence does not show.

  • The implemented source criterion checks a precisely defined structural property of supplied bytes.
  • Correct computation does not establish that those bytes capture the real-world requirement.
  • Current writer/observer adapters do not establish independently controlled native observation.
  • Simulated anchors and same-device storage do not establish hardware-backed nonrollback protection or power-loss durability.
  • Useful toy-task completion is not demonstrated practical usefulness across applications.
  • No comparative advantage over an appropriate simpler control has yet been established.
  • No general AI safety, frontier-model containment, universal non-bypass, production readiness or instantiated ANN is claimed.
  • AI-assisted review is not independent human peer review, external certification or separate institutional validation.

Negative findings remain welcome: simpler mechanisms may suffice, stronger assumptions may be necessary, or parts of the proposed architecture may prove unnecessary.

Next milestone

PREPARATION

Next: prepare and review a bounded native private-file effect implementation, including descriptor custody, current authority, one-attempt handling, separately obtained observation and explicit recovery/uncertainty behaviour.

The first bounded adapter milestone has been reported as achieved. This does not mean the complete architecture or a protected deployment is finished. No completion date is promised.

External scrutiny

Written technical criticism and scoped external evaluation are welcome, particularly on authority boundaries, recovery under uncertainty, requirement-to-evidence correspondence and comparisons with simpler approaches.

contact@sergiumargan.com

Written correspondence by email only.

Status dashboard

Current bounded facts.

Bounded evidence exists

Small, bounded experiments with adversarial testing and AI-assisted review, scoped to the individual records described here. The first two entries are historical: these entries describe earlier bounded implementations. The current fresh-source evaluator and P3-06 adapter composition are reported separately. Earlier test results do not automatically validate the newer composition.

  • Historical bounded Python authority/effect path.IMPLEMENTED / TESTED
  • Historical deterministic verification shell.IMPLEMENTED / TESTED
  • Authenticated effect evidence, ledger and replay machinery exist in bounded form.BOUNDED EVIDENCE
  • M1 adaptive learner exists as a research harness (not integrated into P3-06).BOUNDED EVIDENCE
  • M1 is tabular, not an ANN.BOUNDED EVIDENCE
  • Integrated adapter lifecycle (P3-06) executed with memory writer/observer adapters; separate AI-assisted source and retained-evidence review, which did not rerun the recorded suite. Candidate not adopted.EXECUTED / AI-REVIEWED — NOT ADOPTED

Specified or under test

Written down or actively reviewed. Specification is not evidence.

  • ANN architecture is specified but not currently instantiated.SPECIFIED
  • Private semantic and kernel bodies are not publicly instantiated in this build.SPECIFIED
  • Systematic-closure research remains under adversarial specification review.UNDER TEST

Not established

Open. No claim of any result is made here.

  • Substrate/build preparation is partial (P3-02). Actual enrollment, complete protected composition and native protected effect (P3-07/P3-08) are not established.UNESTABLISHED
  • Systematic-closure research is not upgraded by these implementation milestones.UNDER TEST

A green status means implemented or bounded evidence. It never means safe, validated or production-ready. Private implementation bodies are not exposed; a document marked ready is not externally validated; a test double is not semantic truth; an ANN in a diagram is not an implemented neural architecture.

The experimental programmeGlossary of terms

Review provenance

Multi-model, AI-assisted review

Kernel ANN's research and review workflow has involved multiple AI models, under Sergiu Margan's direction. Models used for review across the programme include:

  • Claude Code with Opus 5.5 — High
  • GPT-6 Astra — Ultra
  • Daybreak — Ultra
  • Grok 4.6 — Expert

Model names and effort settings are reported by the project author.

The workflow uses separate review passes to challenge specifications, inspect implementation and retained execution evidence, identify defects, and scrutinise the scope of research claims. Earlier failures and subsequent corrections remain distinguishable in the research record.

These are programme-level reviewer credits. They do not mean that every listed model reviewed every component, that each reviewer reran the experiments, or that all reviewers reached the same conclusion.

Multi-model AI review is not independent human peer review, institutional certification or endorsement by the model providers.

Planes

Runtime architecture is not research verification.

Runtime authority plane

  • Target architecture: specifiedSPECIFIED
  • Bounded adapter composition: implemented and executedIMPLEMENTED / TESTED

The constitutional path a proposal must cross before it becomes an effect: proposal boundary, K_d, V_d, H_d, authenticated evidence, T_d.

P3-06 exercises a bounded proposal-to-effect lifecycle through memory writer and observer adapters. This is partial realisation of the architectural design, not qualification of the complete protected runtime. The current candidate remains unadopted.

NON-OPERATIONAL

Research / specification verification plane

Proofs, specification review, adversarial reading and reproduction work. This plane authorizes nothing at runtime and grants no effect authority. Progress here is not evidence that a deployed system is safe.

Claim discipline

What we do not claim.

  • Not proven general AI safety.
  • Not AGI safety solved.
  • Not universally secure.
  • Not production-ready.
  • Not proof of C2/C3/C4.
  • Not proof that K1–K4 are necessary or unique.
  • Not proof against arbitrary host, OS or physical compromise.
  • Not a Transformer replacement.

Method

How a claim is allowed to move.

  1. conjecture
  2. explicit proposition
  3. implementation candidate
  4. adversarial review
  5. failure preservation
  6. repair
  7. independent reproduction
  8. stronger test

A failed version stays failed.

Criteria are not weakened after negative evidence. Only public artefact and status progression is shown.

Provenance

Research lineage.

Intellectual and technical provenance — not a proof chain and not a validation chain. Other migrated works appear as related formal, architectural or authority research rather than links in this sequence.

  1. TRO

    The Redemption Optimization

    Foundational work in the research lineage that later led to MKO, closure/kernel research and Kernel ANN.

  2. MKO

    Moral Kernel Optimization

    Non-bypassable constraint structure: preservation, anti-domination, repair.

  3. Closure

    Closure and kernel work

    Margan Closure Theorem, kernel transfer and Coherence Wall results; Ground as an operational closure audit layer.

  4. Kernel ANN

    Protected effect authority

    Architectural separation of a mutable optimizer from the authority governing consequential effects.

  5. M1…

    Bounded experimental programme

    Small, bounded experiments with adversarial testing and AI-assisted review, scoped to the individual records described here.

The number of prior research records does not strengthen any Kernel ANN claim.

Research indexConvergence Science context

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 request

≠

non-mutating effect

PASS label

≠

property established

reproducible result

≠

semantically faithful result

shared constitutional protocol

≠

universal authority root

R(x) is evidence for P(x) does not imply R(x) == P(x)

This is an active research hypothesis, not an established theorem or a solved universal theory.

Open questions

What could falsify or dissolve this work.

  1. 01How much of the architecture is already captured by classical runtime assurance?
  2. 02Which assumptions fail under adaptive adversarial attack?
  3. 03Can complete mediation be maintained in realistic agentic environments?
  4. 04What is the correct trusted computing base?
  5. 05Does the architecture preserve useful agency at greater scale?
  6. 06Do bounded results survive across different optimizers and environments?
EARLY CONCEPTUAL WORK — NOT FORMALIZED

Love and freedom as structural notions

Love here is not sentiment. Tentative idea: love preserves another as an irreducible bearer of standing rather than collapsing the other into the actor's objective.

Freedom is not simply the number of available immediate actions. Tentative idea: freedom concerns preserving the possibility for independent future agency.

An illustrative question: emergency treatment of an unconscious patient, where bounded temporary authority may instantiate rather than violate freedom when its purpose and endpoint are restoration of the patient's capacity for agency.

Not settled mathematics. Not a product principle.

Longer horizon

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

The long-term direction asks whether systems can discover consequential trajectories, expose latent failure paths, preserve evidence, interrupt established harm paths, repair realized harm and reopen previous conclusions under counterevidence — without giving the optimizer final authority over the meaning of harm, evidence, closure or repair. This is a research direction.