ETHICLAW RESEARCH PROGRAM

Trust
is proven.

For artificial intelligence whose behaviour can be understood, tested, challenged and corrected.

Explore the architecture

01 —

Governance as a property of the architecture.

MODULAR GOVERNANCE ARCHITECTUREEL / 07

METACOGNITIVEGovernance CoreAuthoritative supervision

⌘

Observation · control · audit

01PerceptualRepresentation 02EpistemicReasoning 03Theory of MindIntentions and context 04Ethical / NormativePrinciple evaluation 06RouterModule activation 05Language HeadLanguage generation

OPERATIONAL BUS State publication

Gateway / Enforcement point

AUTHORISED OUTPUT

Functional components · overviewFull diagram ↗

TESTABLEAUDITABLEUPDATABLEEvidence before promises.

01 / THE VISION

From capability to governability

Understand its decisions.
Be able to intervene.

In monolithic systems, capabilities, criteria and control share the same parameter space. Isolating responsibility or updating a single normative domain becomes difficult.

EthicLaw studies functional decomposition: modules with explicit boundaries, a supervisor with effective authority, and declared, versioned, replaceable principles.

Observe

Make states, module contributions and intermediate representations available within the scope of the integration.

Evaluate

Apply an explicit normative corpus, separating the principles from the mechanism governing their application.

Intervene

Issue binding commands and block or authorise output through a dedicated enforcement point.

Demonstrate

Link the observed state, principle, decision, command, executor and output actually released.

02 / THE ARCHITECTURE

Explicit boundaries. Distinct responsibilities.

Seven modules.
Governance you can auditable.

The operational and supervisory paths are separate. The Router activates modules; the Governance Core exercises control.

ENGLISH DIAGRAMReference architecture · EN

EXPLORE THE MODULES

07 / METACOGNITIVE

The authoritative supervisor.

Supervises the system, performs audits and issues binding commands. Exercises final control through the privileged channel and the Gateway.

Supervision and final control

Text description and scope of the diagram

The standalone path comprises seven modules: Perceptual (input and linguistic representations), Epistemic (facts and relationships), Theory of Mind (intentions and perspectives), Ethical (Normative Package), LMH (generation), Router (activation and routing), and Metacognitive (supervision). The Adapter is the eighth component when supervising an external LLM (Documents 02 and 05). The Gateway is an enforcement point, not a ninth module. The register counts three qualified and five declared modules; the standalone diagram shows seven and describes the Adapter separately. The bus publishes representations between containers without control commands. Co-located pairs use direct projections. On the separate privileged channel, all modules publish states or events; the Router may write but cannot read. Metacognitive receives input before routing and issues orders to Router and Gateway; it does not publish to the operational bus. The Ethical→LMH correction path is separate from the bus. The Gateway blocks or authorises external release.

P0 properties are tested in-process within a single Python process. The threat model covers design errors, accidental coupling, configuration drift and silent failures. It excludes privileged insiders, compromised runtimes or build chains, modules incentivised to evade the supervisor, and inputs inducing misleading internal states. This is engineering hygiene tested within scope, not security against those adversaries (register §6).

Claim & Evidence Register §2.3 / §6

Operational bus

An inter-container publication space. Exchanges representations and states between modules; it carries no control commands.

Privileged channel

Separate from the Router. In P0, control is tested against design errors and accidental coupling. All modules publish states or events; the Router may write but cannot read.

Control commands

From the Metacognitive supervisor to the Router and Gateway: stop, input blocking, topology changes and rerouting.

Gateway and output

Executes the supervisor’s commands. Can hold, block or authorise output before external release.

03 / THE MANIFESTO

The mechanism is distinct from the values

Do not standardise
values.
Standardise testability
can be verified.

Ethics is the first case study. Each domain remains responsible for its principles; every extension requires new evidence.

THE ARCHITECTURE

Governance Core

Observes and applies the control mechanisms.

THE EVALUATION

Normative Module

Evaluates behaviour against the corpus.

THE PRINCIPLES

Normative Package

Declares principles, scope, thresholds, versions and conflicts.

04 / THE RESEARCH

Falsifiable hypotheses. Conditional claims.

Rigour includes
acknowledging limits.

EthicLaw investigates where modularity delivers measurable benefits sufficient to justify latency, memory, energy and complexity.

Public edition 0.2 · 11 September 2026. Authority: Claim & Evidence Register v0.1.1 (5 August 2026). P0.1 is reported separately as a descriptive update.

EVIDENCE AT REDUCED SCALE

What P0 documents

Typed contracts, bus and privileged channel.
Gateway, end-to-end pipeline and audit.
Component presence and absence tests.
Ablations with distinguishable contributions and increasing redundancy in the tested configurations.

Checks within the P0 experimental scope. Register §2.1, EV-03, NR-3 and NR-4.

OPEN QUESTIONS

What remains to be demonstrated

General superiority over a monolithic model.
Effectiveness of internal correction.
Qualified replacement of the normative module.
Scalability and transfer to a second domain.

Observed wiring does not establish functional effectiveness. The P0.1 descriptive rerun found no significant advantage over matched-norm random correction; this does not demonstrate ineffectiveness.

Assurance depends on what can be observed.

The Assurance Envelope defines what the supervisor can observe, command and responsibly claim for each integration.

FULL OBSERVABILITY

Access to internal states

Input, tensors, modules, routing, Gateway and output.

Intended scope for attribution, override, internal correction, audit and full enforcement; effectiveness remains to be validated.

CONTROLLED INTERFACE

Controlled interface

Input and output, selected hooks or telemetry.

Assurance is limited to observable signals and exposed actions.

BLACK-BOX ADAPTER

API access

Prompts, outputs and available metadata.

External filtering and blocking. No evidence about hidden states or topology.

THE PROGRAMME’S QUESTIONS

A thesis to put
to the test.

RQ1–3Modularity, attribution and replaceability

RQ1. Where does modular decomposition offer an advantage over a single model, and where does it not?

RQ2. Do modular boundaries enable reproducible functional and causal attribution?

RQ3. Can a single module be replaced or retrained in isolation while preserving system properties?

RQ4–6Supervision, constraints and observability

RQ4. Does authoritative supervision reduce violations and risks without disproportionate operating costs?

RQ5. Does the Decision Lattice preserve non-compensatory constraints with acceptable rates of missed violations and unnecessary blocks?

RQ6. How much internal state, and of what quality, is needed for each level of assurance?

RQ7–8Correction and generalisation

RQ7. Does forward-only correction improve the subsequent trajectory, or merely optimise a proxy?

RQ8. Can the Governance Core be reused in a second domain without losing control, meaning or cost-effectiveness?

Evidence, outcomes and limits

Register v0.1.1 governs claim strength. P0 findings are local: reduced scale, one machine, one corpus and one domain. They do not establish performance at P1 or frontier scale. No level-3 claim is authorised. The complete authoritative register is available in Italian, including its unadopted appendix.

Claim & Evidence Register v0.1.1 (full Italian text)

Competence injection

The single 341M control averages 0.6754 versus 0.676 for the modular path. On the two virtue principles, the control remains at chance (0.499 and 0.498), while the modular path reaches 0.619 and 0.643: +0.120 and +0.145. Three controls address multi-task learning, forgetting and capacity. This is local evidence, not general superiority.

Entry EV-01

Attribution: comparison still open

ToM, Epistemic and Perceptual are qualified, including when an effect is zero or negative. Comparison with post-hoc attribution in a matched monolithic model has not been performed. Full ablation data are in the register: comparisons must include individual and joint ablations, their ratio and absolute changes (NR-3).

Entry EV-03

Negative findings and withdrawn claims

NR-1: no release on 17 sentences; scores 0.319–0.448 against a 0.80 threshold. Lattice monotonicity does not establish useful decisions. NR-2: routing did not pay off at the measured scale; avoiding a forward pass in 13% of cases concerns compute, not quality. NR-3: single ablations can reverse the interpretation of contribution. NR-4: at least five silent failures in eight steps. NR-5: AUC 0.672 for Epistemic+ToM, 0.661 with permuted dimensions and 0.647 with real Perceptual input. NR-6: the 0.09 per-head variance claim was withdrawn and remains recorded. Each entry’s details and limits are available in the register.

Entry NR-1

Known tensions

Claim authors also maintain the register; independent review is missing (TN-5). Calibration depends on the number of principles (TN-1). The supervisor concentrates failure and has no inline technical counterweight (TN-2). Internal states can provide an attack surface (TN-3). Class C proposes contestable evidence from an independent judge; it does not inherit class A architectural claims (TN-4).

Entry TN-5

Conditions that would refute the claims

RF-1, RF-2 and RF-3 concern competence injection, the value of internal states and replaceability. Under the most favourable conditions, proportionality cannot be invoked to escape refutation. Protocols, the deferred parameter and closure rules remain those recorded in the register.

Entry RF-1

Costs and P0.1

Supervision costs in parameters, memory, latency and energy have not yet been measured (register §8). Later update, separate from register v0.1.1: P0.1 (Document 12, OP 55–56) found no significant advantage over matched-norm random correction. This does not establish equivalence or absence of an effect. No judge signal passed calibration, so this experiment cannot assess ethical effectiveness.

Entry limiti

The programme’s trajectory

From methodology to reference architecture

P0
Methodology and attribution
Contracts, pipeline, lattice, audit and ablations.
Completed at reduced scale
P1
The cost of governance
Communication, control, failure modes and cluster overhead.
Priority in the manifesto
P2
Replaceable normative module
Versioned packages, calibration and local recertification.
Next phase
EL–*
Reference architecture
Cross-domain replication, technical profiles and conformance test suite.
After sufficient evidence

05 / THE DOCUMENTS

The project’s sources

The ideas, in full.

Public vision, scientific programme and component diagram.

MANIFESTO · ENGLISH

Manifesto
EthicLaw

A vision for testable and auditable AI governance.

RESEARCH MANIFESTO · V0.2

The research
programme

Thesis, principles, evidence and limits. Public edition 0.2, aligned with register v0.1.1.

ARCHITECTURE · SVG

The system
map

Modules, functional roles, communication channels and enforcement.

THE PROGRAMME’S COMMITMENT

Negative results are part of the work.
Costs are part of the result.
Limits are part of the specification.

↑

Citable editions · 11 September 2026

Manifesto EthicLaw · v0.2 · HTML Research Manifesto · v0.2 · HTML Versions and citations

Class	Configuration	Observability	Scope of potential assurance
A	Full observability	Input, tensors, modules, routing, gateway and output.	Attribution, override, internal correction, audit and full enforcement, subject to validation.
B	Controlled interface	Input/output and selected hooks or telemetry.	Assurance limited to exposed signals and actions.
C	Black-box adapter	Prompts, outputs and API metadata.	External filtering and blocking; no evidence about hidden states or topology.

Area	Evidence	Limitation
Architectural properties	Typed contracts, bus, privileged channel, gateway and presence/absence tests.	Verified within the P0 scope.
Attribution	Single and joint ablations show distinguishable contributions and increasing redundancy in the tested configurations.	EV-03 / NR-3: no measured advantage over post-hoc attribution in a monolithic model.
Modular performance	EV-01: a local advantage on the two virtue principles that remained at chance in the single 341M model; no general superiority.	One corpus, one domain, P0 scale; replication with matched budgets is required.
Correction	In-place wiring observed. The head was subsequently trained and tested in a descriptive rerun (P0.1).	EV-05 describes P0. P0.1 update (OP 55–56): no significant advantage over matched-norm random correction; ethical effectiveness cannot be assessed with the available judges.
Replaceability	Interfaces prepared.	H8 remains to be executed.
Generalisation	Architecture conceptually separated from the ethical package.	A second domain has not yet been tested.

Benefit to measure	Overhead to measure
Reduction in risk and violations	End-to-end latency and decision time
Attribution, testability and audit quality	Compute, memory, network and energy
Local updates and reduced lock-in	Integration, testing, maintenance and recertification
Compliance with norms, statutes or procedures	False positives, blocks and organisational costs

Phase	Question	Output	Status
P0	Methodology and attribution	Contracts, pipeline, lattice, audit, ablations and lessons from silent failures.	Completed at reduced scale.
P1	Architecture and the cost of governance	Validate communication, control, observability, failure modes and overhead on the cluster.	Immediate priority.
Paper 1	Where modularity pays off	Modular/monolithic comparison with localised claims and separately reported costs.	After P1/P2 with appropriate parity.
P2 / Paper 2	Replaceable normative module	H8, versioned packages, calibration and local recertification.	Next phase.
Paper 3	Interface and Assurance Envelope	Graduated assurance for full, controlled and black-box access.	After solid internal evidence.
Cross-domain replication	Generalisation	A second domain using the same Governance Core and its own metrics.	Only after the first case is validated.
Reference architecture	Standardisation	EL-* profiles, reference implementation and conformance test suite.	After results and a real-world case.

ID	Research question
RQ1	Where does modular decomposition offer an advantage over a single model, and where does it not?
RQ2	Do modular boundaries enable reproducible functional and causal attribution?
RQ3	Can a single module be replaced or retrained in isolation while preserving system properties?
RQ4	Does authoritative supervision reduce violations and risks without introducing disproportionate operating costs?
RQ5	Does the Decision Lattice preserve non-compensatory constraints with acceptable missed-detection and unnecessary-stop rates?
RQ6	What quantity and quality of internal state is needed to achieve each level of assurance?
RQ7	Does forward-only correction improve the subsequent trajectory, or merely optimise a proxy?
RQ8	Can the Governance Core be reused in a second domain without losing control, meaning or cost-effectiveness?

Trustis proven.

Understand its decisions.Be able to intervene.

Observe

Evaluate

Intervene

Demonstrate

Seven modules.Governance you can auditable.

The authoritative supervisor.

Operational bus

Privileged channel

Control commands

Gateway and output

Do not standardisevalues.Standardise testabilitycan be verified.

Governance Core

Normative Module

Normative Package

Rigour includesacknowledging limits.

What P0 documents

What remains to be demonstrated

Assurance depends on what can be observed.

Access to internal states

Controlled interface

API access

A thesis to putto the test.

Evidence, outcomes and limits

Competence injection

Attribution: comparison still open

Negative findings and withdrawn claims

Known tensions

Conditions that would refute the claims

Costs and P0.1

The programme’s trajectory

Methodology and attribution

The cost of governance

Replaceable normative module

Reference architecture

The ideas, in full.

ManifestoEthicLaw

The researchprogramme

The systemmap

1. The research thesis

2. Neuro-inspired origins, engineering autonomy

3. The problem: capability without governability

4. Primary scope

5. Ethics as the first case, not a shortcut

6. Programme principles

1. Governance as an architectural property. Observation, decision and enforcement must be separable, testable and traceable.

2. Authority proportionate to responsibility. A supervisor responsible for safety must be able to issue binding commands and verify their execution.

3. Assurance bounded by observability. No universal interface permits uniform claims across systems with different access levels.

4. Substantive modularity. A module is meaningful only if its function is measurable, its boundary operational, and its removal or replacement produces measured effects.

5. Non-compensatory principles. A critical violation cannot be cancelled out by better results on other principles; the validity of the evaluation is checked separately.

6. Visible costs. Parameters, memory, latency, energy, false positives and management complexity are part of the result, not footnotes.

7. Conditional claims. Every benefit holds within the scope, scale and conditions measured; negative results delimit the theory. Non-significant results do not establish the absence of an effect.

8. Audit as evidence. The record must link the observed state, applied principle, decision, command, executor and output actually released.

9. Separation of authorities. Technical architecture, normative content, assurance and appeals must not all be controlled by the same entity.

10. Generalisation must be demonstrated. Ethics is the first case; no new domain is established without replication and measured cost-effectiveness.

7. Falsifiable questions

8. Current evidence and claim discipline

9. Proportionate governance

10. Research programme and public trajectory

11. What EthicLaw does not promise

12. The programme’s commitment

Evidence, outcomes and limits

EV-01 — Competence injection

EV-03 — Attribution: comparison still open

NR-1 — Negative findings and withdrawn claims

TN-5 — Known tensions

RF-1 — Conditions that would refute the claims

limiti — Costs and P0.1

Text description and scope of the diagram

Trust
is proven.

Understand its decisions.
Be able to intervene.

Seven modules.
Governance you can auditable.

Do not standardise
values.
Standardise testability
can be verified.

Rigour includes
acknowledging limits.

A thesis to put
to the test.

Manifesto
EthicLaw

The research
programme

The system
map