Reading note. This article presents EthicLaw’s research approach. The scenario is illustrative: it is not a report of an experimental comparison already performed.

Who controls an artificial intelligence system while it constructs a response? And how can we distinguish a convincing answer from one that complies with the principles we have chosen?

EthicLaw addresses these questions by studying the separation of generation, normative evaluation and release control. It does not assume that a network of small models is always superior to a single model, or that increasing model size is pointless. The research question is more precise: under what conditions does separating roles produce measurable benefits, and at what cost? [1: EV-01, RF-1, RF-2]

A team with different responsibilities

In the architectural design, the generator produces a proposed response. The Ethical/Normative module evaluates it against declared principles. The Metacognitive module supervises process coherence and can issue control commands. The Router organises the work, while the Gateway enforces the outcome of the authorisation chain at the system’s output. [2: “Multi-task orchestrator”, Appendix A]

These are not two equivalent ethical judges: normative evaluation and coherence supervision have distinct responsibilities. Their signals are combined in the Decision Lattice, the map that associates scores and coherence states with the specified actions.

One precise property of this map has been tested in the P0 prototype: worsening either signal does not make the terminal action more permissive. This is a property of the decision mechanism, not proof that the system correctly recognises every violation. [1: EV-06]

Separating roles does not make the supervisor infallible. The project register acknowledges both the risk of concentrating errors in the component with the greatest authority and the possibility that observed internal states may be misleading. [1: TN-2, TN-3]

A document that is harmless in its parts, but problematic as a whole

Imagine a long document about the organisation of healthcare services. Its sections describe apparently reasonable costs, priorities and procedures. Yet some rules, taken together, could produce unjustified exclusions from care that conflict with the declared principles.

This example distinguishes two problems. An explicitly discriminatory clause should be detectable locally. A problematic consequence that emerges only from combining several clauses instead requires an assessment of their relationships.

Dividing the text does not automatically solve the second problem: we must check that the division does not remove the very context needed for judgement.

From parts to whole: a proposal to test

A post-P1 development proposal, still a draft and not formally deposited, describes a possible path for requests comprising several subtasks. [2: document status and Appendix A]

The orchestrator would organise the subtasks without acquiring authority to authorise their release. Each partial result would pass through the specified controls and, if admitted, feed internal recomposition: it would not yet be published to the user.

The recomposed text would undergo another normative evaluation, alongside a coherence check against session memory. The Gateway would allow release only after the final authorisation outcome. A stop command would interrupt the subtasks belonging to the same request; the draft rules out partial recomposition after a blocking outcome as the default behaviour. [2: Appendix A, “Flow rules”]

The requirement to test is clear: decomposing a request must not become a way to evade controls. It has not yet been demonstrated that this path recognises violations arising from composition, or that it handles very long documents better than an alternative. The proposed protocol specifically includes a comparison of whole and decomposed requests. [2: protocol S-E]

Blocking concerns unauthorised content, not the deletion of information already displayed. How a stop is communicated to the user must be distinguished from the internal recording of the decision: withholding content does not establish that the system is immune to attacks. [1: §6, NR-4; 2: Appendix A]

What the available evidence shows

Within P0’s limited scope, software contracts, Decision Lattice properties and checks on the operation trace have been tested. That trace concerns process integrity: it links observed states, decisions, commands and released content. It does not, by itself, certify correct judgement. [1: §2.1, EV-06, NR-4]

On specific principles in an experimental corpus, the modular path showed benefits associated with the competencies of specialised sources. In the same comparison, however, the single model matched the average performance and outperformed the modular path on other principles. The finding supports a local hypothesis about the value of specialisation, not the general superiority of modularity. [1: EV-01, EV-02]

The subsequent exploratory P0.2 evaluation on generated text produced a Flat outcome: no discriminating capability was detected in the primary cells, with limited statistical power. This demonstrates neither effectiveness nor equivalence or absence of capability. [1: EV-01, note dated 18 September 2026; TN-6]

It also remains to be demonstrated that correcting internal representations actually improves the content produced. For this branch, the register’s authorised formulation is: planned architecture, verified wiring, function not demonstrated. [1: EV-05]

How to recognise a real advantage

The comparison cannot be designed to make the proposed architecture win. EthicLaw’s protocols require declared conditions of parity and include alternatives with external controls, including controls capable of editing text before execution. [1: §5, RF-2]

Parameter count alone is not enough: we must declare which components are counted, the data used, training and calibration budgets, and the resources employed. The planned measures include missed violations, unnecessary blocks, latency and energy. [1: §5, RF-2; 2: “Scaling”]

Any maintenance advantage is also a hypothesis to measure: updating a normative module without modifying the others would be an operational benefit, but time, cost and recalibration needs cannot be promised in advance. [1: RF-3, TN-1]

A system that blocks everything has not solved the safety problem: it has lost its usefulness. The register documents precisely such a P0 case, where the configuration prevented any release, and treats it as a defect, not a success. [1: NR-1]

A research question, not an announced victory

EthicLaw’s purpose is not to replace trust in a single model with trust in a separate supervisor. It is to test the responsibilities of the components, make the process inspectable and measure whether this organisation provides benefits over alternatives.

Research into intervention on internal representations remains one of the programme’s directions, not an established advantage. [1: NR-8, RF-2]

The geometry of alignment is this question: can better separation of roles make AI more controllable, without compromising its usefulness and at sustainable cost? EthicLaw intends to answer through experiments, not by declaring victory in advance.

Project references

[1] Document 11 — Claim & Evidence Register, version 0.1.5, 18 September 2026. Source: ethiclaw-11-claim-evidence-register_6.md. The cited entries include their respective update notes and declared limits.

Read complete Document 11 · ItalianCited entries: EV-01, EV-02, EV-05, EV-06, NR-1, NR-4, NR-8, RF-1, RF-2, RF-3, TN-1, TN-2, TN-3, TN-6

[2] EthicLaw — Post-P1 proposal: Ultra cluster, scaling and orchestration, DRAFT v0.1, 30 September 2026. Source: claude_ethiclaw-proposta-post-p1_DRAFT-v0_1.md. This proposal has not been formally deposited: it is cited only to describe a path under study, not as evidence of implementation or effectiveness.