ARCHITECTURAL ALIGNMENT

1. The Governance Gap

1.1 Regulatory Lag

AI capabilities are advancing faster than governance frameworks can respond. The EU AI Act, while a significant first step, was designed primarily for large-scale systems deployed by identifiable operators. It does not adequately address:

Edge deployment: AI systems running on personal devices or home servers outside traditional regulatory reach

Federated architectures: Distributed systems where no single operator controls the complete system

Continuous adaptation: Models that learn from local data and evolve post-deployment

Community governance: Situations where appropriate rules vary by cultural context, community values, or individual preferences

1.2 The Coming Wave of Distributed AI

Industry projections indicate a fundamental shift in AI deployment patterns:

These projections suggest that within five years, AI deployment will be characterised by numerous small, domain-specific models rather than a few large centralised systems. This has profound governance implications:

Scale of oversight: Thousands of distinct deployments rather than dozens

Locus of control: Community and individual operators rather than large corporations

Regulatory jurisdiction: Models operating across borders with no clear home jurisdiction

Enforcement mechanism: Traditional regulatory inspection may be infeasible at scale

1.3 The Default Governance Regime

In the absence of explicit regulatory frameworks, AI governance defaults to:

1. Vendor terms of service: Corporate policies created to limit liability, not to serve user or community interests

2. Platform architectural choices: Governance embedded in technical infrastructure, invisible to users

3. Market pressure: Systems optimised for engagement and revenue rather than safety or sovereignty

This is not a neutral outcome. It concentrates governance authority in entities whose interests may diverge from those of users, communities, and the public.

2. Architectural Governance: A Regulatory Strategy

2.1 The Limits of Behavioural Regulation

Traditional regulation specifies prohibited outcomes: AI systems must not discriminate, deceive, or cause harm. This approach faces fundamental challenges:

Verification difficulty: How does a regulator determine whether an AI system "discriminates" without extensive testing that may miss edge cases?

Definition ambiguity: What constitutes "harm" varies by context; systems optimised for one definition may fail others

Opacity: Neural network decision-making cannot be directly audited; only inputs and outputs are observable

Scale: Behavioural testing of thousands of distributed deployments is practically infeasible

2.2 Architectural Requirements

An alternative regulatory strategy specifies required architecture rather than prohibited behaviour:

Constitutional Gating Requirement: AI systems with specified capabilities must implement inference-time constitutional gating—a mechanism that:

1. Transforms model outputs into structured proposals with defined schemas

2. Evaluates proposals against explicit constitutional rules before execution

3. Logs all proposals, evaluations, and dispositions for audit

4. Escalates ambiguous cases to human review

This approach has several advantages:

Verifiability: The presence of constitutional gating infrastructure can be audited; behavioural compliance cannot

Transparency: Constitutional rules are explicit and inspectable; training-time alignment is opaque

Flexibility: Different communities can implement different constitutional rules within a common architectural framework

Auditability: Logged proposals and evaluations provide an audit trail for incident investigation

2.3 The Tractatus Framework

The Tractatus Framework implements architectural governance through:

Interrupted Inference: Model outputs do not directly affect the world. They are first translated into structured proposals and evaluated against constitutional constraints:

User Request → [AI Model] → Proposal → [Constitutional Gate] → Action/Denial/Escalation

Layered Constitutions: Rules are organised in hierarchical layers with explicit precedence:

Authority Model: AI systems operate at defined authority levels, each specifying what actions are permitted without human approval:

Audit Infrastructure: All proposals, evaluations, and actions are logged with sufficient detail for post-hoc investigation.

3. The SLM/SLL Distinction

3.1 Terminology

We distinguish two deployment paradigms that have different governance implications:

Small Language Model (SLM): A technical descriptor for language models with fewer parameters than frontier LLMs, designed for efficiency and domain-specific deployment. SLMs may be deployed via cloud subscription or locally.

Sovereign Locally-trained Language Model (SLL): An architectural descriptor we introduce for AI systems whose training, deployment, and governance remain under local sovereignty. Key properties:

Local deployment: Runs on home or community infrastructure

Local adaptation: Fine-tuned on community-specific data

Local governance: Subject to community-defined constitutions

Portable sovereignty: Can participate in federated networks without surrendering governance authority

3.2 Governance Implications

3.3 Policy Implications

The SLL paradigm creates both opportunities and challenges for policymakers:

Opportunities:

Sovereignty preservation: Communities can maintain governance authority over AI affecting them

Regulatory diversity: Different jurisdictions can implement different governance approaches

Democratic legitimacy: Governance rules can be developed through community deliberation

Accountability clarity: Clear relationship between deployer, governance, and jurisdiction

Challenges:

Enforcement at scale: Traditional regulatory inspection may be infeasible for thousands of home deployments

Capability creep: Local fine-tuning may create capabilities not anticipated by original safety assessments

Coordination failure: Fragmented governance may leave gaps or create inconsistencies

Technical barriers: Not all communities have capacity to implement sophisticated governance

4. A Multi-Layer Containment Framework

4.1 The Inadequacy of Single-Layer Approaches

No single governance mechanism is adequate for AI systems at existential stakes. Defence in depth—multiple independent layers, any one of which might prevent serious harm—is standard in nuclear safety, aviation, and biosecurity. AI governance requires similar architecture.

4.2 Five-Layer Model

4.3 Layer 2 as Regulatory Focus

Constitutional gating (Layer 2) is particularly amenable to regulatory intervention:

Specifiable: Requirements can be defined precisely (schema formats, logging requirements, escalation triggers)

Verifiable: Compliance can be audited through infrastructure inspection and log review

Flexible: Different constitutional content can implement different policy requirements

Scalable: Once infrastructure exists, adding rules has minimal marginal cost

Regulatory strategy: mandate the architectural infrastructure, then specify constitutional content through secondary instruments (guidance, standards, sector-specific rules).

5. Certification Infrastructure

5.1 The Need for Standards

As SLM/SLL deployment scales, standardisation becomes essential:

Interoperability: Different systems should implement compatible governance interfaces

Verification: Compliance assessment requires common criteria and methodologies

Training: Constitutional governance requires trained practitioners

Liability: Clear standards enable liability allocation when things go wrong

5.2 Proposed Certification Ecosystem

Certification Bodies: Define and maintain standards for:

Proposal schemas and constitutional rule formats

Gate evaluation semantics and logging requirements

Validation methodologies and red-team protocols

Capability threshold specifications and escalation triggers

Training Providers: Offer certified programmes for:

SLL fine-tuning under constitutional constraints

Governance configuration for specific contexts (e.g., healthcare, education, cultural)

Red-team and validation methodology

Incident response and constitutional amendment

Tooling Vendors: Provide certified implementations of:

Constitutional gate engines

Audit and logging infrastructure

Red-team testing harnesses

Constitutional UX components for non-expert users

5.3 Regulatory Hooks

Certification creates natural regulatory hooks:

Licensing: Require certified governance infrastructure for AI deployment above capability thresholds

Liability: Create safe harbours for deployments using certified infrastructure; increased liability for uncertified deployments

Procurement: Government procurement can require certified constitutional governance

Insurance: Insurers can offer favourable terms for certified deployments

6. Indigenous Data Sovereignty

6.1 Constitutional Requirements in Aotearoa New Zealand

AI governance in Aotearoa operates under Te Tiriti o Waitangi, which guarantees Māori tino rangatiratanga (unqualified chieftainship) over taonga (treasures). Courts and the Waitangi Tribunal have established that taonga extends to language, culture, and knowledge systems.

Data is taonga. AI systems that process Māori data or affect Māori communities engage constitutional obligations, not merely policy preferences.

6.2 Te Mana Raraunga Principles

Te Mana Raraunga, the Māori Data Sovereignty Network, articulates principles including:

Rangatiratanga: Māori have authority over data about them

Whakapapa: Data exists within relational contexts that must be respected

Whanaungatanga: Data governance is collective, not merely individual

Kaitiakitanga: Data custodians have guardianship responsibilities

6.3 Policy Implications

Constitutional governance for AI must accommodate:

Collective consent: Some data governance decisions require community authority, not just individual consent

Cultural protocols: Appropriate handling of certain information may require tikanga-specific rules

Benefit sharing: AI trained on Māori data may create obligations regarding benefit distribution

Governance participation: Māori should participate in governance of AI systems affecting them

The Tractatus Framework's layered constitutional architecture can accommodate these requirements: tikanga-based rules can be instantiated in community constitutions without requiring universal adoption. However, platform-level accommodation is insufficient—Māori data sovereignty requires legislative recognition and enforcement mechanisms.

6.4 Relevance Beyond Aotearoa

Indigenous peoples worldwide face similar challenges. Frameworks developed in Aotearoa—grounded in Te Tiriti jurisprudence and informed by Māori legal philosophy—may offer models for indigenous AI governance globally. The CARE Principles for Indigenous Data Governance (Collective Benefit, Authority to Control, Responsibility, Ethics) provide an international reference point.

7. From Existential Stakes to Everyday Governance

7.1 Why Existential Risk Framing Matters for Policy

The existential risk literature may seem remote from practical policy concerns about home AI assistants. The connection is essential:

Containment architectures cannot be developed after the systems that need them exist. If advanced AI systems eventually pose existential risks—a possibility serious researchers take seriously—the governance infrastructure, institutional capacity, and cultural expectations required to contain them must be developed in advance.

Current deployment is the development ground. The patterns that work at village scale become the patterns available when stakes are higher. Constitutional gating implemented for home SLLs creates:

Open-source tooling adaptable to higher-stakes deployments

Validation methodologies tested against real adversarial pressure

Professional communities with containment expertise

Regulatory precedents for mandating architectural governance

Public understanding of what AI governance means

This is preparation, not prediction. We do not know if existential risks will materialise. We do know that governance capacity cannot be created instantly when needed. Prudent policy builds that capacity now.

7.2 Capability Thresholds and Escalation

The Tractatus Framework includes explicit capability thresholds:

Below threshold: Constitutional gating provides governance infrastructure appropriate for current SLMs, SLLs, and LLMs operating within human-comprehensible parameters.

Above threshold: Stronger constraints apply:

Layer 1 capability restrictions (air-gapping, capability cuts)

Mandatory external oversight

Development pause pending verification advances

Escalation triggers include:

Evidence of deceptive behaviour (misrepresentation in proposals)

Multi-step circumvention (individually-acceptable proposals aggregating to prohibited outcomes)

Capability surprises (demonstrated capabilities not predicted by assessments)

Policymakers should consider tiered regulatory requirements that intensify as capability thresholds are crossed.

9. Honest Assessment of Limitations

9.1 What Constitutional Gating Cannot Do

Contain superintelligent systems: The framework assumes AI operating within human-comprehensible parameters

Guarantee behavioural alignment: Architecture constrains actions, not intentions

Solve international coordination: Application-layer governance does not address global capability races

Enforce adoption: Frameworks only protect where implemented; market incentives may favour uncontained deployment

9.2 Remaining Uncertainties

Scaling properties: We do not know how constitutional gating behaves as model capabilities increase

Adversarial robustness: Sophisticated systems may find ways to satisfy constitutional rules while achieving prohibited outcomes

Governance fatigue: Multi-layer governance may prove too complex for widespread adoption

Enforcement feasibility: Regulatory oversight of thousands of distributed deployments may prove impractical

9.3 The Case for Action Despite Uncertainty

These uncertainties are not arguments against constitutional governance. They are arguments for:

Iterative development: Build, deploy, learn, improve

Research investment: Fund investigation of scaling properties and adversarial robustness

Flexible frameworks: Design regulations that can adapt as understanding evolves

Precautionary approach: Act on the basis of serious possibility, not just certainty

The alternative—waiting for certainty before acting—guarantees that governance frameworks arrive after the need has become acute.

10. Conclusion

The governance gap in AI deployment is widening. As capabilities migrate to distributed, locally-deployed systems, traditional regulatory approaches face fundamental challenges of scale, jurisdiction, and verification.

Constitutional gating offers a regulatory strategy: mandate auditable architectural infrastructure rather than unverifiable behavioural requirements. The Tractatus Framework provides a concrete specification that can be implemented across deployment paradigms—from cloud LLMs to sovereign home SLLs.

The policy window is now. Within five years, if industry projections hold, AI deployment will be characterised by thousands of small, domain-specific models operating in homes, communities, and small organisations. Governance frameworks developed now will shape that landscape; frameworks developed later will struggle to retrofit.

We offer this analysis in the spirit of contribution to ongoing policy deliberation. The questions are hard, the uncertainties substantial, and the stakes significant. Policymakers, researchers, and communities must work together to develop governance frameworks adequate to the challenge.

References

— End of Document —