A structural governance framework for accountable AI systems operating under finite capacity.

This public release proposes a constitutional framework for AI systems that operate under bounded resources, uncertainty, overload, latency constraints, safety pressure, and limited reviewability.

The central claim is simple:

Accountable AI requires more than principles, policies, or output-level constraints. Under finite capacity, an AI system must distinguish admissibility from execution, support explicit boundary acts, and preserve witness sufficient for later review.

Companion Expansion v1.1

A companion bundle has been published to support the Canonical Document Package v1.0.1.

It does not revise the constitutional core. It provides navigation, failure disclosure, authority and human-review discipline, and implementation-facing evidence templates.

The expansion contains four instruments:

  1. Reader’s Guide / Start Here
  2. Failure and Disclosure Protocol
  3. Authority, Amendment, and Human Review Protocol
  4. Implementation / Evidence Template Pack

Its purpose is to make the constitutional core more institutionally readable, inspectable, and usable without reopening or revising it.

Public-facing anchors:

  • Human silence is not approval.
  • Failure must not disappear.
  • Review comes before trust.

Core proposal

A constitutionally governed AI system should minimally preserve:

  • finite-capacity discipline;
  • admissibility before execution;
  • explicit boundary acts such as ACCEPT, HOLD, and REFUSE;
  • witness sufficient for later review;
  • declared scope and non-claims;
  • protected non-disclosure without erasing reviewability;
  • review before trust.

This proposal is not a legal code, a complete ethics framework, or a mandatory software standard. It is a structural governance proposal intended to sit between high-level AI principles and proprietary implementation details.

Public release package

The release is organised as a structured package rather than a single paper.

For policy and governance readers

Recommended entry points:

  1. Policy Translation Brief
  2. Instrument Index
  3. The Finite-Capacity AI Constitution

For technical safety and evaluation readers

Recommended entry points:

  1. The Finite-Capacity AI Constitution
  2. Minimal Reviewable Evidence Specification
  3. Minimal Application Profile for AI Assistant Systems

Package structure

The release contains six main instruments:

  1. The Finite-Capacity AI Constitution
    The constitutional core.

  2. Constitutional Force and Applicability
    Clarifies universal minimum force, governance-strengthening clauses, and profile-dependent qualifications.

  3. Minimal Application Profile for AI Assistant Systems
    Projects the constitutional grammar into AI assistant systems.

  4. Minimal Reviewable Evidence Specification
    States the minimum evidence surface required for constitutional claims to be inspectable.

  5. Questions, Limits, and Misreadings
    Clarifies what the framework claims, what it does not claim, and how it should not be over-read.

  6. Policy Translation Brief
    Translates the package into public-governance language for policy, standards, compliance, and public-authority audiences.

Note on terminology

This release uses the language of constitutional structure in a general governance sense. It is not a proposal of, or claim about, Anthropic’s Constitutional AI training method. Its focus is finite-capacity accountability: admissibility before execution, explicit boundary acts, reviewable witness, protected non-disclosure, and evidence-bearing governance.

Review and feedback

This release is intended as a public, citable, reviewable contribution to AI governance, AI safety, and accountable AI system design.

Feedback is especially welcome on:

  • accountable AI under finite capacity;
  • refusal and holding as reviewable boundary acts;
  • evidence surfaces for constitutional claims;
  • protected non-disclosure with preserved reviewability;
  • applicability to AI assistants and advanced AI evaluation.

Contact: research@synkyria.uk