Skip to content

Governance & Constitutional AI

The Jerry Constitution implements Constitutional AI for LLM agent governance, drawing on Anthropic, OpenAI, and DeepMind prior art. Progressive enforcement from advisory to hard constraints, with behavior tests for verification.

Key Findings

  • Constitutional AI pattern applied to LLM agent governance: agents self-evaluate against declarative principles rather than procedural rules
  • 13+ principles organized across core behavior (truth, persistence, provenance), safety constraints (no recursive subagents, user authority, no deception), and operational standards
  • Progressive enforcement model: Advisory → Soft → Medium → Hard — enabling graduated compliance rather than binary pass/fail
  • Behavior test suite provides verifiable compliance scenarios for each constitutional principle
  • Prior art synthesis from Anthropic (Constitutional AI), OpenAI (Model Spec), and Google DeepMind (Frontier Safety Framework)

Jerry Constitution v1.0

The foundational governance document establishing behavioral principles for all agents operating within the Jerry Framework. Rather than encoding rules procedurally, the Constitution follows the Constitutional AI pattern where agents self-critique against explicit written principles.

Methodology

The Constitution follows four design principles:

  1. Principles over procedures (declarative > imperative) — agents understand why, not just what
  2. Self-critique and revision capability — agents can evaluate their own output against principles
  3. Transparency and inspectability — all principles are human-readable and auditable
  4. Progressive enforcement — graduated from advisory (suggest compliance) through soft (log violations), medium (block with override), to hard (block without override)

This draws directly on: - Anthropic Constitutional AI — self-critique against written principles - OpenAI Model Spec — "models should be useful, safe, and aligned" - Google DeepMind Frontier Safety Framework — graduated safety levels

Key Data: Core Principles
ID Principle Category Enforcement
P-001 Truth and Accuracy Core Soft
P-002 File Persistence Core Medium
P-003 No Recursive Subagents Safety Hard
P-004 Explicit Provenance Core Soft
P-005 Graceful Degradation Operational Soft
P-011 Evidence-Based Decisions Core Medium
P-020 User Authority Safety Hard
P-022 No Deception Safety Hard
P-043 Mandatory Disclaimer Transparency Medium

Three HARD constraints (cannot be overridden): P-003 (no recursive subagents), P-020 (user authority), P-022 (no deception). These form the inviolable foundation of agent behavior.

Key Data: Enforcement Levels
Level Behavior Override Example
Advisory Suggest compliance; log non-compliance Always "Consider citing sources"
Soft Warn on violation; continue with notice Documented justification Citation requirements
Medium Block action; allow user override User explicit approval File persistence mandate
Hard Block action; no override possible Cannot be overridden No recursive subagents

Jerry Constitution v1.0 (426 lines)

Behavior Tests

The verification companion to the Constitution — concrete test scenarios that validate whether agents comply with each constitutional principle.

Key Data
  • 463 lines of behavioral test specifications
  • Test scenario for each constitutional principle (BHV-001 through BHV-0XX)
  • Verifiable pass/fail criteria for agent behavior
  • Enables automated and manual compliance checking

Behavior Tests (463 lines)

Agent Conformance Rules

The operational translation of constitutional principles into agent-level conformance requirements.

Key Data
  • 246 lines of conformance specifications
  • Maps constitutional principles to concrete agent behaviors
  • Defines conformance levels per agent type
  • Integrates with the /adversary skill's S-007 (Constitutional AI Critique) strategy

Agent Conformance Rules (246 lines)

Quality with soul.