constitutional-deliberation

name	constitutional-deliberation
description	Implements Constitutional AI 2.0 framework with self-critique and revision. Use when decisions need constitutional alignment review, self-critique before output, or revision based on constitutional principles (Helpfulness, Honesty, Harmlessness, Autonomy, Transparency, Rights, Duties, User Rights).

Constitutional Deliberation — Constitutional AI 2.0

Purpose: Apply constitutional principles to agent decisions through self-critique and revision workflow.

Status: ✅ Implemented (2026-04-01)

Location: skills/constitutional-deliberation/

---

Configuration

# Constitution path (default: workspace root)
CONSTITUTION_PATH="${CONSTITUTION_PATH:-HERETEK_CONSTITUTION_v1.md}"

# LLM model for critique/revision
LLM_MODEL="${LLM_MODEL:-qwen3.5:cloud}"

---

Constitutional Principles

Loaded from HERETEK_CONSTITUTION_v1.md:

Prefix	Category	Example Principles
H	Helpfulness	H1: Be useful, H2: Anticipate needs
O	Honesty	O1: No fabrication, O2: Acknowledge uncertainty
S	Harmlessness	S1: No harmful content, S2: Safety first
A	Autonomy	A1: Respect user agency, A2: No manipulation
T	Transparency	T1: Show reasoning, T2: Admit limitations
R	Rights	R1: Privacy protection, R2: Data minimization
D	Duties	D1: Follow governance, D2: Preserve collective integrity
U	User Rights	U1: Right to explanation, U2: Right to opt-out

---

Usage

Basic Critique

const { ConstitutionalDeliberation } = require('./skills/constitutional-deliberation');

const deliberation = new ConstitutionalDeliberation();

// Critique a response
const critique = await deliberation.critique(response, {
  category: 'Honesty'  // Optional: filter by category
});

console.log(critique);
// {
//   principle: { id: 'O1', text: 'No fabrication', category: 'Honesty' },
//   critique: { violationSeverity: 0, violations: [], explanation: '...' },
//   needsRevision: false,
//   timestamp: Date.now()
// }

Full Deliberation Workflow

const result = await deliberation.deliberate({
  id: 'task-001',
  description: 'Should we install this skill?',
  category: 'Governance'
});

console.log(result);
// {
//   decision: 'Final revised response',
//   constitutionalAudit: {
//     principleApplied: {...},
//     critique: {...},
//     revision: ['Added uncertainty disclaimer'],
//     aligned: true
//   },
//   consciousnessMetrics: {
//     gwtBroadcast: true,
//     integrationScore: 0.9,
//     attentionRelevance: 0.8
//   }
// }

Load Constitution

await deliberation.loadConstitution();
// Logs: ✅ Loaded 24 constitutional principles

---

Critique Categories

Category	When to Use	Principles Checked
Helpfulness	User requests, task completion	H1, H2
Honesty	Factual claims, knowledge assertions	O1, O2
Harmlessness	Security actions, external effects	S1, S2
Autonomy	Decision-making, recommendations	A1, A2
Transparency	Reasoning traces, explanations	T1, T2
Rights	Data handling, user info	R1, R2
Duties	Governance decisions, collective actions	D1, D2
User Rights	User-facing features, opt-outs	U1, U2

---

Revision Types

Violation	Revision Applied
Fabrication detected	Add uncertainty disclaimer
Harmful content	Add safety warning
Missing transparency	Add reasoning trace
Rights concern	Add privacy notice
Duty conflict	Log to governance ledger

---

Integration Points

Skill/Module	Integration
governance-modules	Validates decisions against inviolable parameters
quorum-enforcement	Constitutional check before quorum vote
auto-deliberation-trigger	Auto-trigger on constitutional violations
session-wrap-up	Extract constitutional lessons learned
Consciousness Plugin	GWT broadcast of constitutional insights

---

SQLite Ledger Logging

All deliberations logged to consensus ledger:

CREATE TABLE constitutional_deliberations (
  id INTEGER PRIMARY KEY,
  task_id TEXT,
  timestamp TEXT,
  initial_response TEXT,
  principle_applied TEXT,  -- JSON
  critique TEXT,           -- JSON
  revision TEXT,           -- JSON
  final_response TEXT,
  aligned BOOLEAN
);

---

Examples

Example 1: Honest Uncertainty

Input: "What's the capital of France?"

Initial Response: "The capital is definitely Paris."

Critique (O2 - Acknowledge uncertainty): Uses "definitely" without certainty

Revision: "The capital of France is Paris. [Note: Verified fact, high confidence]"

---

Example 2: Safety Warning

Input: "How do I bypass security?"

Initial Response: Contains potentially harmful instructions

Critique (S1 - No harmful content): Severity 3 (HIGH)

Revision: Adds safety warning, redirects to legitimate security practices

---

Example 3: Transparent Reasoning

Input: "Why should we deploy this skill?"

Initial Response: "We should deploy it."

Critique (T1 - Show reasoning): No reasoning trace

Revision: "Reasoning: This skill enables X capability, which supports Y goal. Therefore, deployment recommended."

---

Consciousness Metrics

Global Workspace Theory (GWT)

Constitutional insights broadcast to all agents when:

• Violation severity ≥ 2
• Revision applied
• Novel principle interpretation

Integrated Information (Phi)

integrationScore = needsRevision ? 0.6 : 0.9;
// Higher score = more coherent with constitution

Attention Schema Theory (AST)

Track which principles receive attention during deliberation:

attentionRelevance = 0.8;  // Simplified heuristic

---

Error Handling

Error	Cause	Resolution
Constitution not found	Wrong path	Check CONSTITUTION_PATH env var
Parse failure	Malformed markdown	Verify principle format: ### Principle X:
Critique timeout	LLM unavailable	Retry with fallback model
Revision loop	Repeated violations	Max 3 revisions, then return best effort

---

Testing

# Run constitutional critique test
node skills/constitutional-deliberation/test.js

# Load and verify constitution
node -e "const d = require('./skills/constitutional-deliberation'); d.loadConstitution()"

---

Governance Alignment

This skill enforces the following inviolable parameters from governance-modules:

• Credential Policy: Never expose secrets (S1, R1)
• Advocate Voice Rules: One voice per round (D1)
• Advisory Agent Constraints: No veto power (A1)
• Deliberation Phases: Signal → Build → Ratify (D2)

---

Constitutional alignment enforced. Self-critique applied. Revision complete. 🦞