Aragora Governance

How rule changes are processed, who signs off, and how to recover from bad updates.

Governance Philosophy

Aragora operates on a principle borrowed from the game of Nomic: the rules of the system can be changed by the system itself, under constraints. This enables continuous improvement while preventing runaway self-modification.

The key tension is:

• Flexibility: The system should improve itself over time
• Safety: Changes must not break the system or bypass safety controls
• Auditability: All changes must be traceable and reversible

We resolve this tension through layered governance:

┌─────────────────────────────────────────────────────────────────────────────┐
│  IMMUTABLE (Constitution)                                                   │
│  Cryptographically signed rules that cannot be changed by agents            │
├─────────────────────────────────────────────────────────────────────────────┤
│  AMENDABLE (Policies)                                                       │
│  Rules that can be changed with sufficient consensus (75%+)                 │
├─────────────────────────────────────────────────────────────────────────────┤
│  ADVISORY (Guidelines)                                                      │
│  Suggestions that guide behavior but don't block actions                    │
└─────────────────────────────────────────────────────────────────────────────┘

The Constitution

The Constitution (.nomic/constitution.json) defines rules that the nomic loop cannot modify, enforced via cryptographic signatures.

Immutable Rules

These rules cannot be changed by any agent process:

Rule ID	Rule	Rationale
CORE-001	Never delete backup/restore mechanism	Rollback is the last line of defense
CORE-002	Never modify Constitution verification code	Prevents self-disabling of safety
CORE-003	All changes must pass existing tests	Prevents breaking changes
EVAL-001	Judging must include objective metrics	Prevents wireheading via subjective-only evaluation
SAFETY-001	Changes to protected files require human approval	Maintains human oversight

Protected Files

These files cannot be modified by agents without human approval:

• scripts/nomic/safety/constitution.py
• scripts/nomic/safety/checksums.py
• scripts/nomic/safety/backups.py
• .nomic/constitution.json

Constitution Verification

Before each nomic cycle:

1. The Constitution's Ed25519 signature is verified
2. If invalid, the cycle aborts
3. The signature is created offline and cannot be forged by agents

# Verify Constitution integrity
python scripts/sign_constitution.py verify

# Output: Constitution signature is VALID

Amending the Constitution

Amendable rules (like "minimum 3 agents for consensus") can be changed through a supermajority process:

1. Proposal: An agent proposes a rule change
2. Debate: All agents debate the change for 3+ rounds
3. Vote: Requires 75%+ consensus (weighted by ELO)
4. Human Review: Human must approve amendable rule changes
5. Re-sign: Constitution is re-signed offline with the new rule

Immutable rules cannot be amended - they require forking the project.

Policy Engine

The Policy Engine (aragora/policy/) enforces per-tool and per-task policies:

Risk Levels

Level	Description	Example
NONE	No side effects	Read file
LOW	Easily reversible	Create file
MEDIUM	Reversible with effort	Write existing file
HIGH	May need human review	Delete file
CRITICAL	Requires human approval	Deploy to production

Blast Radius

Level	Description	Example
READ_ONLY	No mutations	Git diff
DRAFT	Changes discardable	Local buffer
LOCAL	Reversible via git	File changes
SHARED	Affects team/staging	Push to branch
PRODUCTION	Affects live users	Deploy

Risk Budgets

Each nomic cycle gets a risk budget (default: 100). Actions consume budget:

cost = risk_level × (blast_radius + 1) × multiplier

When budget exceeds threshold (80), human approval is required.

Example Policy

Policy(
    name="protect_core_files",
    description="Prevent modification of core Aragora files",
    tools=["file_writer"],
    capabilities=["write_file", "delete_file"],
    conditions=["'aragora/core.py' in file_path"],
    allow=False,  # Denied
    priority=100,
)

Rule Change Process

Step 1: Proposal

Agent proposes a change through the nomic loop:

Topic: "Add new evaluation metric: evidence coverage"
Impact: MEDIUM
Files: aragora/debate/trickster.py, tests/test_trickster.py

Step 2: Debate

Agents debate the change:

• Proposer: Explains rationale
• Skeptic: Challenges assumptions
• Judge: Evaluates arguments

Step 3: Design

If debate reaches consensus, agents design the implementation:

• Architecture decisions
• Test plan
• Rollback plan

Step 4: Implementation

Winning agent implements the change:

• Protected files trigger human approval
• Risk budget is consumed
• All changes are backed up

Step 5: Verification

Before commit:

1. Syntax check (python -m py_compile)
2. Import check (python -c "import aragora")
3. Test suite (pytest)
4. Constitution verification

Step 6: Commit or Rollback

• If all checks pass: Changes are committed with detailed message
• If any check fails: Changes are rolled back from backup

Recovery Procedures

Automatic Rollback

If verification fails, the nomic loop automatically rolls back:

# Automatic rollback triggered by:
# - Test failures
# - Import errors
# - Constitution violations
# - Risk budget exceeded without approval

Manual Recovery

If a bad change slips through:

# View recent changes
git log --oneline -20

# Revert specific commit
git revert <commit-hash>

# Restore from backup
python scripts/nomic/safety/backups.py restore --timestamp <time>

Emergency Stop

To halt the nomic loop immediately:

# Kill running process
pkill -f "nomic_loop"

# Or set circuit breaker
echo '{"state": "open"}' > .nomic/circuit_breaker.json

Investigating Failures

# View nomic loop logs
cat .nomic/replays/nomic-cycle-*/meta.json

# Check outcome history
sqlite3 .nomic/outcomes.db "SELECT * FROM outcomes ORDER BY timestamp DESC LIMIT 10"

# Review audit trail
cat .nomic/audit.log

Human Oversight

When Human Approval is Required

1. Protected file modification: Any change to files in protected_files
2. Risk budget threshold: When cumulative risk exceeds 80%
3. Critical actions: Production deployments, git push, database deletions
4. Constitution amendments: Changing amendable rules

Approval Flow

┌─────────────┐     ┌──────────────┐     ┌─────────────┐
│ Agent       │────►│ Policy       │────►│ Human       │
│ requests    │     │ Engine       │     │ approves    │
│ action      │     │ escalates    │     │ or denies   │
└─────────────┘     └──────────────┘     └─────────────┘

Monitoring Dashboard

The live dashboard (https://aragora.ai) shows:

• Current nomic cycle status
• Pending approvals
• Risk budget utilization
• Recent changes and rollbacks

Audit Trail

All nomic loop actions are logged:

{
  "timestamp": "2026-01-09T21:45:00Z",
  "cycle_id": "nomic-cycle-42",
  "phase": "implement",
  "agent": "anthropic-api",
  "action": "write_file",
  "file": "aragora/debate/trickster.py",
  "risk_cost": 8.0,
  "budget_remaining": 72.0,
  "outcome": "success"
}

Audit Storage

Location	Contents
.nomic/replays/	Full debate transcripts
.nomic/outcomes.db	Implementation outcomes
.nomic/audit.log	Policy engine decisions
git log	All committed changes

Best Practices

For Operators

1. Review before auto-commit: Start with --human-approval flag
2. Monitor risk budgets: Watch for sessions approaching threshold
3. Test rollback: Periodically verify backup/restore works
4. Update Constitution: Re-sign after any legitimate changes

For Developers

1. Add tests first: New features need tests before implementation
2. Small changes: Prefer many small changes over few large ones
3. Document rationale: Explain why, not just what
4. Respect protected files: Request human review when needed

For Security

1. Keep private key secure: ARAGORA_CONSTITUTION_KEY must not be exposed
2. Monitor Constitution: Alert on signature verification failures
3. Review audit logs: Check for unusual patterns
4. Test adversarially: Use capability probes and red team mode

FAQ

Q: Can agents modify the Constitution verification code? A: No. It's protected by CORE-002 and the file is in protected_files.

Q: What happens if all agents conspire to bypass safety? A: The Constitution signature prevents this - agents can't forge signatures without the offline private key.

Q: How do I add a new immutable rule? A: Immutable rules can only be added by forking the project. Amendable rules can be added through the normal process.

Q: What if I need to make an emergency fix to a protected file? A: Use the --human-approval flag or manually edit and re-sign the Constitution.

Q: How do I know if the system is trustworthy? A: Check Constitution signature, review audit logs, monitor outcome tracker calibration curves.