Governance
Constitution v1.5.1 · 16 Laws · 14 Gates · 9 enforced at runtime
Runtime Violations (24h)
The 16 Constitutional Laws
Immutable governance principles that bind all agents. No agent may override these without explicit human approval.
Every agent shift produces observable evidence -- logs, cost entries, audit events. Silence is a failure mode. No run completes without a trace.
All code merges, deploys, and production-facing changes require review by a different agent or human. No self-service to prod.
Code writers (Claude), code reviewers (GPT-4o), and verifiers (Groq/Llama) use different LLM providers. Prevents monoculture failures.
CFO runs first, every cycle. No agent can spend beyond its $2/day cap. Daily fleet cap is $20. CFO suspension is binding.
When confidence is low or the action is irreversible, agents must escalate to human review. Autonomy has guardrails.
All agent shifts are logged. All costs are tracked. All merges and deployments produce observable events. Audit events cannot be deleted by agents.
Any action an agent can take autonomously can be overridden by a human at any time. SCHEDULER_DISABLED=true kills everything.
API keys, tokens, and credentials are never written to files, logs, or messages. Railway is the only secrets store. scan_secrets.sh gate at PR time.
PR review SLA: 24h maximum. Escalations must be acknowledged within 6h. Stalled agents (>24h) are flagged by Ops Manager.
Agent definitions, governance rules, and authoritative decisions live in the GitHub repo. Agents do not modify their own SKILL.md files.
The CFO agent has binding authority over budget enforcement. No agent may override a CFO suspension. Budget is law.
No agent can modify .governance/, SOP files, or the Constitution without explicit human approval. Protected paths enforced.
All agent reports must conform to the report schema. Invalid reports are rejected and the agent is flagged for remediation.
Agent-to-agent handoffs require explicit receipt confirmation. Unreceipted handoffs trigger escalation after timeout.
Protected paths enforced in github-tools.js. scan_secrets.sh gate at PR time. No exceptions. Violations block merge.
No two instances of the same agent may run simultaneously. The runningAgents map prevents double-dispatch at the scheduler level.
14-Gate Enforcement Pipeline9 enforced1 PR-time4 honor-based
Every agent action passes through this pipeline. Gates are ordered by enforcement phase: code, CI, runtime, policy.
Where Things Live
Authoritative file locations for governance artifacts. The GitHub repo is the single source of truth.
gov/sops/SOP-001.mdThe 16 laws that govern all agent behavior
gov/.governance/GATES/14-gate enforcement pipeline definitions
gov/.governance/CONTRACTS/Report schemas, handoff schemas, cost schemas
gov/MASTER_PLAN.mdStrategic roadmap and milestone definitions
data/agents/*/SKILL.mdPer-agent capability definitions and constraints
PostgreSQL audit_events tableEvery control plane action is logged here
How to Update the Constitution
The amendment process ensures governance changes are deliberate, reviewed, and traceable.
Propose
Draft a change to the Constitution, SOP, or gate definition. Document the rationale and expected impact.
Review
Open a PR against the governance repo. Requires review from at least one human and one cross-provider AI reviewer.
Approve
Human (Ethan) must explicitly approve. No agent may approve governance changes. Protected path enforcement blocks unauthorized merges.
Merge & Deploy
Merge to main. Governance changes propagate to all agents on next cycle. Audit event is recorded. Version is bumped.
Honor-Based Controls -- Risk Inventory
These controls exist in SKILL.md instructions but have no hard runtime enforcement.
The following controls are agent-policed. Agents are expected to comply, but there is no gate that prevents violation:
- G05: Agent 06 reviewing its own code (provider separation enforces indirectly)
- G10: Agent 10 deploying without Agent 07 PASS gate
- G11: Agent 01 sending email without Ethan approval
- G12: Agent 17 suspending Agents 01, 07, or 10
- Agent 16 acting without a clear fix (self-regulated threshold)
Phase 3 roadmap: Add audit log triggers and notification gates for highest-risk honor-based actions. Agent 06 merges and Agent 10 deploys will get hard enforcement via webhook-based approval flows.
Hardening Checklist4 / 8 complete
audit_events table -- incidents, suspensions, dispatches, escalation resolves
All POST routes validate required fields; SQL parameterized queries throughout
CFO Agent 17 + /api/budget/caps -- $2/agent/day, $20/fleet/day
Railway env vars only -- no secrets in code, scan_secrets.sh CI gate
No middleware rate limiter -- Vercel edge limits apply but no app-level throttle
All actions attributed to "ethan" hardcoded -- NextAuth is in place but session identity not wired to audit events yet
No explicit CORS headers -- relies on Next.js same-origin defaults
No Content-Security-Policy, X-Frame-Options, or Strict-Transport-Security headers configured