The operating idea
The core principle is simple: The model does not constitute the system. AI may classify intent, reason about a path, and produce a recommendation, but the surrounding control plane owns governance, approval, execution scope, and audit evidence.
Plan, approve, execute, audit
The reference implementation is organised around a Plan, Approve, Execute, Audit flow. The orchestrator structures the request, governance evaluates deterministic rules, approval keeps authority with humans, the gateway enforces scoped execution tokens and allowlists, and audit evidence records what happened.
Implemented today
- Governance evaluator covering GR-001 to GR-011.
- Execution token validation for actor, tenant, action, resource, expiry, approval state, and rollback reference checks.
- Approval workflow scaffolding with approved-token issuance paths.
- Domain-agent framework and reference domain agents.
- OpenAPI contract, HLD, ADRs, threat model, RBAC model, and runbooks.
- CI checks for tests, lint, OpenAPI validation, security scan, UI type-check, and build.
Current evidence
The public repository keeps implementation evidence and maturity boundaries visible through STATUS.md. Current evidence includes verified check-suite expectations across pytest, linting, security scanning, OpenAPI validation, UI build checks, Docker Compose configuration, migration validation, Kubernetes manifest validation, production-runtime validation, and control-plane smoke testing.
Docker Compose smoke evidence covers service health, approval quorum, post-approval token issuance, gateway execution by server-side token ID, required audit events, audit-chain integrity, and fail-closed token issuance before approval quorum.
Boundary
EAAP is a production-oriented reference implementation and personal architecture experiment. It is not presented as a certified standard, a production-ready enterprise platform, or an autonomous infrastructure operator.