Earned Autonomy Protocol

An agentic OS that only installs OPA/Cedar/Keycloak still needs a doctrine for when execution scope expands. EAP is the IAM-cell research pointer for earned trust, perception vs action loops, and constitutional invariants (money, filings, unprompted client contact) that must not move on a “good streak.” Prefer OPA/Cedar to evaluate permissions; prefer Keycloak for identity; prefer Headroom (Guardrails cell) for measuring HITL escalation fatigue on coding actions.

Operators run agents in observation-first mode, evidence-lock decision paths (Trust Composite Scores), automate perception loops that only produce checkable claims, and keep action loops human-gated until genesis/ceremony expands scope. Constitutional boundaries change only by explicit logged human ceremony.

1. Default: observation / no execution autonomy.
2. Log and verify bounded decisions; run perception automation.
3. Gate world-binding actions until evidence justifies a grant.
4. Expand or revoke scope per domain; never self-grant.

Related: cedar · keycloak · open-policy-agent · topics/13-governance-policy

• Use when: designing autonomy governance (scope, genesis, constitutional invariants) for high-stakes installs.
• Skip when: you need an implementable IdP (keycloak) or policy engine (opa, cedar) alone without methodology.
• Prefer instead: OPA/Cedar for runtime allow/deny; EAP for the rules of when those allows may widen.

• Spec, not a binary: you still implement gates in OPA/Cedar/product code.
• Vendor narrative risk: chronology and production claims are author-asserted — verify against your corpus before treating as measured fact.
• Tiny star / consultancy origin: treat as doctrine input, not CNCF infrastructure.
• Not the identity store or policy DSL.