Doctrine

The 4 autonomy levels N0–N3

ACF® classifies agents by autonomy level. The recommended progression is N0 → N1 → N2 → N3. Each upgrade triggers the controls of card ACF-00 (Sovereignty Score).

!Warning

N2 is the default target for the majority of production use cases in 2026-2027. N3 is the exception, not the rule. Skipping N1 to land directly at N2 is acceptable; skipping N2 to reach N3 is not, save for documented edge cases.

N0 — Classical automation

Fixed rules, no learning. Every change goes through a human change request.

Expected controls : Static config diff review · change management · no log requirement beyond classical IT operations.

N1 — Assisted agents

The agent analyses, ranks, and recommends. The final decision remains systematically human.

Expected controls : Signed recommendation log · human accept/reject capture · monthly drift review · DDAO designated but light cadence.

N2 — Governed agents

The agent decides within a strict frame (agentic constitution + locked non-delegable zones + tested kill switch + signed register). This is the default target for production agentic commerce in 2026-2027.

Expected controls : Full ACF-04/06/08/12 stack · quarterly simulation (ACF-15) · 30-day audit (ACF-10) · DDAO with veto authority.

N3 — Supervised autonomy

The agent decides AND learns. Reserved for specific cases. Requires the most mature governance: monthly review, dual DDAO, biannual external audit, embedded model board approval before any model swap.

Expected controls : Everything from N2 + continuous model evaluation pipeline + biannual external audit + dual-control kill switch + canary deployments.

Transition rules

N0 → N1 — card ACF-04 produced, DDAO designated, recommendation register active.
N1 → N2 — ACF-06 tested, ACF-08 Ed25519-signed, ACF-15 simulation passed, ACF-10 30-day audit green.
N2 → N3 — dual DDAO in place, initial external audit passed, continuous model evaluation pipeline operational, board signs off.