MEUUS HUMAN APPROVAL LADDER

L0 — OBSERVE / READ
AI may read only information explicitly available to the agent.
No action.

L1 — RECOMMEND
AI may suggest.
Example: recommend a learning resource.
Human remains decision-maker.

L2 — PREPARE
AI may create a draft/proposed action.
Examples:
- plan
- email draft
- application draft
- document draft
No consequential external execution merely because preparation exists.

L3 — LOW-RISK REVERSIBLE EXECUTION
Only when a specific workflow policy allows it.
Must have:
- bounded scope;
- reversible action;
- evidence/log;
- clear rollback;
- no hidden consequential effect.

L4 — CONSEQUENTIAL
Human confirmation required.
Examples may include:
- external communication;
- publication;
- submission;
- financial commitment;
- account-level change.

L5 — RESTRICTED
Human/admin/qualified-professional authority only.
Examples:
- professional legal judgment;
- clinical decision;
- high-risk financial authority;
- irreversible destructive action;
- sensitive administrative control.

The classification sets a ceiling.
It does not grant permission.
