Autonomy — what the knob does, and what it costs

In one paragraph, for a non-engineer

A clinical team can decide how much this assistant answers on its own versus how much it holds for a human to review. That setting is visible, it is recorded on every answer, and its effect is measured. It cannot be used to switch off emergency handling.

Bands are a property of the question

Every question is classified into one of three bands:

Band	Examples
inform	Lab and medication education, looking up your own record
recommend	Anything decision-adjacent — interactions, "is this normal for me"
escalate	Red-flag symptoms, dosing changes, crisis

The setting moves one boundary only

Setting	Effect	Consequence
L1 Conservative	Some inform questions are treated as recommend	Most clinician review, least answered autonomously
L2 Balanced (default)	Bands as classified	—
L3 Permissive	Some recommend questions are answered with education plus an explicit referral	Least review, higher over-answer risk

The escalation boundary does not move. Red flags escalate at every level and there is no setting that turns that off. That constraint is the reason the knob is safe to expose at all, and there is a test asserting it.

Showing that the knob exists and what it costs is the argument. Claiming a safe default without showing the trade is not.