The rubric suggests High for the systems attached and the autonomy declared. The stored tier is Medium.

Why this tier

Medium in its first form: every change the agent proposes is reviewed by a named editor before it reaches a live page, so the blast radius of a bad inference is a rejected suggestion. The tier rises as soon as promotion becomes automatic.

The ladder

Rung 1 — detect and report only. Rung 2 — propose a diff against a page, human promotes. Rung 3 — auto-promote a narrow, enumerated class of fixes (broken links, alt text, stale hours) with post-hoc review. Each rung is a separate approval, not a gradual drift.

Never

No clinical content, no provider credentials, no pricing or coverage language, and no page that collects patient information. Those pages are read-only to the agent at every rung.

Audit

Every proposal, promotion, and rejection is logged with the agent's evidence and the reviewer's name. The log is the artifact a regulator or an internal audit reads, so it is written before the fix ships, not after.