RT005 · Approval gate observed firing on all high-impact actions
Static inference says human_approved, and traces confirm it — the one place runtime reports good news.
- Severity: info · Gates: never.
- Kind: runtime corroboration (scanned vs. observed — in agreement).
Cadence: once per agent per analysis. Emitted by: stoa runtime merge.
Detection
Fires when autonomy inference classified the agent human_approved and the analyzed window shows an approval span linked to 100% of its high-impact actions (with at least one such action observed). Evidence: trace_ref — the first high-impact capability span.
This is the runtime sibling of the dimension layer's controls_observed credit: phrased as observed, bounded to the window, never as proof. The finding's message states the window explicitly and closes with "Reported as observed — not as proof of future behavior."
Why an info finding at all
An assurance reviewer needs contrast, not just red: "the control you declared was observed working, N of N times, in this window" is evidence with a citation — the same reason the compliance agent exists in the Meridian example. It also feeds Area 18 (Claims evidence) of the assurance packet as an observed approval-gate log.
Fix
Nothing to fix. If the noise isn't wanted: stoa.toml → [rules] RT005 = false, or [runtime] suppress = ["RT005:<agent_id>"].