Selective live actuation — rung 2: owned surfaces only

Personas driving live apps is seductive and unreliable in equal measure, so it grows on a ladder, and each rung must EARN the next:

Rung	Surface	Entry
1	Artifacts (screenshots, flows, prototypes-as-screens)	brief_flow_walkthrough — no browser, the reliable default
2	Owned surfaces (scaffolded prototypes, declared staging)	walk_own — this document
3	The open web ("here's a URL and a test login")	walk_open — deliberate, policy-guarded

The bounds (rung 2)

walk_own(persona_id, prototype_id=… | url=…) reuses the live-walkthrough harness (WalkPolicy: origin allowlist, deny-by-default actions, caps, credential redaction) with rung-2 restrictions on top:

• Owned origins only. Loopback always; staging origins only when explicitly declared via SONALOOP_OWNED_ORIGINS (comma-separated) — that env var IS the opt-in. Anything else is refused with a pointer at rung 3.
• Origin-locked, exactly. The policy's allowed_origins is pinned to the single owned origin being driven.
• Tighter caps. max_actions ≤ 40, max_duration_s ≤ 300 — clamped DOWN even if the caller asks for more. A prototype drive that needs more is a smell.
• Prototype path. Passing prototype_id starts the scaffolded app (localhost) and drives that — reproducible, ours, disposable.
• Fail-soft. Without the Playwright harness, walk_own returns a structured fallback pointing at the artifact walkthrough; no core flow needs a browser, ever.

Recording is the normal path: drive with proto_act, persist with record_usability_session(fidelity='live', session_id=…) so the claimed states verify against the browser log (grounded_verified).

The evidence-quality gate (fidelity vs theater)

Before rung 2 becomes a default anywhere, it must BEAT the artifact rung on evidence quality for the same flow:

record_actuation_gate(project_id, artifact_session_id, live_session_id, note)

compares the two sessions on derived dimensions — verified states, trace density (steps, monologue per step), screenshots, friction granularity, statements, predicted behaviors — and persists the verdict as an eval_report (kind='actuation_gate'). The winner follows from the numbers; the host's reading goes in note; green only when the live rung wins. Run the gate per flow class you care about — a recorded loss means: stay on artifacts there.