Sandbox policy/risk services: an observation, not a proof. observation.json records the /fabric/try scenarios run two ways on 2026-09-21: (a) an isolated gateway instance pointed at a real OPA server and the real adversarial-fuzzer, (b) the public production endpoint at the same moment. It shows why the public "prompt-injection" sandbox scenario is NOT blocked today (policy and risk stages unavailable in production) and that it is blocked once those services are reachable. Also recorded: safe-read returns REQUIRE_APPROVAL (trust for a fresh principal is unknown), not the advertised ALLOWED. You can re-run (b) yourself: POST /fabric/try?scenario=prompt-injection on any of the three sites and read the "policy" and "risk" stages. (a) cannot be reproduced by outsiders: the runtime source is private. Same-author,