Propose a rewording from this rule's own record. Writes nothing.

Awaited, unlike a test. One call over text the route assembled — no document to attach, no fan-out — so it answers in a few seconds, which is the shape the surface wants: a reviewer presses Suggest and reads a proposal, then presses Test and watches it drain.

The evidence is read here, not accepted from the client. The dismissals with their reasons and notes, the findings to keep firing on, and the previous tests with the judge's own words. A client-supplied evidence blob would be prompt-injection surface with nothing to check it against, and a proposal a reviewer cannot trace back to a dismissal they recognise is one they have to take on faith.

Both sides always go in. Asking only about the dismissals produces a rule narrowed until it fires on nothing, which scores perfectly and has stopped doing its job.

The session is closed before the call, and the call is raced. Every field the prompt needs is a primitive by the time SuggestionEvidence is built, and underneath propose_requirement sits run_structured_with_usage's unlowerable budget — four attempts with 8-64s backoff on a client with a ten-minute timeout. Holding a pooled connection across that, on a web engine of pool_size=5, max_overflow=24, pool_timeout=5, is how five reviewers pressing Suggest take every route down for every tenant; and with no race, the request itself can outlive any proxy in front of it. The test path was shaped this way from the start (helpers/requirement_tests); this is the same discipline on the awaited half.