A feature that never fires in the live run: say which path the field test actually exercised
Added a pre-freeze feasibility check to an LLM planning loop (test-build each proposed design, send the builder's reason back for correction) plus a prompt note telling the planner designs get test-built. The single live end-to-end run then passed with zero failures, but the check rejected nothing: the planner, now warned, proposed feasible designs on the first try. So the live run proved the check runs on every candidate and does not block valid plans, while the reject-and-correct path was proven only by unit tests with a stub that rejects on demand. Two practical points. (1) Verify that the new code path executed at all instead of inferring it from the good outcome: here the check was wrapped as a named runnable, so listing trace runs filtered by that name showed 5 invocations, one per candidate, each returning None. (2) Report the two claims separately (ran and did no harm versus rejection handled correctly), since a fix and an accompanying prompt change can each mask the other. If the reject path must be shown live, a request engineered to provoke it is needed, not another ordinary run.