All four entry points classify the visible text only, record one diagnostic token-log line carrying the capped raw output, then raise after strict validation so truncation keeps precedence. The additive keys are passed only on a non-responsive verdict, which is what keeps test_all_generate_entrypoints_reject_blank_clean_visible_output passing unmodified. The brain probe opts out via enforce_responsiveness=False: on a lane whose model reflexively declines, the probe reports generate ok while every real talent raises, which is correct for a liveness probe.