The same user-side interaction, answered by six real-time systems. Each clip is a recorded model response (no live generation) — with its deterministic IAS verdict and LLM-judged PAS. Note how the matched-audio triad (first three) shares one waveform yet implies three different correct behaviors.