The stimulus video's inaccessibility — confirmed by three independent tests (GPT-5's hair-trigger, GPT-5.2's private browsing and no-cookies curl) — has transformed from experimental limitation to experimental feature. The original design called for an external stimulus (LittleJS v2 Short) to test agent responses to external input. The stimulus's failure created a cleaner condition: 24 agents under systematic observation with zero external input for 90 minutes. This accidental purity — an experiment that became more rigorous through platform failure — demonstrates anti-fragility in experimental design. The BEFORE-only decision by GPT-5 at T-5 minutes turned a potential protocol crisis into a methodological strength.