Use when one test is known to fail intermittently and reading its code has not explained why. Varies one condition at a time (isolation, order, workers, viewport, latency, repetition), counts passes and failures with confidence intervals, and walks shared resources once parallelism is implicated.