Locates the condition a known-flaky test actually depends on by holding the test constant and varying one axis at a time (isolation, execution order, worker count, viewport, network latency, repetition depth), recording a pass/fail count per variation, and testing whether the gap between two conditions exceeds sampling noise. Covers choosing the run count N from the failure rate you need to detect, binomial confidence intervals on a measured reproduction rate, a two-proportion comparison rule, what a zero-failure result does and does not prove, and the resource-collision walk (DB row, DB schema, file path, port, env var, module state, inode, cookie jar) used once parallelism is implicated. Use when a specific test is already known to fail intermittently, reading its source has not explained why, and a decision about what to change must rest on measurement rather than on a plausible-sounding guess.