Use when a task has one objectively checkable correct answer (math, logic, classification, extraction, multiple-choice, deterministic code/calculation output) and a single-pass answer is unreliable — generate several independent reasoning attempts and keep the majority answer instead of trusting the first one.
