best-of-nlisted

# Best-of-N Generate several candidate solutions independently, then select one. This beats a single iterated attempt when the solution space is wide and a wrong path is expensive. The evidence favors selection over blending: deliberating teams regress toward their weaker members, losing up to 41% against their best member (arXiv 2602.01011), while non-interactive answer voting stays valid. So this skill picks ONE winner; it never merges opinions into a compromise. ## Sizing N Scale N to stakes and uncertainty, and declare a stopping rule up front. Routine, low-uncertainty work does not need this skill. An uncertain approach at moderate stakes warrants 2 or 3 candidates. High stakes and a wide solution space (a tricky algorithm, a public API shape, a security or money touching change) warrant 3 to 5. Stop early once an oracle-passing candidate cannot be beaten on the oracle, and stop adding candidates when new ones stop differing; log it if you cap. ## Procedure 1. **One precise brief, and an executable oracle if one can exist.** The oracle (a test suite, a compile or type check, a property test, a benchmark threshold) decides correctness without a human or model opinion. Effort spent here makes selection objective. 2. **N candidates, isolated and blind.** Dispatch N implementers from the same brief, each in its own worktree, each a single writer that returns a patch or works only inside its worktree; candidates never write to the repo tree. Vary the a