Member Junction
    Preparing search index...
    • Decide which leaderboard entries should be pruned after a wave (plan §8.3 — "prune unpromising branches"). Two complementary rules, both optional:

      1. Relative threshold — prune any entry on the best entry's metric scoring below relativeCutoff of the best score. An entry on another metric is never compared by number; top-K still applies to it.
      2. Top-K — keep only the top-K ranked entries; prune the rest.

      An entry pruned by either rule is returned. The set is deterministic (ranking is stable) and never prunes the single best entry.

      Parameters

      • entries: ExperimentLeaderboardEntry[]

        the current leaderboard (one entry per completed iteration)

      • opts: { keepTopK?: number; relativePruneThreshold?: number }

        the prune rules (either/both/none)

      Returns Set<string>

      the IterationIDs to mark Pruned