Tournament selection for model populations

Why tournaments

Model-breeding systems need a selection mechanism. A single global score is tempting, but it hides trade-offs. Tournament selection compares candidates across task, robustness, calibration, latency, cost, diversity, and safety dimensions.

A tournament does not mean every candidate fights every other candidate in production. It means each candidate faces a structured evaluation bracket against the current champion, no-op, and relevant specialists.

Bracket design

FUNCTION run_selection_tournament(candidates, champion, ecology, policy)
    bracket <- []
    FOR each candidate IN candidates
        scorecard <- EVALUATE(candidate, policy.suites)
        scorecard.noop_delta <- COMPARE_TO_NOOP(candidate, ecology)
        scorecard.champion_delta <- COMPARE_TO(champion, candidate)
        scorecard.complementarity <- MEASURE_NON_OVERLAP(candidate, ecology)
        bracket.ADD(scorecard)
    END FOR

    eligible <- FILTER(bracket, hard_invariants_pass AND noop_delta > policy.minimum_gain)
    ranked <- SORT_BY_NET_VIABILITY_AND_DIVERSITY(eligible)

    RETURN ranked.FIRST_OR_NOOP()
END FUNCTION

Avoiding bracket gaming

Keep holdout sets private from candidate generation. Freeze suite versions during a tournament. Record all failed candidates, not only winners. Failed candidates are useful evidence because they reveal which operators are wasting budget.

Multi-winner tournaments

Sometimes the right output is not one winner. A candidate can lose as a general champion but win as a narrow specialist. Use lifecycle states to reflect that: candidate, shadow-specialist, canary-specialist, champion-route, or retired.