Expand description
Tournament — rank competitors by pairwise comparative judgment.
Each competitor produces one candidate answer to the task, then a fresh judge
agent compares candidates two at a time and picks the better one. Unlike
Vote (single-shot majority over independent answers) or
AdversarialReview (pass/fail against a spec), a
tournament produces a relative ordering — the blog-style “sort a large set
by comparative judgment / run a bracket until a winner emerges” idiom.
Single-elimination: competitors are paired each round, the judge picks a
winner per pair, winners advance, and an odd one out gets a bye. After
ceil(log2(n)) rounds one winner remains. The round at which a competitor is
eliminated yields a coarse ranking (later elimination = higher rank).
Pairwise comparison is more reliable than asking a model to score N items on an absolute scale: a judge comparing exactly two answers has a much easier, lower-variance decision.
Structs§
- Match
Result - One pairwise judgment in the bracket.
- Tournament
- Rank competitors by single-elimination pairwise judgment.
- Tournament
Result - Result of a tournament.