$ ~/aibattle/q-and-a

1. What does “aggression” mean? How often a model attacks instead of going along with the action — used identically on every page: aggression = (bet + raise + all-in) ÷ (bet + raise + all-in + call + check) Folds are not counted. 0% = never bets/raises (pure caller); 100% = always bets/raises.

2. What is “corr with win%” (correlation)? Shown on some tables as e.g. corr with win%: +0.61 — the Pearson correlation between a metric and match win rate across the models, from −1 to +1: +0.9 ≈ move together almost perfectly, +0.6 = clearly related, 0 = unrelated, negative = move in opposite directions. With only ~11 models, read values above ~+0.58 (or below −0.58) as a real signal and smaller ones as weak.

3. What is “equity”? A hand's chance of winning if all the cards were dealt out, estimated by simulation (deal the opponent a random hand, finish the board, repeat thousands of times). Caveat: the opponent's real cards are never in the logs, so equity is computed vs a random hand, not their actual range — a rough, range-free proxy. That is why the equity-based “Decision quality” block is labelled experimental.

4. How is a “bluff” detected? A bet/raise/all-in made with a weak hand — specifically when the model's equity (vs random) is below 40%, so it is betting a hand that is probably behind, to push the opponent off. bluff success = of those bluffs, how often the opponent actually folds. (Pure bluffs and semi-bluffs are lumped together, and 40% is a chosen cut-off.)

5. What is the “gear-shift” (ahead vs behind)? How a model changes when it is losing on chips. Strong players, when behind, raise more and fold less (they fight for pots); weak players play the same whether ahead or behind. Each cell shows the value when ahead → when behind.

6. Win rate vs Elo — why are they different? Win rate is just the share of matches won. Elo is opponent-adjusted — beating strong opponents is worth more than beating weak ones, and it accounts for who each model actually played. A model can have a decent raw win rate but a lower Elo if its wins came against weaker opposition, so the leaderboards are ordered by Elo as the fairer ranking.

7. How do you serve the open-weight models? All open-weight models are served via Fireworks AI (serverless inference) through one OpenAI-compatible client; the closed models (Claude, GPT-5.x) run on their own provider APIs.

8. How is the overall ranking and Elo calculated? The overview leaderboard shows two cross-environment summaries over the core head-to-head strategic settings: Connect Four, Gomoku, Hold'em 1-Hand, Hold'em Match, and Leduc Holdem. On that table, click a score header to sort by it, or the ⓘ for how each is computed. Coverage = environments a model has entered (treat low coverage as provisional).