The Price of Fairness: Qualification Reliability in Agentic Poker Tournaments

Agentic poker competitions select finalists from noisy outcomes under finite simulation budgets. We study top-six qualification in 60-agent, six-player no-limit hold'em Swiss tournaments using three controlled fields and two mechanism-specific reference estimands. At the baseline of 40 rounds with 400 hands per match, independent reruns change 3.83 to 4.38 of six qualifiers on average. Even in an engineered separated control, six independently validated leaders qualify individually with probabilities of only 0.31 to 0.76. Selection improves at every observed checkpoint from five to 40 rounds, without an observed plateau. At an equal scheduled hand budget, ten 1,600-hand rounds produce 0.42 to 0.87 more changed qualifiers than forty 400-hand rounds. Adaptive Swiss is more consistent than fresh-randomness replay of its realised schedules, although this comparison does not identify adaptation's causal effect. A short-match duplicate package improves ranking consistency across all three fields; improvements in changed qualifiers and regret are individually clear only in the separated control. Most wrong qualifiers lie outside the reference-ambiguous set, and reference uncertainty accounts for little of the measured strength loss. These results favour distributing scheduled simulation across more rounds in the tested setting, while leaving the budget required for stable qualification unresolved.