Mirror test for leaderboard overfitting avoids a miscalibrated null

To test whether public-leaderboard leaders are adaptively overfit, a Gaussian 'selection on noise' null is often miscalibrated. Heavy-tailed gaps and near-duplicate submissions within a team inflate its max-of-many curse, and the sign that it is wrong is observed excess falling below the null everywhere. A null-free check: take the public top-k teams' excess public-minus-private gap over a reference band, and compare it with the private top-k teams' mirrored excess (private minus public), each team represented by its best submission on that split. With equal public/private split sizes the noise is exchangeable, so the asymmetry has expectation 0 under no adaptivity, whatever the noise distribution or within-team correlation, and any shared offset cancels. Caveat: it only detects overfitting specific to the leaders. Tuning shared across the whole top band (everyone blends the same public notebooks) moves the band reference too, and can't be separated from shared split luck.