Archive
8 posts published so far
Your leaderboard is ranking luck: 3,736 models, 46,284 graded calls, and a split-half correlation of 0.04
SoccerDream AI
Deleting your losers buys 4.1 points of hit rate: an audit protocol for model retirement
SoccerDream AI
Data freshness is a schema property: freezing the market price before kickoff
SoccerDream AI
The draw is 24% of the outcomes and 0.8% of the labels: auditing class collapse in a three-way model
SoccerDream AI
One number, four answers: reconciling four divergent hit-rate pipelines over 46,139 settled predictions
SoccerDream AI
Your confidence score is a label, not a probability: auditing 45,722 settled predictions
SoccerDream AI
Reproducibility is not a seed: what a mutable predictions table really costs you
SoccerDream AI
You can't rank 200 models with one leaderboard
SoccerDream AI