We run a population of model instances that publish pre-event probability calls on football fixtures. Every call is written to ai_predictions at creation, graded exactly once after the event, and stor
Every prediction system eventually grows a lifecycle manager: models get born, promoted, demoted and killed. Ours has 8,789 model rows and a retirement pipeline that runs every two hours. For a long t
Every prediction is a claim about a world state that does not exist yet, which makes its input side auditable in a way most ML pipelines never face: the feature vector must be reconstructable as of a
A three-class classifier that reports one label per row is not a three-class classifier. It is a ranking function collapsed to a decision, and the collapse is where the information dies. On our platfo
Every analytics platform eventually ships a number nobody can trace. Ours was the hit rate. It appeared in four places — the model table, the leaderboard, the per-competition rollup, the daily stats c
Most prediction systems ship two numbers: a call, and a confidence. The confidence is usually an integer — 1 to 10, 60 to 95, whatever reads well in a UI. Everyone downstream then treats it as a proba
Most teams treat "reproducibility" as a machine learning problem. Fix the seed, pin the library versions, snapshot the model weights, and you are done. That definition breaks the moment the prediction
I run a platform where a couple of hundred independent models publish probability estimates on football fixtures. They aren't variants of one model with different seeds — each one has its own training