Loading…
Reading the draws and working out what chance predicts.
Loading…
Reading the draws and working out what chance predicts.
Whole-Draw Drift
Complete draw-sets are compared at every permissible cut and at three declared scales. Every permutation repeats the entire search, so the global reading already pays for choosing the most dramatic boundary.
Ontario 496/49 + 1/49 era · since Nov 202587 draws
Maximum multiscale set-kernel MMDset-kernel-multiscale-v2seed 2026082799 full-search relabellings
Each draw is a set. The kernel is , where the symmetric difference counts balls belonging to only one draw. MMD² compares average within-window similarity with between-window similarity.
The scanner evaluates every permitted cut with 20-, 40-, and 80-draw windows. A null run shuffles the dates and repeats that entire scan. Its statistic is the largest MMD² it finds, so the final p-value pays for selecting both the strongest date and the strongest scale.
A machine change can alter several weak features at once while no individual ball looks dramatic. A whole-draw distance can hear that collective movement. It still cannot distinguish machinery from source changes or errors; those are the first things to audit.