Whole-Draw Drift

Scan every boundary, pay for every look.

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.

Strongest multiscale boundary

Québec Max7/52 era · since Apr 202644 draws

Maximum MMD²: 0.0694
Global scan p: 0.410
95% null maximum: 0.0800
Winning scale: 20 + 20 draws

2026-07-03 → 2026-07-07 · 20-draw sides0.0694

2026-07-14 → 2026-07-17 · 20-draw sides0.0677

2026-07-10 → 2026-07-14 · 20-draw sides0.0675

2026-07-07 → 2026-07-10 · 20-draw sides0.0656

2026-06-30 → 2026-07-03 · 20-draw sides0.0633

Maximum multiscale set-kernel MMDset-kernel-multiscale-v2seed 2026082799 full-search relabellings

The maths, in plain language

Distance between complete draws

Each draw is a set. The kernel is $K(A,B)=\exp(-\gamma|A\triangle B|)$ , 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.

The story behind it

Listening for collective movement

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.