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

Hoosier Lotto +PLUS6/46 era · since Apr 2022300 draws

Maximum MMD²: 0.0827
Global scan p: 0.830
95% null maximum: 0.0974
Winning scale: 20 + 20 draws

2024-09-04 → 2024-09-07 · 20-draw sides0.0827

2026-08-01 → 2026-08-05 · 20-draw sides0.0821

2026-03-14 → 2026-03-18 · 20-draw sides0.0818

2026-03-18 → 2026-03-21 · 20-draw sides0.0813

2026-08-08 → 2026-08-12 · 20-draw sides0.0805

2026-03-07 → 2026-03-11 · 20-draw sides0.0796

2026-03-21 → 2026-03-25 · 20-draw sides0.0794

2024-08-31 → 2024-09-04 · 20-draw sides0.0789

2026-08-05 → 2026-08-08 · 20-draw sides0.0781

2026-03-11 → 2026-03-14 · 20-draw sides0.0781

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.