2025-12-30
We rank values across st1–st5 using gap percentiles, split the ranked list into buckets, and backtest where the next draw lands.
Read: Lab Report — Percentile Ranking Across Columns (Bucket Test Results) →Lab Blog
New concepts, datasets, and “try this at home” experiments. Same goal every time: more structure, fewer random guesses.
2025-12-27
A spacing-based method for rich-state features (like sum and OSLCS): build a contrarian candidate corridor and a forbidden “avoid” list using follower-exclusion + spacing stats. Download full docx file.
Read: Contrarian & Forbidden Spacing Strategies for High-Cardinality Features →2025-12-26
A practical approach as to test the covering with different mappings. Get the script. Download here.
Read: Random Mapping of a min Covering → Download code →2025-12-26
We scan 5–6 x-columns and 0/1 patterns, filter tot_df + hist_df in lockstep, then rank candidates using heavy-lift (ratio_of_ratios), delay-percent due-ness, plus a support penalty so tiny subsets don’t fool us.
Read: Heavy-lift x-pattern filter (ranked search) → Download code →2025-12-15
A compact set of 5-number lines built so every possible winning draw shares at least 4 numbers with at least one line in the set.
Read: C(50,5,4) Min-Covering - a “4-hit safety net” dataset →