Scoring Audit

AI Analysis

Flagged Criteria

• Even Spacing: -13.1pt drift (≈2.0σ) is borderline notable but within 2σ threshold (~±14); monitor next 2-3 weeks for pattern confirmation

Suggestions

• Track Even Spacing and Spread (which spiked +11.8pt) over next 3 weeks to confirm whether drift persists or reverts to mean
• No evidence of unfairness; all variation explainable by random sampling in 50-draw windows

Current Weights

Split-avoidance release — the score is unchanged; what changed is which qualifying tickets subscribers receive. (1) No shared pool: each subscriber's picks are drawn independently from the full set of qualifying combinations, so subscribers no longer hold copies of one another's tickets (with the old 500-ticket pool, about a third of tickets were duplicated across the subscriber base). (2) Crowd gate: of the 275,193 combinations that pass all nine criteria, only the 187,303 with at most two numbers from 1–31 are eligible. Per-tier winner counts over 1,412 draws show sets with four or five such numbers are held by about 1.5× as many players as average, and sets with at most two by about 7% fewer (0.93×). (3) Powerball tilt: the red ball is drawn in proportion to how under-played it is (index^−6), so the most over-played ball (7, 1.13×) is drawn about a third as often as the least played (1, 0.93×) — a mild lever, since Powerball popularity is nearly flat. All three change only how many people would share a prize, never the odds, which are identical for every combination. The crowd constants are a dated snapshot and change only with a versioned revision.

Analytic bands and fixed weights. Sum Range and Spread now score against bands derived exactly from the 5/69 matrix (sum: mean 175, σ 43.2; spread: mean 46.7, σ 11.9) instead of the historical mean and σ, so a combination's score no longer shifts as draws accumulate and the model has no dependence on past results at all. Retired the weekly audit's automatic weight adjustment: weights are set by combinatorial filter strength in code and change only with a versioned revision — the audit remains a fairness monitor. The selection pool now keeps only combinations that pass every criterion (about 2.4% of all combinations) and de-duplicates them. Weights unchanged from v7.0.

Current-matrix calibration + honest reweighting. All analysis now scoped to the current 5/69 era (draws from 2015-10-07 on); ~591 older draws from the retired 5/59 matrix were being mixed in, biasing every frequency, σ-band, and drift figure. Verified against a 5M-combination Monte-Carlo: on current-era data all criteria sit within their 95% combinatorial bands (the lottery is fair). Recomputed every weight from exact combinatorial filter strength — weight ∝ (1 − pass rate), normalized to 100 by largest-remainder rounding; the prior v6.0 weights deviated from this formula by up to 7 pts (Even Spacing was the 2nd-hardest filter yet weighted lowest). Dropped Range Coverage as redundant (φ up to 0.41 with Sum Range, 0.33 with Tens Diversity) — down to 9 criteria.

Purely combinatorial model. Replace Consecutive Pair (empirical, 26.6% pass rate) with Modular Balance (all 3 residues mod 3, 64.3% pass rate). Remove Drift Rebalance overlay (empirical mean-reversion assumption). Reweight all criteria by filter strength — weight proportional to (1 - pass rate), normalized to 100. Unique Digits (35.1% pass rate) now highest at 17 pts. Protected criteria updated to Parity, High/Low, Unique Digits. All 10 criteria now derivable from pure combinatorial math with zero reliance on historical patterns.

Add Drift Rebalance overlay (+0-7 bonus pts). Compares last 20 draws against full-history baselines across 8 structural metrics (parity, high/low, sum, digit sum, primes, gap entropy, consecutive pairs, decades). Picks that lean against metrics drifted beyond 1σ earn up to 1 pt each, capped at 7. Operates at the distribution level only — no individual number targeting. Base 100-pt model unchanged.

Remove gambler's fallacy criteria. Replace Co-occurrence (5 pts, worse than random prospectively at 36% vs 50%), Hot/Cold Mix (5 pts, 35.6% vs 50%), and PB Weighting (5 pts, 70.1% vs 75%) with three combinatorially-grounded criteria: Range Coverage (6 pts, all 3 thirds represented, 62%), Tens Diversity (5 pts, 4+ tens groups, 66%), and Even Spacing (4 pts, all gaps ≤ 2x ideal, 59%). All criteria now verifiable from pure combinatorial math with zero reliance on historical patterns persisting.

Major model overhaul. Remove Decade Spread (97% pass rate, no discrimination) and Drought Bonus (gambler's fallacy, −20 drift). Add Consecutive Pair (6 pts, 28% of draws have one) and PB Weighting (5 pts, PB drought above 25th percentile). Graduate Sum Range (18/9/0 at 1σ/1.5σ), Spread (13/6/0), and Unique Digits (7/3/0 at 5/4 unique). Loosen Hot/Cold Mix to include 2+ hot or 2+ cold. Reduce Primes 8→5 pts, bump Co-occurrence 4→5 pts.

Replace Anti-Birthday (4 pts) with Co-occurrence (4 pts) — scores picks containing historically co-occurring number pairs. Recalibrate Drought Bonus thresholds (35→20 floor, 80→60 divisor) to fix −21.4 drift.

Redistribute points to maintain 100-point scale. Parity/High-Low/Sum Range 15→18 pts each, Spread 10→13 pts. Max stays at 100.

Drought Bonus 10→5 pts, Decade Spread 8→4 pts, Anti-Birthday 7→4 pts. First AI audit-driven recalibration.

Initial scoring model with 10 criteria, max 100 pts.