1. The Validation Principle
A statistical tool is only as good as its ability to detect the truth. To validate the Lucky Picks Fairness Score, we don’t just run it on real lotteries—we run it on synthetic lotteries where we control the truth.
We developed a simulation engine (trust-score-experiment.js) that generates thousands of draws under specific, “rigged” conditions to prove that our algorithm correctly identifies them.
2. Stress Test Scenarios & Results (v3.1)
We subjected the production Fairness Score engine (v3.1) to our synthetic stress-test suite (trust-score-experiment.js). This version enforces the “Weakest Link” integrity policy with a 10.0 failed-component floor and a +10.0 critical-cap recovery buffer (ScoreFinal = min(ScoreWeighted, ScoreFailingComponent + 10)), ensuring that a critical failure in a direct randomness component (such as a biased bonus drum or pattern bias) caps the overall score.
For a full breakdown of the v3.1 algorithm, please read our Methodology Report.
- Verification Run Date: November 15, 2026
- Algorithm Version: Fairness Score Engine v3.1 (
lotteries/stats/trust-score.js) - Lottery & Sample Configuration: 5/69 main drum + 1/26 bonus drum (Powerball-style), 1,000 synthetic draws per iteration (3-day UTC draw cadence starting 2020-01-01) repeated over 10 independent iterations per scenario (50 draws in the small-sample check).
- Test Parameters: Scan windows
[100, 200, 500, Overall], 500 Monte Carlo permutations per scan statistic, significance thresholdα = 0.05, failed-component score floor10.0, critical-cap buffer+10.0.
Scenario A: The Control Group (Fair Game)
- Setup: Pure uniform PRNG generation (
Math.random()) for all main and bonus numbers across 1,000 draws × 10 iterations. - Expected Result: Pass (Score 90–100).
- Observed Score (Mean / Range): 99.3 (range
97–100) - Verdict: ✅ VALID. The system consistently scores a fair game near 100 without false-positive penalties.
Scenario B: The “Hot Number” (Biased Main Drum)
- Setup: One specific number in the 69-ball main drum is given a 3x selection weight while the bonus drum remains fair.
- Expected Result: Fail (Score < 50).
- Observed Score (Mean / Range): 38.0 overall (range
38–38; Main Drum component score: 10.0, combined 80/20 Frequency dimension: 28.0, capped at28.0 + 10.0 = 38.0). - Verdict: ✅ VALID. The Main Drum hits the 10.0 failure floor and triggers the Weakest-Link cap across every iteration.
Scenario C: The “Hot Bonus” (Biased Bonus Drum)
- Setup: The 5/69 main numbers are completely fair, but one Bonus Ball in the 1/26 drum is given a 5x selection weight.
- Expected Result: Critical Fail (Capped Score ≤ 20).
- Observed Score (Mean / Range): 20.0 overall Fairness Score (range
20–20; Bonus Drum component score: 10.0). - Why both 10.0 and 20.0 appear: In v3.1, the failing Bonus Drum hits the 10.0 component floor, and the Bonus Drum Integrity Check caps the overall lottery score at
10.0 + 10.0 = 20.0. (In our earlier December 2025 v2.1 build, before the 10-point component floor was added, a failed drum returned0.0, capping the overall score at10.0.) See Section 3 for the full walkthrough. - Verdict: ✅ VALID.
Scenario D: Pattern Bias (80% Even)
- Setup: Individual numbers remain spread across the range, but each draw slot is forced to pick Even Numbers 80% of the time.
- Expected Result: Critical Fail (Score < 50).
- Observed Score (Mean / Range): 19.9 (range
19–20; Pattern component score: 10.0, capped via Weakest-Link propagation). - Verdict: ✅ VALID. The Pattern Analysis module flags the combinatorial parity distortion in 100% of runs.
Scenario E: Seasonal Parity Shift (Temporal Scan)
- Setup: “Even” numbers are forced 80% of the time only in December, leaving the other 11 months completely uniform.
- Expected Result: Temporal Anomaly Flagged for Monitoring.
- Observed Score (Mean / Range): 78.9 overall (range
71–81) with an active Temporal Anomaly Monitoring Flag (Temporal subtest drops into the critical zone while Frequency and global Patterns stay near 100). - Verdict: ✅ MONITORED (By Design). In v3.1, Temporal Analysis already applies its own internal weakest-link cap across multiple subtests; rather than double-applying the global critical cap on an isolated calendar-month anomaly, the engine lowers the composite score out of the “Excellent” band and attaches an explicit “Temporal anomaly flagged for monitoring” warning.
Scenario F: False Alarm Stress Test (FWER & Small Sample)
- Setup: Fair uniform simulations tested at both full scale (1,000 draws × 10 iterations) and small-sample scale (50 draws × 10 iterations, Scenario I). Without multiple-testing protection, ~5% of window scans would trigger false alarms at
p = 0.05. - Expected Result: 0 False Positives (100% ≥ 90).
- Observed Result: 0 False Positives (mean 99.3 at 1,000 draws; mean 99.4, range
94–100at 50 draws). - Verdict: ✅ VALID. Monte Carlo Permutation Testing and Holm-Bonferroni pattern correction prevent false alarms across overlapping windows.
Scenario G: The “Subtle Nudge” (Sensitivity Calibration)
- Setup: A mild single-ball bias (
weight = 1.2in the generator, roughly doubling one ball’s appearance rate versus a 3x spike in Scenario B). - Expected Result: Monitor Band (Score 55–80).
- Observed Score (Mean / Range): 69.4 (range
56–77, 100% below 80). - Verdict: ✅ VALID. Dynamic Cramer’s V effect-size scaling places mild drift in the monitoring tier rather than either ignoring it or misclassifying it as a total RNG breakdown.
Scenario H: Extreme Pattern (Consecutive Sequences)
- Setup: The generator forces 5-number consecutive runs (e.g., 12-13-14-15-16) in 50% of draws.
- Expected Result: Critical Fail (Score ≤ 20).
- Observed Score (Mean / Range): 20.0 (range
20–20). - Verdict: ✅ VALID. Both the Consecutive Pattern test and the Duplicate-Draw Collision check (~436 collisions vs. 0.04 expected) fire immediately, capping the score at 20.0.
Scenario J: The “Eddie Tipton” (Targeted Date Sniper Attack)
- Setup: Modeled on the historic “Hot Lotto” fraud, where the RNG behaves fairly all year except on 3 target dates per year (May 26, Nov 23, Dec 29) when a fixed 5-number set is forced (~9 rigged draws across 1,000 draws).
- Expected Result: Fail (Score < 60).
- Observed Score (Mean / Range): 53.8 (range
52–54). - Verdict: ✅ DETECTED. Even though 99.1% of draws are fair, the v3.1 Birthday-Problem Collision Test detects 9 exact 5/69 repeats (0.04 expected by chance, Poisson upper-tail
p < 10-4), applying a 70-point main-drum integrity penalty (Main score 30 → Frequency 44 → Overall capped at 54).
3. Case Study: Reconciling 10.0 vs. 20.0 in the “Hot Bonus” Fix
During internal auditing of Version 2.0, we uncovered a dilution flaw: in Scenario C (Hot Bonus), the pre-cap algorithm returned 79.2 / 100 because fair Main Drum (5 balls), Temporal, and Pattern scores averaged away a rigged Bonus Drum.
How the Integrity Cap Works in v2.1 vs. v3.1 We introduced the Component-Based Integrity Check, which treats the Main Drum (when pool > 20) and Bonus Drum (when pool > 20, such as Powerball’s 1/26) as critical dependencies:
- Component Score = 10.0 (v3.1 floor): Under v3.1, any single drum or pattern test that suffers maximum statistical and effect-size penalty is floored at 10.0 (
max(10, 100 − penalty)) rather than0.0. In Scenario C, the rigged Bonus Drum therefore scores 10.0 / 100, which also pulls the Frequency dimension to 10.0 / 100. - Final Overall Fairness Score = 20.0 (v3.1 capped result): Because the Bonus Drum is below the
50.0critical threshold, v3.1 caps the overall composite Fairness Score atScoreFinal = min(ScoreWeighted, ScoreBonus + 10) = 10.0 + 10.0 = 20.0 / 100across all 10 simulation runs. - Why earlier documentation showed 10.0: In the December 2025 v2.1 prototype (prior to adding the
10.0component floor), a failed drum scored0.0, so the same+10.0cap formula yielded0.0 + 10.0 = 10.0. Both figures reflect the same Weakest-Link cap operating before and after the10.0component floor was standardized in v3.1.
4. Reproducibility
We believe in open verification. The core statistical libraries used in our analysis are standard, open-source packages:
chi-squared: For goodness-of-fit testing.jstat: For distribution analysis.
The simulation parameters described above can be replicated by any researcher to verify our findings.