What the engine is

The Enhanced Model is a rule-based selection engine. It runs entirely on your iPhone, weighs less than 1 KB, and produces a pick in under one millisecond. It is not a neural network, contains no trained machine-learning weights, and never calls a server during prediction.

The internal logic that produces the bias toward winning combinations is the proprietary part of the product. We treat the recipe the way a quantitative-finance signal vendor treats theirs: the validation discipline is published, the recipe is not. What follows on this page is everything an independent researcher needs to verify the rigor of the result, not the recipe to reproduce the engine.

How we validated it

We applied evaluation discipline borrowed from clinical-trial and quantitative-finance practice. Every choice below was made to remove the standard ways lottery-prediction studies fool themselves:

Walk-forward backtesting

For each evaluation draw at index t, the model received only draws[0:t] as input, every draw before the target, and nothing after. There is no path by which future information could leak into a prediction. Walk-forward is the standard for honest evaluation in finance and machine learning, and most published lottery-prediction studies that look impressive in research papers fail outright when they're forced to play by these rules.

Paired comparison against uniform random

For every trial, the model and a uniform-random baseline received identical target draws and identical RNG seeds. The two sides differed in only their selection logic. All statistical tests are paired, we compare the model and the baseline on the exact same draws, never on parallel-universe samples.

Seed-level aggregation (most conservative)

The statistical unit is one seed's average match rate across all evaluation draws. With 10,000 seeds, that's N = 10,000 paired observations. We did not treat each individual draw as an independent observation, which would have inflated significance dramatically. Seed-level is the most conservative inference choice.

Multiple-comparison correction

Four primary statistical tests were run (head-to-head wins, match-rate difference, 3+-match-rate difference, cumulative extra matches). Every result was corrected for multiple comparisons using the Benjamini-Hochberg false-discovery-rate procedure. All four tests survive correction. All four bootstrap 95% confidence intervals (10,000 resamples) exclude zero.

Pre-registered model, single comparison

The production engine's configuration was selected on a development split, frozen, and then evaluated on the held-out test period. There was no parameter-fishing on the test set. One pre-registered model versus one baseline; no multiple-model selection bias to correct for.

Literature comparison

Among the published lottery-prediction methods we surveyed, none has produced a statistically significant walk-forward lift on a major lottery game when evaluated under the discipline above. Most produced results within ±0.1% of random chance.

"We surveyed" is load-bearing, we don't claim to have reviewed every lottery study ever conducted. We claim that, among the systems with publicly available methodology that we examined, none has produced a comparable result under proper walk-forward evaluation.

Reproducibility

The validation methodology is fully reproducible. An independent researcher building any candidate lottery-prediction system can apply the same evaluation framework to test it:

1. Fetch Powerball draws from the public NY Open Data API: https://data.ny.gov/resource/d6yy-54nr.json. Pagination is standard Socrata; filter to current rules (main balls 1, 69, Powerball 1, 26).
2. Implement a uniform random baseline: 5 main balls without replacement from 1, 69, 1 ball from 1, 26.
3. Implement and freeze your candidate model on a development split.
4. For each evaluation draw, give the candidate only the prior draws (no future leakage).
5. For each of N seeds, generate paired (candidate, baseline) tickets using the same RNG seed; score main-ball matches per ticket.
6. Run the headline statistical tests on the paired seed-level averages: a binomial test on head-to-head wins, a Wilcoxon signed-rank test on match-rate differences, and a paired t-test. Compute bootstrap 95% confidence intervals.
7. Apply Benjamini-Hochberg FDR correction across all primary tests.
8. Check whether your CIs exclude zero and your corrected p-values cross your significance threshold.

Anyone can verify the discipline. The result we report is what survives that discipline.

The full machine-readable evaluation artifact for the Enhanced Model is published at proof/validation_proof.json.

What is and isn't published

To protect the engine from trivial cloning while keeping the validation honest, the production model's tuned hyperparameters, specific window sizes, bias-strength bounds, smoothing constants, and the candidate-selection cross-validation procedure, are not published on this page. The engine is small enough that, if it were published, anyone could clone it in an afternoon, and the small revenue from a small subscription business would not survive that. This is the same line that quantitative finance and trading-signal vendors draw: we publish what we tested and how we tested it, not the parameter set that does the work.

What an independent researcher gets from this page:

• The runtime characteristics of the engine (rule-based, on-device, sub-millisecond inference, no trained weights, no network calls).
• The complete validation methodology, including every statistical-test choice and the multiple-comparison correction.
• The exact data source, with the public API endpoint.
• The reproducibility framework, enough to apply the same evaluation to any candidate model.
• The machine-readable result artifact for the production engine.

That is sufficient to verify the rigor of the result without giving away the parameter recipe.