How Evidraw works.
Evidence, explained.

These methods describe the current Learning Lab Explorer, its historical replay and account-linked experiment tracking. Other reports may use different samples or metrics; check the method and dates shown on each report. No score on this page is a calibrated probability of winning.

The Lab uses Evidraw's official-results archive. Check the displayed source, date range and available draw count before interpreting a chart. A cached archive is not a guarantee that the latest draw has arrived.

Replay validates dates, unique main numbers, number ranges and duplicate draw dates. It separates the current Lotto Max format from draws before April 14, 2026, and the current Powerball format from draws before October 7, 2015. These are explicit rule cutoffs—not automatic detection of machine changes or hidden regimes.

Sums, parity, gaps and pair analysis use main numbers only. Bonus and special balls do not enter these charts. A main-number match score must not be read as a jackpot or payout result, particularly for games with a separate special ball.

The generator builds lines incrementally, favouring less-used numbers and less-repeated pairs, with seeded random tie-breaking. It rejects duplicate full selections. This is a practical greedy heuristic, not a search of every possible portfolio or a guarantee of the global optimum.

Typical sums, odd/even balance, previous-draw repeats and selected historical signals influence choices within that process. In the Explorer, a final guard compares the custom portfolio with pure coverage at the same seed and line count. If unique-number or pair coverage falls, or maximum overlap increases, it restores pure coverage. A preference can therefore have little or no visible effect.

The usage map, unique-number and pair counts, maximum overlap and average overlap describe how exposure is distributed. They are not winning-probability scores. In a fair draw, equal numbers of distinct valid complete selections have equal jackpot chances; broader pair coverage does not change that.

Explorer charts offer 30-, 60- and 150-draw windows, subject to available history. The signal mix uses recent frequency over up to 30 draws, longer frequency over up to 150, pair support over up to 60, and gaps within the supplied history. Selected signals are normalized and equally averaged; these are descriptive weights, not learned predictive probabilities.

Sum and odd/even reference distributions come from counting valid combinations under fair sampling without replacement. Common categories contain more combinations—not individually luckier lines. Gaps for numbers unseen in a window are censored at that window, not their lifetime absence.

For a fixed pair in a game drawing k numbers from N, its per-draw probability is k(k−1) / [N(N−1)]. The pair table uses an upper-tail binomial test across independent draws and a Bonferroni adjustment over all N(N−1)/2 possible pairs. This controls that one pair family, not every window, game or strategy you might try. An adjusted p-value is not the probability that a pair will appear next.

For each target draw, replay rebuilds signals using strictly earlier draws, constructs the portfolios, then scores the target. Your custom mix, pure coverage and random comparison use the same line count and target dates. Random lines use seeded shuffling without replacement within a line and reject duplicate full selections within a portfolio. The random comparison is not filtered to copy your preferences.

Five deterministic seed trials per target show some seed sensitivity; they are not five independent lottery draws. Repeats-enabled runs skip targets whose immediately preceding scheduled draw is missing. Reports show skipped draws and results by seed and period. Removing one option at a time is a sensitivity check, not a full factorial experiment or proof of what caused a result.

Strict chronological inputs prevent one form of data leakage. They do not undo hindsight from choosing settings after seeing past results, repeatedly rerunning tests, or selecting the most favourable window. These reports remain exploratory—not a certified untouched holdout or evidence that the next draw is predictable.

Earlier draws required

30

Target-draw options

15 / 30 / 60

Seed trials per target

5

Practice lines per portfolio

1–25

Account-linked tracking requires an owned saved experiment, the matching game, and eligible generated lines recorded as played before the app's draw cutoff. Creation must precede the played timestamp; the recorded history must precede the target. Missing results stay pending. Drafts, ineligible entries and older label-only associations do not become verified experiment outcomes. Recording a line is not independent proof of buying a ticket.

Eligible scored portfolios are compared with 200 seeded random portfolios using the same played-line count. The metric is the best main-number match in each portfolio, averaged across random trials—not profit or jackpot probability. Runs on the same game and draw date are averaged together before summarizing across draws, so multiple portfolios do not masquerade as independent draws.

The current 95% interval is a descriptive normal approximation: mean difference ± 1.96 × sample standard deviation / √n, withheld until 30 independent draw dates. It is not corrected for repeatedly checking results or testing many strategies. An interval above zero is an observed lead on this metric, not proof of a persistent edge. An interval including zero does not prove equivalence. Tracking currently displays up to 100 recent linked or legacy records and indicates truncation.

Saved experiment definitions preserve the selected settings; make a new version to change an idea. Historical report exports retain settings, seeds, dates and an archive fingerprint. Matching these inputs lets a run be compared or repeated; the fingerprint is a reproducibility aid, not a cryptographic audit signature. Browser-generated reports are not independently certified.

The animated pipeline is an explanation replay, not a live trace of autonomous agents discovering winning numbers. Advanced exploration does not silently change your active generator settings. Review and explicitly apply an experiment when you want to use it.

Evidraw has not established a predictive jackpot advantage. We do not present historical rankings as calibrated next-draw probabilities, claim comprehensive correction across every experiment, or promise globally optimal portfolios. Broad multiple-testing controls, independent holdouts and longer prospective samples would be needed before making stronger statistical claims.

Use a fixed entertainment budget. A more detailed model is not a reason to buy more lines. The useful outcome can be a clearer record, a better-organized portfolio—or evidence that an idea does no better than random.