LottoExpert Original Research

What 38,005 Scored Lottery Predictions Can—and Cannot—Teach an AI System

A transparent census of LottoExpert prediction records scored against their recorded official drawing. The central finding is simple: a prediction can show how it performed without proving which setting caused that result.

Published August 31, 2026 Data through August 31, 2026 Aggregate data only Version 1.0

The short answer

Scoring tells us whether a saved prediction matched an official result. Learning requires an extra proof: the system must also know exactly which settings produced that prediction. LottoExpert keeps those two facts separate so that a large record count cannot be mistaken for reliable settings evidence.

38,005saved pre-draw prediction records with an official score
142canonical lottery game IDs represented
1,934distinct game-and-drawing combinations
4,581distinct settings fingerprints recorded
Original finding Across the SKAI-labelled portion of the census, 33,760 records had official scores, but only 4,272 records—12.65%—currently met the stricter contract required to influence Hive learning. The rest remain useful as performance history, but they are not allowed to claim that a particular setting caused the outcome.

Two kinds of evidence, explained plainly

1. Scoring evidence

“What happened?”

  • The prediction was saved before its target drawing.
  • The official result for that drawing was received.
  • The saved candidate list was compared with that result.
  • The record can show hits and misses without proving why they occurred.

2. Learning evidence

“What setting produced it?”

  • The canonical game ID and exact target drawing are proven.
  • The saved record is tied to the exact experiment assignment.
  • The settings fingerprint matches the settings that actually controlled the run.
  • Only then may the result support or challenge a settings hypothesis.

The SKAI evidence funnel

These 33,760 records were identified by both the SKAI analysis-method key and the SKAI prediction source. The categories below describe learning status—not whether a prediction was “good” or “bad.”

4,272Learning eligible · 12.65%
28,691Excluded from causal settings learning · 84.99%
797Pending classification · 2.36%
Why exclusion is not deletion: an excluded record remains part of the historical scorecard. It simply cannot vote for or against a settings recipe unless its settings lineage is proven.

Where the scored evidence is deepest

This table reports record depth, not winning probability. Games use different rules and candidate-pool sizes, so their results should not be compared as if they were the same experiment.

Canonical gameGame IDScored recordsDistinct drawingsSettings fingerprints
Powerball1016,12192838
Texas Cash 5TX22,604112333
Florida Fantasy 5 EveningFL32,360129350
EuroMillions8012,2413490
Washington Hit 5WA62,208119349
Louisiana Easy 5LA52,16741140
Mega Millions1131,9704468
Texas LottoTX11,9376484
Oregon MegabucksOR11,27097314
California Fantasy 5CA29926221

How this census was built

  1. Inclusion rule: the record had a non-empty canonical game ID, an evaluation status of scored, and a stored main-number hit count.
  2. Unit counted: one saved prediction made for a recorded target drawing and later compared with its official result.
  3. Drawing count: one canonical game ID plus one official actual-or-target drawing date. Multiple predictions for the same drawing remain multiple records but one drawing.
  4. Settings count: distinct non-empty learning-settings fingerprints. A fingerprint identifies a settings recipe; it does not prove that recipe was causal unless the learning contract also passes.
  5. SKAI subset: records whose saved analysis method is skai and whose source is skai_prediction.
  6. Privacy: the published files contain aggregates only. They contain no member names, email addresses, user IDs, or prediction-number payloads.
Snapshot period: the included records run from December 27, 2025 through August 31, 2026. This is a database census, not a randomized trial and not a claim that any model changes lottery odds.

What this report proves—and what it does not

Supported by this census

  • LottoExpert stores a substantial pre-draw prediction history and scores it against recorded official results.
  • The evidence covers many game IDs, drawings, and settings fingerprints.
  • The learning gate is intentionally stricter than the scoring gate.
  • Excluded learning records are retained instead of silently rewritten or discarded.

Not supported by this census

  • It does not prove that SKAI predicts random drawings better than chance.
  • It does not claim that more records automatically produce better predictions.
  • It does not compare hit averages across games with different rules.
  • It does not turn an unproven settings assignment into causal learning evidence.

Responsible-use statement: Lottery drawings remain random and official odds do not change because a prediction system analyzes prior results. LottoExpert provides structured analysis, ranking, tracking, and evidence review—not a guarantee of winning numbers or prizes.

Download the aggregate evidence

The two CSV files let readers inspect the cohort totals and the game-level coverage used in this report. They are intentionally aggregate and privacy-safe.

Download cohort data (CSV-formatted text) Download game coverage (CSV-formatted text) Read how SKAI works

Main Menu