Original data analysis
What our retained simulations actually show
Audited · Publisher: Football Sim Lab
We recounted the outputs in three stored batches rather than generating a new forecast. Each file has 1,000 runs. The results differ substantially, which makes the inputs and any fixed results essential context for interpreting percentages.
Champions counted from each batch
| Batch file | Most frequent champion | Count / runs | Frequency |
|---|---|---|---|
| simulation-results.json | Argentina | 355 / 1000 | 35.5% |
| simulation-results-v2.json | Spain | 159 / 1000 | 15.9% |
| simulation-results-v3.json | France | 158 / 1000 | 15.8% |
A fixed-looking pairing is a warning to check the inputs
In simulation-results.json, the M97 quarter-final is France vs Morocco in all 1,000 outputs. That frequency describes the retained batch. It does not establish that the match was certain before the tournament, or that the model predicted it perfectly. Locked inputs, a partially completed bracket or other assumptions can produce such a pattern. The batch does not retain enough input metadata to distinguish those explanations.
How these numbers are produced
The audit counts each stored champion once per run. For a pairing, it selects the match ID, sorts the two team names so their order does not matter, and counts the pair. A displayed frequency is count divided by the file's run count. It does not average incompatible batches. Some stored round labels incorrectly call later matches “round-of-16”; this audit uses match IDs 73–88 for the round of 32 and 97–100 for the quarter-finals.
The repository script scripts/generate-research-summary.mjs recomputes these aggregates from the three files and records each source's SHA-256 checksum. The downloadable audit data includes all champion counts, match-pair counts, host appearances and source checksums.
What cannot be recovered
Retained output batches, not current forecasts. Original seed, complete input snapshot, model commit and locked-result settings were not retained alongside these files. Filenames are identifiers, not certified generation dates.
Consequently these batches do not support a retrospective accuracy score, a claim that the current engine is calibrated, or the previously published “June 2026” estimates. Those historical articles now explain the evidence and correction rather than presenting untraceable percentages as verified forecasts.
See the corrected archive analyses of knockout pairings, host nations and champion frequencies. For the engine's current rules, read Methodology.