World Cup 2026 Champion Simulations: An Audit of Retained Runs
Publisher: Football Sim Lab · Published · Corrected
Correction, 18 October 2026: The earlier June-labelled estimates could not be attributed to a retained input snapshot. They have been withdrawn. This article now reports counts from named stored output files, not verified forecasts or actual tournament results.
Champion counts differ across batches
| Retained file | Most frequent champion | Wins / 1,000 | Frequency |
|---|---|---|---|
| simulation-results.json | Argentina | 355 | 35.5% |
| simulation-results-v2.json | Spain | 159 | 15.9% |
| simulation-results-v3.json | France | 158 | 15.8% |
These are counts of the stored champion field, not a new forecast. The leading team changes between files, which makes their assumptions essential. The filenames alone cannot establish that one batch reflects May and another June.
Why the earlier prose was corrected
The earlier article said quarter-final pairings rarely exceeded 12%, then listed 43.5%. It described an 11.7% semi-final frequency alongside a 30% entry, and approximately 10% for a final pairing alongside 20.6%. These statements could not be reconciled with a documented single batch.
Those percentages and their interpretation have been removed. The table above makes a narrower claim that can be checked against the retained files. A team appearing most often does not establish calibration, and a large change between batches does not demonstrate a change in the team's real strength.
Use the archive to ask better questions
A useful comparison records the team list, draw, manual scores, real-results setting and model version. Changing any of them can change the winning frequencies. The current simulator can help explore those scenarios, but cannot retroactively recover the missing inputs of old runs.
Evidence and limits
The research audit and downloadable data contain source filenames, SHA-256 checksums, run counts and aggregates. The repository script scripts/generate-research-summary.mjs recomputes the counts. All three batches contain 1,000 outputs.
Original seeds, complete rating inputs, locked results and model commits were not retained alongside these files. Filenames cannot certify a generation date. These outputs cannot establish forecast accuracy or a calibrated real-world probability. Some stored round labels are wrong, so the audit uses match IDs to identify stages. Batches with unknown assumptions must not be pooled.
Read Methodology for the current game engine and How to play for a controlled example. Official tournament information is available from FIFA. These are independent game outputs, not FIFA predictions.