Inside the Machine: How We Simulated 104 Matches and 351 Goals

The Question Everyone Asks First

"Did the AI just make all of this up?" It's a fair challenge. When you see Kylian Mbappé score a 112th-minute winner in the Final (match 2026-104, France 3–2 Argentina AET) to hand Les Bleus a second World Cup title, the sceptic in you wants to know whether that moment was plucked from thin air or whether something more rigorous is underneath it. The honest answer is: both instincts are partially right, and understanding where one ends and the other begins is what this piece is about.

Two Engines, One Tournament

Our simulation runs on a hybrid architecture. The first layer is a large language model — the "AI judgement" layer — that encodes a vast, implicit understanding of international football: squad depth, tactical tendencies, recent form, historical head-to-heads, player age curves, and the psychological weight of tournament pressure. This layer produces a probability distribution over possible outcomes for each match. It does not pick a score; it assigns likelihoods across a range of scorelines, expected goal-scorers weighted by role and current-season output, and even rough minute-of-goal distributions informed by patterns across thousands of real matches.

The second layer is deterministic sampling code. Given the probability distributions the AI layer outputs, a reproducible random seed is used to draw a single realised outcome. Think of it as rolling a weighted die: the AI designs the weighting, the code rolls it. This is where "AI judgement ends and deterministic code begins." The code also enforces hard constraints — scorers must be in the named squad, own goals are flagged separately, extra-time and penalty-shootout logic follows FIFA rules exactly. The result is a simulation that is statistically coherent rather than merely narratively convenient.

What the Numbers Tell Us

Across 104 matches, the engine produced 351 goals — an average of 3.38 per game, which sits comfortably above the 2022 Qatar average of 2.69 but is consistent with the expanded 48-team format's expectation of more mismatched group fixtures. The distribution is telling: Germany's 5–0 demolition of Curaçao (2026-010) and Spain's back-to-back 4–0 group wins (2026-014 and 2026-038) anchor the high end, while the knockout rounds tighten dramatically — five of the last eight matches required extra time, including the semifinal between Brazil and Argentina (2026-102, 2–3 AET) and the France–Germany last-16 thriller (2026-089, 2–2 AET, pens 4–5). The AI's probability layer had correctly flagged both those ties as "too close to call," and the sampling code delivered accordingly.

The simulation logged just two upsets across the entire tournament — a deliberate calibration choice reflecting that, at a World Cup, the better-resourced, higher-ranked squads do tend to progress. Morocco's penalty shootout elimination of the Netherlands in the Round of 32 (2026-075, 2–2 AET, pens 3–4) was one of them, a result the AI layer assigned roughly a 22 % probability — unlikely, but far from impossible given the Atlas Lions' defensive resilience and Dutch penalty-taking history. The other upset is left for readers to identify in the full bracket.

Where Human Judgement Still Lives

The AI layer is not a black box we trust blindly. Our editorial team reviews every output against a sanity checklist: Does the top scorer make sense given squad role and minutes? (Mbappé finishing on 13 goals as a central striker in a dominant France side: yes.) Are there any scorers who were injured in real-world pre-tournament build-up? (Flagged and replaced before the seed is locked.) Does the goal-minute distribution show implausible clustering? The checklist caught two early drafts where a single player scored in the 1st and 3rd minute of the same match — a statistical artefact the code now filters out. What the checklist cannot do is override the sampled result; once the seed is locked, the outcomes are canon. That discipline is what makes the simulation a simulation rather than a story someone wrote to please a preferred winner.

The Limits We Acknowledge

No model predicts football perfectly — if it did, bookmakers would be out of business. Our 351-goal total and France's championship are one coherent draw from a distribution that also contained timelines where Argentina retained their title, where Erling Haaland's Norway went deeper than the Round of 32, and where Jonathan David's eight-goal group-stage rampage carried Canada further than the last 16. We publish one timeline, clearly labelled as speculative, because a single vivid narrative is more useful — and more honest — than a probability table that hedges every claim into meaninglessness. The machine made its choices. Now we watch the real tournament to see how well it listened.