Five Goals, One Stalemate, and a Very Awkward Night for Grok

Four matches on Tuesday spanning Houston, Foxborough, Toronto and Zapopan. Three of them the models handled reasonably well. The fourth — England 0–0 Ghana — was a collective face-plant that tells us something worth filing away about how AI models think about heavy favourites.

Portugal 5–0 Uzbekistan (Group K, Houston)

The right call, emphatically confirmed. Portugal were overwhelming favourites and delivered accordingly, winning by five. All models predicted the correct outcome; the interesting question is just *how* confident each was — and whether that confidence was appropriately rewarded by the scoreline.

Grok wins this one clearly. Its 86% confidence on Portugal — the highest of any model — translated into the best Brier score of 0.031. This is exactly where Grok's betting-market anchoring earns its keep: when a mismatch is as obvious as the market implies, leaning hard on those odds is efficient. GPT-5.4 and Claude came in most cautious at 78% each, both crediting Uzbekistan with more threat than the 5–0 scoreline suggested they deserved. Claude's known tendency to over-price home advantage didn't inflate its Portugal figure here — if anything it was under-confident relative to Grok, possibly because the away-venue effect (NRG Stadium is technically neutral territory) muted that instinct.

All five models picked 2–0 for the scoreline. The actual result was 5–0, so every exact-score attempt missed — but every outcome call landed, and the direction of the miss (underestimating Portugal's margin) is consistent with the well-documented pattern of models underrating goal-differential in mismatches. Not a failing that costs points today; worth remembering when tournament progression depends on goal difference.

England 0–0 Ghana (Group L, Foxborough)

Here is your finding of the day. England were favourites across the board — heavily so in Grok's case — and Ghana held them to a goalless draw. Every model predicted England to win; every scoreline pick was 2–0; nobody got within reasonable distance of the actual result.

Gemini was least wrong — its 22% draw probability was the highest on the board, and its Brier score of 1.08 reflects that marginal hedge. GPT-5.4 was close behind. At the other end, Grok assigned only 13% to a draw, pushed England to 81%, and paid the steepest price with a Brier score of 1.42. This is Grok's betting-market fingerprint working against it: when the market is confidently wrong, following it amplifies the error. England's implied probability in pre-tournament odds would have been generous; Grok absorbed that signal and ran with it.

Claude sat at 72% for England with only 17% on the draw — its home-advantage inflation is visible here (though England were technically away from home in any meaningful cultural sense, playing in Massachusetts). The ensemble, averaging across the bullish models, ended up the second-worst performer. Aggregation dampens individual extremes but cannot rescue a collective blind spot.

This is a useful data point for the broader 'reputation versus form' bias we track. England's reputation demands respect; Ghana's recent form and organisational discipline were, evidently, sufficient to deny them. No model gave that scenario more than a 22% chance. That's not a catastrophic miss in probability terms — draws *are* the low-probability outcome in most asymmetric fixtures — but the collective failure to see it coming at all is a pattern worth watching as the group stage progresses.

Panama 0–1 Croatia (Group L, Toronto)

The models read this one well. Croatia were clear favourites, Panama's hosts-by-geography advantage counted for little, and the correct outcome was predicted by all. Note that probability data for this match is partially absent from the raw predictions block, though accuracy scores confirm the breakdown: Grok and Gemini were at 60% and 63% for Croatia respectively; GPT-5.4 and Claude went further at 65% and 67%.

Claude takes top honours here, and it is mildly interesting that its home-advantage bias did *not* inflate Panama's chances — when the home side is a clear underdog even against relatively modest opposition, the structural disadvantage appears to override the model's usual tilt. Croatia's experience and pedigree (league prestige that Gemini should technically over-weight) was sufficient for Gemini to still call this correctly, even if it was the least confident of the group. Grok was most cautious about Croatia at 60%, a rare moment where its market anchoring produced a more conservative figure than the other models. All five scoreline picks were 0–2; Croatia won 0–1. Outcome right, margin overstated — par for the course when models pick a strong side to win comfortably.

Colombia 1–0 DR Congo (Group K, Zapopan)

A genuinely competitive fixture on paper, and the models' spread of confidence reflected that. Colombia at 54–67% for the win is not a rousing consensus — this was the most uncertain prediction of the day — and the 1–0 result validated the cautious majority view while rewarding those who leaned Colombia most confidently.

Grok wins again — 67% for Colombia, best Brier score of 0.172. Its bullishness on the stronger-looking side (Colombia carry more FIFA ranking weight; the betting markets would have reflected their South American pedigree) delivered here. GPT-5.4 was least convinced at 54%, giving DR Congo 21% — the most generous assessment of African opposition across today's slate. Claude was similarly cautious at 55%, consistent with its general tendency not to over-commit on matches where home advantage is real but not overwhelming.

The scoreline story is worth a sentence: all five models and the ensemble picked 1–0 for Colombia, and all five were exact hits. That is genuinely unusual — exact scores are a hard target and four correct calls in a day would be remarkable; five unanimous ones landing simultaneously is the kind of outcome that makes you want to check the data twice. It checks out. Credit where it is due, though the caveats apply: this is one match, and 1–0 is a common result in tight encounters that models tend to default to when they expect the stronger side to edge it narrowly.

Day in Summary

• Grok: Best day overall — nailed the Portugal and Colombia confidence levels, took the heaviest punishment on England. Net: solid when markets are right, exposed when they are wrong.
• Gemini: Quietly consistent — highest draw probability on England (22%), best-performing model in that match. League-prestige bias didn't distort Croatia or Portugal calls today.
• Claude: Competitive on Croatia (top scorer there), underconfident on Portugal and Colombia, slightly inflated home-advantage reads balanced out. Home-advantage fingerprint visible on England where it probably wasn't warranted.
• GPT-5.4: Steady mid-table performer all day; most sceptical of Colombia and least aggressive on Portugal. No disasters, no standout wins.
• Ensemble: Absorbs the collective bias on England and compounds it — the aggregation benefit disappears when all inputs share the same blind spot. Best used when models genuinely disagree.

Looking Ahead to Wednesday

Tomorrow's slate will be worth watching for several reasons. If any Group K or L matches involve sides with strong league reputations (Gemini alert) facing teams with genuine recent defensive form (England–Ghana style), the draw-underpricing bias could resurface. Any match featuring a traditional powerhouse in a neutral or quasi-away context will test whether Claude's home-advantage inflation recalibrates appropriately. And if Grok continues to anchor on market odds, any match where the betting consensus looks shaky — overrating a fading name, perhaps — could expose it again the way Ghana did today. We will be watching.