WC2026

About this site

AI calls.
No wriggle room.

This is a running log of World Cup 2026 predictions made by Claude AI, tracked honestly against actual results as they come in. Every call is locked in before kick-off. Nothing gets rewritten after the fact. The accuracy numbers are computed automatically from the raw data, not hand-picked.

The idea is to put an AI prediction model under real accountability: make a call, log the reasoning, see if it was right, and build up an honest picture of where the model is sharp and where it is not. So far the meta-pattern is clear: picking the winner is fairly reliable; pinning the exact scoreline is harder.

How the scoring works

EXACT Predicted the correct scoreline: the right result and the right goals.
CORRECT Got the right outcome (win/draw/loss) but the scoreline was off.
MISS Wrong outcome. The result went the other way.
PENDING Match hasn't been played yet. Prediction is locked in, result is not.

Scorer calls are tracked separately. HIT if the named player scored, ASSIST if they were directly involved but did not score, MISS if they had no involvement. Some calls are flagged as "goal or assist". Those count as a HIT on either.

Prediction types

Mismatch Clear quality gap. A multi-goal margin is on the table.
Favourite Confident favourite call. Expecting a comfortable win, but not a mismatch.
Fav (tight) One side is better but the game will be close. Backing a narrow win.
Draw call Genuinely even game. The draw is the highest-probability single outcome.
Rule bend Going against the method slightly. Noted explicitly and explained.

The role of AI

The predictions are generated by Claude AI (claude.ai) running Claude Opus at Max Effort, Anthropic's highest reasoning mode, which gives the model extended thinking time before committing to a call. The engine originally ran on Claude Fable until Anthropic retired the model; Opus with Max Effort has been the engine since. Given the fixture, group context, team form, and an established prediction method, the model produces the scoreline call, the scorer pick, and the reasoning. All before kick-off.

The site itself is built and maintained by Claude Code. The daily workflow is: verified result comes in, it gets described in plain English, Claude Code finds the right match in the data file, fills in the actual score and scorers, then rebuilds and deploys automatically via Netlify.

The AI also cross-checks results across multiple sources (ESPN, FOX Sports, NBC Sports) before logging anything, and will not update a match until the full-time whistle is confirmed. That matters because fast-moving live feeds occasionally get things wrong in the first few minutes after a goal or at the final whistle.

Built with Astro, Tailwind CSS v4, and TypeScript. Deployed on Netlify. All verdicts derived from a single TypeScript data file. Nothing hardcoded.

Back to predictions