ARCHITECT WORLDCUP

I wanted a World Cup forecast I could actually defend, so I built the verification first and let the forecast follow. Architect WorldCup re-simulates the 2026 tournament from the full history of international football: it rates every team, fits a goal model, runs the real bracket thousands of times, and reports each team's odds of reaching every round. Every number it prints is backed by a test that runs, and the experiments that failed stay in the record beside the ones that worked.

I did not hand-write the model. I directed it. I ran a two-agent method by hand: a strategy chat to reason through each problem and Claude Code to execute against the repository, with me carrying decisions between them, reviewing what came back, and killing proposed fixes by reading the code they claimed to repair. Directing that system and catching its errors was the work. The written protocol, the session log, and a co-author trailer on every commit are all in the repo, so the method is on the record as plainly as the results.

Data leakage fails the build here. If a model can see a match it is trying to predict, the test suite goes red and the run stops. Out of sample, across eight non-overlapping backtest windows, the goal model scores a mean Ranked Probability Score of 0.1575 against a base-rate baseline of 0.2292, and it beats that baseline in all eight. Two ideas I expected to help, downweighting friendly matches and a gradient-boosting ensemble, never cleared the noise floor, so they stayed in the repository as documented, rejected experiments.

The last layer is the one closest to what I build. It puts the math model and a language model against each other on the same knockout tie, scored the same way, one tie at a time, each prediction committed before kickoff. The language model is confined to a frozen, timestamped dossier, given no tools when it predicts, and refused if it states anything it cannot cite. A reconciler reads both forecasts and issues a final call it has to justify against the evidence. That is the Architect principle running on a single question: a layer that reasons, a layer that executes, and a check at every handoff between them.

This is one domain instance, built on the same verification principle the Architect AI is built on. It is public, Apache-licensed, and maintained, and the repository is the full and current record.

The repository is the full record: the seven-layer pipeline, the tests that enforce every claim, the two rejected experiments, and the forward-only math-versus-language-model comparison, committed tie by tie before each kickoff.