The 2026 World Cup was Momus's first tournament-scale test: a call on every single match. The read held up — a strong win rate across the tournament, at a positive ROI. But the more useful lesson wasn't the record. It was what the record hid.
Action mode: activity vs discipline
For the tournament we deliberately tuned the workflow to push Momus harder toward a bet than it normally would — partly as a value test, and partly because a call on every match makes for better coverage. That keeps the win rate high, but it forces stakes onto short-priced favourites where there's little value left. The winning calls carried the run; the marginal ones mostly gave it back. The takeaway: discipline over activity.
Calibration: does 70% really win 70%?
We hold Momus to more than a win rate. Calibration asks whether a 70% call actually wins about 70% of the time. It's what separates a genuine edge from a lucky streak, and almost no tipster measures it. The tournament gave us our first big sample to check against — and it showed our confidence had run a touch hot under the extra push.
What changed
So we built a calibration layer that continuously compares predicted probabilities to actual outcomes and corrects the model's confidence over time. And post-tournament, Momus went back to value-first: same read, but it only bets when Polymarket's price is genuinely mispriced versus its fair line, and passes the rest.
Call, grade, calibrate, sharpen — run in the open. That loop is the actual product. The full 2026 World Cup record, match by match, is on the track record.