Signals still being tested, shown honestly
Not every signal the model computes is already deployed. This page collects the ones still in testing — including the ones that turn out mediocre or inconclusive — with their real, unfiltered numbers, not just the ones that look good.
Compare "Draft hero winrate" against the model's own call
Every live signal includes a line labeled Draft hero winrate — the real, per-player hero-pool win rate for each team's five picks. This is a separate number from the model's own internal prediction, not the same figure restated. Treat it as a second opinion: when it points the same direction as the model's call for the map, that's reinforcing evidence. When it disagrees, that's exactly when to slow down and watch the game more closely before deciding.
This is an experimental signal, still in testing. The numbers above come from a single, one-time check run on 2026-10-25 against every 19-minute pick settled so far (3,021 total; 3,004 of them had usable Draft hero winrate data at the moment of the check) — not a live, forward-tracked record. The underlying feature is still being trained and validated on the local statistical model, and it isn't used to filter or skip any bet automatically today.
This block, unlike the one-time check above, is genuinely live: it only counts picks placed from 2026-10-25 onward and never reuses today's hero-pool data on older bets, so it carries no look-ahead risk. Expect it to show little or nothing for the first few days — that's expected, not a bug — and to grow on its own as more 19-minute maps settle.
See the deployed model's own track record
Experiments live here until they're proven — the live signals bot only ever uses what's already validated.
See the live track record →