The training flywheel
The bot now retrains itself: it measures which fights it plays worse than humans, collects counterfactual evidence (playing both sides of its hardest decisions to the end of the fight), fine-tunes that encounter's model, and only ships a change that passes every safety gate. This page watches the loop live.
Encounter campaign
| Encounter | Status | Training examples | Latest result |
|---|---|---|---|
| loading… | |||
How a round works
1 · Triage. Replay real human fights from identical starting decks and compare damage taken and deaths. Only encounters the bot plays measurably worse than people earn a round — healthy models are left alone.
2 · Counterfactual collection. Fully-C++ searchers walk hundreds of those fights. Wherever the model admits it can't separate two moves — or shows a known bias (under-blocking a big hit, wasting block on a safe turn, ending a turn with energy unspent, mis-ordering a combo) — both moves are played out to the end of the fight under identical randomness. The difference in outcome is a measured, causal training label.
3 · Self-calibrating fine-tune. The encounter's model is nudged from its deployed weights with three balanced objectives: match the measured outcomes, order the contested moves correctly, and stay anchored to everything it already does well. Its death-probability calibration is refit and sealed into the model file every round.
4 · The gate ladder. A candidate must hold every previously-passing tactical case, then prove itself on a paired battery walk against the deployed model on identical fights. Pass → staged for human review. Fail → at minimum its improved danger calibration is salvaged and installed, with zero change to fight behavior.