PYTOOLZX · WSRF · REPRODUCIBLE PLAY
The training table
Complete hands, changing policies, and results you can inspect.
A fresh 1.6-million-hand examination
The player and control models were frozen before this test. Independent match blocks use fresh seeds and fresh observation priors, with both seats and stack depths balanced. Intervals use variation across blocks.
| Pass | Opponent | Policy | Hands | BB/100 | 95% block interval |
|---|
Full extension receipt · Registered evaluation design
Read the 300-page field guide · 16-page table-side companion · Download the book
| Checkpoint | Hands learned from | TAG BB/100 | LAG BB/100 | Selected then |
|---|
Fresh-deal results
Average net big blinds per 100 hands. Intervals group paired seats on the same deck. Each opponent receives only its own cards and public observations.
| Policy | Opponent | Hands | BB/100 | Approx. 95% interval |
|---|
These are strong local sparring policies built from range equity and betting logic. Their strength has not been ranked against professional players or an equilibrium solver.
Likely and unlikely, separately
The table below groups whole hands by the pocket zone assigned before betting. Dynamic zones can change during play; action and commitment counts for those zones are included in the downloadable report.
| Opponent | Pocket zone | Hands | BB/100 | Total positive BB | Total negative BB |
|---|
What did the zone features add?
The controlled comparison uses the final-round replay data for both fits, with equal total tree capacity. The comparison removes distance features and likely/unlikely forest routing together. The served checkpoint is selected separately by validation return.
How learning works
- Play complete 40 BB and 100 BB hands against Fern and Rook, with exploratory actions.
- Collect public decision features, guide-defined zones, chosen actions and completed returns.
- Refit real WSRF forests on a reservoir from recent rounds. Keep the learned behavior attached to actual outcomes.
- Compare the new checkpoint on validation deals. Use the current champion for the next round, reduce exploration, and give the harder opponent more training exposure.
- Freeze the selected checkpoint before fresh final evaluation. The browser uses that fixed forest and updates its public opponent observations locally.
The equity calculation samples hidden pockets and runouts independently of the actual deal. Guide figures are never inserted as reward labels. A forest’s estimated action value is an estimate; complete-hand returns provide the reported results.
Download the full training receipt · Guide definitions & provenance · Training notes