PulseGTO
Blueprints & databases
Accuracy & limitations
A blueprint can differ from the true equilibrium for three separate reasons. Its accuracy number measures only the first.
- Convergence. Training runs a finite number of iterations, with sampled opponent actions. The regret bound shows how far this has progressed.
- Abstraction. Hands and boards share strategies. A best response computed inside the same abstraction cannot see this error, so a low number means "converged", not "faithful to the full game".
- The tree. The blueprint solves the game you configured: its bet sizes, raise levels and rake. Sizes that are not in the tree are not considered.
Displayed postflop EVs are a separate matter: they are estimates for groups of hands, not per-combo values. See Blueprint postflop EVs are approximate for how to read them.
How to judge a blueprint
- Trust the frequencies more than the postflop EVs. For per-combo EVs, re-solve the node.
- Judge by EV loss, not by whether frequencies still move. Some nodes have several nearly equal mixes, so their frequencies can keep moving after their EV loss has settled.
- Check the tree and the rake before solving longer, and the regret bound before trusting a frequency.
- Re-solve the postflop nodes you care about. A RangeCFR re-solve removes the hand-abstraction error on that node and reports its exploitability for the configured betting tree. You choose when to stop or which accuracy target to pursue. Heads-up covers every turn and river; three-way also covers every turn and river by default when the full tree fits in RAM. A three-way flop uses reduced turn coverage only for quick preview or when RAM requires it.
The current published blueprint is HU · 20bb · cEV (no rake), trained for 50,000 samples. Its preflop game value was compared with an independent reference; postflop play and EVs remain limited by hand abstraction. See the download page.
Diagnose the question before extending a run
| Observed issue | Useful check | What more iterations alone cannot fix |
|---|---|---|
| Mix moves between checkpoints | Compare average, current and consistent action values | A missing action in the game tree |
| One postflop hand looks implausible | Re-solve the node per combo with its actual reaching ranges | Hand-level restrictions of the abstraction |
| Two tools disagree | Match trees, ranges, rake and EV baselines first | Different starting games |
| Three-way flop looks accurate only on some turns | Inspect coverage and solve an omitted turn separately | Unsolved cards |
A low number is useful evidence about its defined metric, not permission to ignore the model. Choose the smallest spot that answers your question, preserve the inputs and report the result with its limits.