Benchmarks & research

Technical limitations

• Blueprint postflop strategies use an abstraction. The "!" badge marks places with limited hand-level detail. Repair or re-solve them.
• Blueprint postflop EVs are approximate, not per-combo values, and can rank actions differently from the strategy. Trust the frequencies; re-solve for per-combo EVs. Older blueprints have less reliable postflop EVs.
• The database is only partly re-solved. Nodes without a saved re-solve show the blueprint. A re-solve improves one node; it does not make the whole game a preflop-to-river equilibrium.
• Every solve is an approximate equilibrium of its own betting tree, reached after a finite number of iterations. No engine considers bet sizes outside its tree.
• Re-solves cover 2 and 3 players. Pots with four or more players keep the blueprint strategy.
• Three-way flops cover all remaining cards by default when RAM permits. If RAM is insufficient or quick preview is selected, fewer turns are solved. The reported accuracy does not include the effect of omitted turns.
• Three-way accuracy is the average single-player gain. It says nothing about two players deviating together.
• Saving a re-solve stores only the node on screen, not the whole subtree below it.
• The SB is the only seat that may limp in the preflop engine.
• No ICM: payoffs are chip EV with rake.
• Windows with AVX2 only; there is no macOS build.
• RAM limits everything. Training stops at 85% of physical RAM, and re-solves do not start if they need more than 90% of free RAM.