Turnstone

An answer comes from one of three places, and the grade names which.

1. Exact solutions it already holds, for 575 flops. They were solved ahead of time until neither player could gain more than 0.75% of the pot by changing strategy, and a grade read from one is marked exact.
2. Exact turn and river solves, done while you play, for the board, pot, stacks and bet in front of you, at the trainer's own bet sizes.
3. A flop search, done while you play, for any other flop in its 13 spots. It takes about 2 to 3 seconds once the engine is running, and the grade is marked engine estimate.

Solving a flop exactly means playing out every turn card and every river card after it. Here that takes from half a minute to over two minutes on a laptop, too long to wait for at a table. Turnstone searches the flop's betting only. At the moment the turn card would be dealt, it asks a neural network what every hand is worth from there, for both players, instead of playing out every turn and river. The network learned this from 100,000 turn positions, each one solved exactly. The search runs 30 rounds over the flop, asking the network at each turn card, and returns a strategy for the flop.

The name comes from that step. A turnstone is a small shorebird that flips stones over to find what is hidden underneath, and poker is a game of hidden cards. This engine does its work at the turn. Chess engines often have nature names, such as Stockfish and Leela, and Turnstone follows that habit.

It builds on published ideas and is not affiliated with the people behind them. The self-play that finds a strategy is counterfactual regret minimization (CFR), the method behind the University of Alberta and Carnegie Mellon poker programs such as DeepStack, Libratus and Pluribus. Searching only a few streets ahead and asking a trained network for the value of the rest is DeepStack's idea, depth-limited search with a value network. AlphaZero-style chess and Go engines have the same shape: a search, with a network to judge the positions where it stops. Turnstone solves turns and rivers exactly with its own solver, using discounted CFR, a variant from Carnegie Mellon. The flop solutions the app ships were made with postflop-solver, an open-source Rust solver by b-inary, and are being re-made with Turnstone's solver.

On 2026-10-13 Turnstone was run on all 575 flops the app has exact solutions for, and its strategy at the first decision was compared with the exact one. The loss is how much of the pot Turnstone's strategy gives up against the exact strategy. In a pot of 6 big blinds, 0.3% is about a fiftieth of a big blind.

• It knows one flop bet size a street, the trainer's own, and answers no other. On the turn and river it solves the size that was bet.
• It covers 13 spots, all heads-up pots after a raise and a call or a 3-bet and a call. Pots with three or more players, limped pots, 4-bet pots and all-ins before the flop are graded by a rule of thumb, and the grade says so. The trainer ships solved flops for a small blind limp against the big blind, but Turnstone's loss there has not been measured, so that spot is not one of the 13.
• Its flop answers are estimates: a search with a measured error. The exact answers are the shipped flops, the turn and the river.
• It has not been measured against any other poker tool.

At the cash table and the tournament table, a heads-up decision after the flop with no exact solution goes to Turnstone for its grade. On the analysis board you set up a spot, a board and the action so far, and ask Turnstone about it. The solver flops drill grades against the exact solutions. All of these are invite-only for now.