How accurate is the solver?
Every board in the library is attacked by the strongest opponent we can train, at every seat. What that opponent can win back is the accuracy number on each board. This page shows those numbers live, by game and table size, and explains exactly what they mean and how we get them.
The numbers, by game and table size
Pick a game. Each row covers every library board of that table size. Lower is better: 1% means the strongest exploiter we can train wins back 1% of the starting pot per hand, every seat added up.
| Players | Boards | Certified | Typical board | Best to worst | Bar | At the goal |
|---|---|---|---|---|---|---|
| measuring… (live from the app) | ||||||
Typical board is the median: half the boards of that size measure better, half worse. Best to worst is the full range. The bar is the line a board must be under to be published as certified; the goal is where the solver keeps pushing every board after that (1% heads-up and 3-handed, 1.5% at 4 to 5 players, 2% six-handed and up). The table covers 2 to 6 players, where the library lives; full-ring tables (7 to 9 players) are solved on demand on the Elite plan, and each of those solves shows its own number the same way.
What the percentage means
A solved strategy is only as good as it is hard to beat. So after a board is solved, we build an opponent for every seat whose only job is to beat it. That opponent knows the strategy exactly and adapts to it perfectly. Then we measure how much more it wins than the strategy itself would win from the same seat.
Add that up across every seat and divide by the starting pot, and you have the percentage. It answers one question: how much could a perfect, all-knowing opponent take from this strategy?
It is not your win rate, not your equity and not how often the solver is "right". Real opponents are nowhere near a perfect exploiter, so in practice the gap is much smaller than the number. Zero would mean unbeatable; small means very hard to beat.
How we measure it
- Solve the board. The solver plays itself billions of hands on that exact pair of flops, every seat, both boards to the river, until the strategy stops improving.
- Train an exploiter for every seat. On the GPU, a best response is trained against the finished strategy: it sees the strategy exactly and learns the most profitable way to play against it. Its training budget grows with the size of the game tree, so a six-handed board gets far more than a heads-up one.
- Measure what it wins. Each exploiter then plays hundreds of thousands to millions more hands to measure its edge, and the edges of every seat are added up.
- Repeat five times. Five independent passes, each with its own random seed, give the number and its standard error, so a lucky pass cannot make a board look better than it is.
- Publish only what passes. A board goes into the library only once its number is under the bar for its table size. The number stays on the board, so you can see how far to trust it before you study it.
The ruler: on September 5, 2026 we raised the exploiter's training budget to 16 times what it was, and every board was measured again. The check converges there (doubling the budget again moves the readings by a few percent). Numbers went up that day because the measuring got stricter, not because the solves got worse.
Certified, bar and goal
Certified
A board is certified when its measured percentage is under the bar for its table size. Every library board you can open is certified, and its own number is shown on it in the app.
The bar
Bigger tables mean bigger pots and much bigger game trees, so the bar is set per table size, at the level the certified library reaches under the current check. It only moves down over time.
The goal
Whenever the solver has spare time it keeps solving library boards deeper, toward 1% heads-up and 3-handed, 1.5% at 4 to 5 players and 2% six-handed and up. The "at the goal" column shows how many boards are already there.
What it does not tell you
It measures the game we solve. Each library board is one flop pair, at one stack depth, ante and set of bet sizes, with a raise cap. Within that game the number is a direct measurement, with its standard error; a spot the game does not include (a bet size it does not offer, for example) is not covered by it.
Omaha hands are grouped by strength. In PLO, PLO5 and PLO6 there are far too many hands to store a strategy for each one on every runout, so hands are grouped by how strong they are on each board, and hands in a group play the same way. The exploiters play the same grouped game.
Multiway has no single right answer. With three or more players there can be many strategies that are all hard to beat. The solver finds one of them, a low-regret strategy, which is the practical standard for multiway games; the number tells you how hard that strategy is to beat, not that it is the only one.
More questions: see the FAQ or contact support.
Study the boards behind the numbers
Twelve sample solves open fully on the free plan, each with its own accuracy number. The Library plan opens every certified board.