# Accuracy

What each answer stands on, what we measured, and what nobody has measured yet. From our own audit of 2026-11-05.

## Preflop

325 chart sets solved with HoldemResources Calculator, with each game's base rake. Its method samples, so charts carry noise.

- On one 3-handed tree, two runs at the same settings had a different top action on 11.5 % to 12.2 % of decisions. Most of those are close: 7 % to 13 % of them cost over 2 % of the pot, and the worst cost 7.53 %.
- Cash trees have not been measured this way.
- The solver's convergence figure has no measured link to how exploitable a chart is, and its card grouping has no measured cost.
- Three sets have no solve record: the 0.4bb small blind ones.
- A jackpot fee on bigger pots is not in the solves, so lines that build big pots may be slightly too loose.

## Flop

A fast approximation, not a full solve. Measured on 238 stored test spots against a reference solver at about 0.5 % exploitability:

| Measure | Fast flop against reference | Two reference solves |
| --- | --- | --- |
| Top action differs, share of range weight | 23.7 % | 7.5 % |
| Frequency distance | 0.249 | 0.069 |

- The EV gap averages 1.26 % of the pot, and the worst spot is 16.08 %. Our own values price this, so treat it as a size and not a truth.
- Those test spots have at least 90bb behind and a stack-to-pot ratio of at least 4.29. 43.9 % of recent real flop solves sit below that, where we have no measurement.
- The test covers the first decision of a solve. Facing a bet or a raise has no grade.
- There is no rake in flop answers.

## Turn and river

Fresh solves, on ranges carried over from the flop.

- Flop error feeds the turn, and turn error feeds the river. No number covers that chain.
- A known gap: the turn engine is missing two fixes for boards with a suit symmetry. 34.4 % of 1,154 logged turn solves had ranges that this can touch. The size of the error is not measured.
- There is no rake after the flop yet. How much that moves answers is not measured.

## Spins and ICM

- ICM charts use the Malmuth-Harville model. It has no term for future games, and its error is not measured.
- After the flop, spin answers use chip EV, and each one says so.

## What we refuse

- Pots with three or more players after the flop: no answer until our multiway solver is done.
- A game with no close chart: we say so, and it can be requested.

## How to read an answer

Each answer carries `source`, `quality` and `match`, so a model can say how much weight to give it. See [the docs](https://pokersolver.app/docs).

## Study, not live play

For study away from the tables: using a solver or any assistance during a hand is against most poker sites' rules. PokerSolver is independent and not affiliated with any poker site.