Poker Assistant

Accuracy

How accurate the solutions are

A solver result is only as good as the measurement behind it. This page says what we measure, in units you play in, and what we do not claim.

The number

After every preflop solve, the engine measures how much a perfect opponent could still win against the result, seat by seat. The engine records it in milli-big-blinds per game, the unit the research literature uses. One milli-big-blind per game is a tenth of a big blind per hundred hands. The shipped packs measure in single digits, which is under one big blind per hundred hands for any seat. A solid winning regular runs at three to five big blinds per hundred. We publish that measurement on its own. An error bar would take longer to compute than the solve did, and it would have to be redone for every pack.

Why the ladder starts at 15bb

Below 12bb a shove-or-fold model is close to right, so the push/fold cells stop there. We tested leaving 15bb and 30bb on that model too. A perfect opponent gained about 190 big blinds per hundred hands against the 15bb push/fold strategy and about 340 at 30bb. On the full raise ladder the same measurement is under one. That gap is why the ladder starts at 15bb and the push/fold pack stops at 12bb.

How it solves

The solver runs counterfactual regret minimisation with discounting (Brown and Sandholm, 2019; Farina, 2021), external-sampling Monte Carlo CFR (Lanctot, 2009), exhaustive board enumeration with variance reduction (Burch, 2018), and a lock-free parallel solve across cores. Each is a published, cited method. The measurement above is how we check that our implementation of them converged.

What it does not pretend

The builder shows a confidence band while a preflop solve runs, so you see how settled the result is at each checkpoint. A clean falling line would be lying about the precision. From any preflop line you can also deal a board and solve that exact spot through the turn and the river; that resolve reports its progress, which is not a convergence measurement. Tournament equity is not in the product.

Solving

Screenshot of a postflop resolve in progress, reporting a running iteration count rather than a convergence measurement.

Solved

Screenshot of the same flop resolved, with a full action ladder and range grid in place of the progress readout.

Re-solved

Screenshot of a second flop re-solved on the same line, confirming the resolve runs independently each time.
The same postflop line resolving on two different boards, each run reporting its own progress.

Where your data is

Poker Assistant stores everything on your device and sends nothing. That is a commitment we make in the privacy statement, and we intend to back it with an independent review or a public data-flow description. Until then, treat it as our word, and hold us to it.