How we check the answers

Which hands we read

No-Limit Hold’em cash games from GG and PokerStars, and hands you describe in your own words. Tournaments, play money and other games are skipped, and the upload says how many.

The advice is built for 6-max. At bigger tables the early seats are read as UTG; heads-up tables get no solver answer.

Preflop: charts

Preflop answers come from fixed 6-max charts at 40bb and 100bb deep: opens, calls, 3-bets, 4-bets, squeezes and play against limps. The charts are our hand-built estimates of standard play, not a solver’s output.

Under 20bb, all-in spots use Nash push/fold charts, with and without antes. A stack between depths uses the nearest chart and says so; a raise more than twice the chart’s size, or a stack over 250bb, gets no answer.

At 100bb a few of your own preflop answers are tightened for how the player pool really plays. Those answers carry a note saying so.

Flop: a library solved in advance

Flops come from a library we solved ahead of time: 37 heads-up lines (pairs of seats in single-raised, 3-bet and 4-bet pots, at 100bb and 40bb), each across all 1,755 strategically different flops, solved to within 0.1–1% of the pot.

The trees bet 33% or 75% of the pot (some lines 75% only) and raise 60%. A pot without its own line borrows the nearest one and is marked approximate.

Turn and river: solved live

Turns and rivers are solved when they matter: for the 50 biggest pots of each upload, and for any hand you open or ask about. A solve stops at 0.3% of the pot on the turn and 0.1% on the river, or after 8 seconds, whichever comes first. The river adds a 1.5× overbet.

Where the solver stops

The solver plays heads-up. A multiway pot is solved against the one opponent who matters, with a note that says so. Under 20bb after the flop, the tree is only all-in or check.

Limped pots, limp-reraised pots, cold 4-bet pots, 5-bet pots and some squeezed pots get no solver answer.

How a decision is graded

We look at how often the strategy takes the action you took. 25% or more is good, 5% to 25% is questionable, under 5% is a leak. A bet or raise is matched to the nearest size in the tree, with a note when yours was more than 10% off.

When the engine had to answer a spot too far from yours (a multiway pot read as heads-up, or a stack half again deeper or shallower than the one solved), the decision is graded rough: you still see the numbers, but it counts as neither good nor a leak and stays out of the leak report.

Grades judge the decision, not the result: a good call that lost is still good.

The leak report ranks spots by a leak score: how far each decision was from the strategy, weighted by the pot, per 100 hands. It is not money won or lost; your actual result in the spot is shown next to it.

Tex, where the solver has no answer

Spots the engine can’t answer go to Tex, an AI model that reads the spot and estimates a balanced strategy. Before his reads were switched on, he was tested on 100 spots the solver had answered: he had to give the same grade at least 70% of the time, call a good play a leak (or the other way round) no more than 5% of the time, and repeat his own grade when asked twice at least 85% of the time.

His reads are graded the same way and count in your leaks, and they are always labelled “Tex's read · experience, not solver”.

Keeping the coaches honest

The coaches are AI models: they explain, the engine supplies the numbers. Every percentage and bb figure in a reply is checked against what the tools returned and what you wrote. A figure with no source gets a dotted red underline (“No tool gave this figure”), and a reply with no engine answer behind it is stamped “Not solver-backed”.

Tex, the Kid and the Duchess speak from experience: their replies are marked “Read”, and their figures are always shown as unverified.

The check proves a number came from the engine, not that the coach used it well. For that, open the hand and see the engine’s answer for yourself.