Read the hand.
Find your moment.

1. A small map of a big deck

Number the ranks from 0 to 12: 2 is 0, and ace is 12. From two reference cards, mark their ranks, their sum modulo 13, and the absolute difference. A rank’s distance is its shortest circular distance to any of these four points.

References = {a, b, (a + b) mod 13, |a − b|}

LIKELY ranks are at distance 0 or 1. UNLIKELY ranks are at distance 2 or farther. For the whole reference pair, this room follows the executable LRO definition: more than half of the 13 ranks must be near, so 7 or more means LIKELY; 6 or fewer means UNLIKELY.

The map can be anchored to your pocket, or updated with the last two revealed cards. The pocket map stays anchored while new cards change its unseen-card counts. The dynamic map can change its reference points at every street. Board alignment is a separate question: are a majority of the board cards near your pocket’s reference points?

The selected equity model uses separate forests for heads-up or multiway play and each street. LIKELY and UNLIKELY examples share those forests: that design won validation among the zone-aware candidates. Pocket and dynamic coverage, board alignment, and zone changes are also inputs. Both zones are trained: an UNLIKELY label alone never means you must fold a strong hand.

2. Give the idea a price

At a river decision with betting ending after the call, a call costing C into a pot P breaks even at a showdown share of C / (P + C). Calling 30 into 90 needs 25%. Earlier streets add future bets, folds and possible raises; raw showdown share is only part of that decision.

A pure bluff that wins only when the opponent folds has expected value F × P − (1 − F) × B. Here B is your bet and F is the fold probability. The break-even fold rate is B / (P + B).

For a 2.5-pot bluff, the required fold rate is about 71.4%. That matters when applying the exploit notes: 75% folds would make that simplified bluff profitable, while 65% would not. A multiway bluff needs the whole relevant field to fold; one opponent’s fold rate is insufficient.

3. Read the people and the street

The guides’ useful organising questions are concrete: who entered the pot, where is the button, how coordinated is the board, and who has been calling or folding? Treat observed habits as clues with a sample size. A few folds do not establish a permanent personality.

“Patient,” “Balanced,” and “A little bolder” are gentle nudges to the action policy. They do not alter the cards or the trained equity value. The range slider lets your adviser explore stronger assumed opposing pockets. It is a synthetic range model, not a read of the actual hidden cards.

Exploit a measured tendency, with an appropriate bet size. We do not label the practice opponents GTO or transfer the PDF’s advertised win rates to this game. Research poker systems use different methods and evaluation: see Brown and Sandholm’s original Pluribus paper.

4. The bell curve belongs to the averages

A single hand can lose, win or split; its outcome is not a bell curve. Under repeated independent sampling, averages can become more concentrated and approximately bell-shaped. The spread of outcomes is standard deviation. The uncertainty of an estimated mean is measured by standard error, which scales as SD / √N under those assumptions.

In the game’s Monte Carlo lab, 4,000 possible opposing pockets and boards are drawn without replacement within each sample. The lab then resamples those outcomes into 600 groups of 100 repeats and plots their average showdown shares. The bars describe repeated versions of that one spot with betting held fixed, not your next 100 live hands.

The displayed interval covers simulation sampling uncertainty under the selected range model. It does not cover a wrong opponent range, future betting, or imperfect model predictions. A profitable decision can still lose; larger samples make an average easier to estimate, not individual cards easier to foresee.

5. What the new forests learned

The training uses the separate pytoolzx copy of pytoolz/WSRF. Monte Carlo showdown shares train the equity forest. Millions of complete simulated betting continuations train a second WSRF forest to compare legal actions by chip return. The guides supply explicit, inspectable features; both LIKELY and UNLIKELY zones receive balanced training.

Results within the zones

Played matches

These checks compare local practice opponents on fresh decks, rotating the candidate through every seat. Benchmark stacks alternate between 40 and 100 big blinds each hand. The games themselves keep stacks between hands. The confidence intervals group seat rotations by deck.

Read the full training notes · Equity receipt · Match measurements · Guide source receipt

A new training experiment

The zone-atlas pass added 16,384 decision positions and 10,047,360 complete training continuations. It used an approximation that weights possible pockets by public calls or raises. Neither candidate cleared the recorded promotion rule across the validation opponents, so the games retain v3.

The separate final comparison completed 92,160 hands. The research model showed gains against v3 and several local policies, but still lost against tight play and earned less than v3 against random play. The full table belongs beside those conclusions.

New fitting receipt · Promotion rule and validation · Fresh paired tournament · Read the training chapters

6. A gentle practice loop

Settle into the new Lantern Lounge ↗

Choose heads-up, six seats, or nine seats. Open Table options for a faster pace, automatic next hands, or practice peeks. A peek marks that hand as practice; it never feeds hidden cards to an adviser. HUD statistics describe completed hands on this device: watch the sample count before trusting a pattern. Download your notebook to back it up, and restore it on the same table size.

1. Begin heads-up with Gentle company. Read your pocket zone and the price before acting.
2. Ask for a hint and open one tree. Notice which features it actually used.
3. At the flop, compare the pocket and dynamic labels. Name what changed.
4. Run the Monte Carlo lab on a difficult spot. Try a narrower range assumption.
5. Finish the hand and write down the decision you would repeat. Then try six or nine seats.

Your pace, your curiosity. The next hand is ready whenever you are.