One hand through the whole system

What does PokerAlpha actually do with a hand? One real river decision, followed end to end: the state, the modelled opponent belief, the abstract information set, the confidence gate, and the learned mixed strategy.

The spot

1 · The modelled opponent belief (street by street)

MODELLED BELIEF — NOT GROUND TRUTH. A position/line prior updated by Bayes’ rule after each observed action, under a behavior archetype. This is the same computation the decision engine uses (same function, truncated per street).

After River: 7.64 bits · ≈200 effective combos — each passive action narrows the modelled range.

2 · The abstract information set

infoset key

3|0|p800b2|urd|ior0fnoneans3|c.b33.b75.b150.a

training visits

226

Full Hold’em has ~10¹⁶⁰ states, so the engine maps the real state into a compressed key — street, position, a river-strength percentile bucket, and the shape of the betting history. Strategically different hands CAN share a key; that risk is measured, not ignored (see the gate below).

3 · The confidence gate

ACCEPTED this key has 226 training visits and low cross-seed disagreement, so the trained strategy is used directly. On most preflop nodes the same gate REJECTS and the engine falls back — refusing is a feature.

4 · The learned strategy

check 83% · bet_33 9% · bet_150 5% · bet_75 2% · all_in 0%

A DISTRIBUTION, not a single answer: the solver checks 83% here and mixes bets at several sizes. In hidden-information games, deliberately mixing is what makes a strategy hard to exploit. Hero equity under the modelled range: 92% ±1%.

5 · The receipt

Every analysis can carry a reproducibility record — the state fingerprint, strategy artifact SHA, configuration and seeds — with a stable ID. This one is PA-F668F249; feeding the receipt back to the engine re-runs the analysis and confirms the identical result.