A working notebook on poker AI in the ACR Poker and WPN ecosystem.

What "poker bot" means in 2026

The term "poker bot" is used so loosely in 2026 that most public discussions start from the wrong place. For the purposes of these notes, a poker bot is a piece of software that receives the visible game state for one seat — its own hole cards, the community cards, the action sequence and stack sizes — and emits an action under a real-time latency constraint. A bot does not read packets, does not know opponent hole cards before showdown, and does not see into the deck. Every input available to the bot is also available to a focused human; the bot's edge comes from consistency, speed, and the willingness to grind tens of thousands of hands in a calibrated way.

Three layers stack under that definition. The first is a solver-anchored baseline: pre-computed strategies from counterfactual regret minimisation (CFR) variants — PioSolver or GTO+ for heads-up and 6-max trees, MonkerSolver for multiway and PLO. These outputs are too large to run online and are compressed into action-frequency tables keyed on bucketed game state. The second is the opponent model. On ACR specifically, this layer is easier than on a fully anonymous network because screen names are persistent and hand histories are stable; PokerTracker 4 and Hold'em Manager 3 work in the conventional way (against the operator's terms of service, but technically functional). The third is a policy combiner that decides how far to deviate from the baseline given the current opponent estimate, and overlays detection-aware behavioural noise.

The ACR and WPN ecology in one page

ACR Poker — short for Americas Cardroom — became the dominant US-facing online room after Black Friday in April 2011, when the U.S. Department of Justice indicted PokerStars, Full Tilt and Absolute Poker and effectively closed the US market for the prior generation of operators. ACR runs on the Winning Poker Network alongside three sibling skins (BlackChip Poker, TruePoker and YaPoker), all sharing a player pool, a security team and an account graph. The network operates under a Curaçao licence and accepts crypto deposits and withdrawals, which is part of why it remains usable from most US states despite the federal grey-zone status.

Phil "OMGClayAiken" Galfond is the most visible ambassador in ACR's recent history; he later refocused on his own Run It Once project and the relationship is no longer active, but the Venom MTT series (named after his nickname) survives as the headline weekly schedule with a million-dollar guarantee on the main event. From a bot-economic perspective, Venom-tier MTTs are interesting marketing but bad ROI — the variance is high enough that a sample of thousands of buy-ins is needed to evaluate edge, and the operator's attention on the field is correspondingly higher.

Three structural facts matter for anyone studying the network. Fixed screen names mean long-horizon HUDs work — both for the player and for the operator's security model. Public bust reports are an enforcement signal: WPN published details of the 2015 bot-ring takedown including approximate funds returned (~$1.4M across roughly thirty accounts), and has confirmed several 2019–2020 cleanups. Cross-skin detection joins accounts across ACR, BlackChip, TruePoker and YaPoker on device fingerprint, IP, deposit method and KYC documents — a bot farm spread across skins under one fingerprint is the easiest possible catch.

How WPN detection actually works

WPN's detection topology is layered, asynchronous, and weighted. There is no single trigger event. Signals accumulate over time under an account-specific risk score with a tunable false-positive budget; when the cumulative score exceeds threshold, the account is queued for human review. The 2015 bot-ring case is the publicly best-documented example: roughly thirty accounts identified through hand-history pattern matching, frozen, and rolled back over a multi-month investigation with funds returned to affected opponents.

Hand-history forensics

WPN's most public signal. Statistical pattern matching across long samples: VPIP / PFR distributions clustered too tightly on solver mass, bet-sizing histograms concentrated on exact pot fractions, fold-to-3bet frequencies that match population by ±0.5pp across thousands of hands. The 2015 bust was driven by exactly this layer — the bot operator's accounts shared distributional fingerprints that humans do not produce.

Behavioural fingerprinting on the client

Input-timing distributions, mouse-path curvature, action-confirmation latency, idle-behaviour patterns between hands. ACR's client is desktop-primary (Java in the early generations, native rewrite in the mid-2020s) which gives the operator a wide telemetry surface compared to mobile-first networks.

Cross-skin account graph

Account joins across ACR, BlackChip Poker, TruePoker and YaPoker by IP, device fingerprint, deposit method (including crypto wallet clustering), KYC document, and table co-occurrence. This layer catches multi-accounting and chip dumping directly; botting falls out as a side-product whenever a farm runs under a single fingerprint across skins.

Human review on flagged accounts

The decisive layer. The mathematical models propose; humans decide. A reviewer reads hand history, checks chat behaviour, looks at withdrawal patterns and KYC consistency. Most botting bans are signed off here, not in an automated rule — which is why the gap between first-flag and visible action runs from weeks to many months.

Areas covered on this site

Open questions for developers

These notes are not finished. Five threads I think are genuinely open in the ACR / WPN context:

1. Cross-skin fingerprint isolation. Running accounts on ACR and BlackChip Poker from the same operator without joining the account graph. The naive failure mode is a shared deposit wallet or KYC document; the harder problem is independent device-fingerprint generation that survives passive observation across the network.
2. Hand-history forensics as adversarial classification. WPN's most public catches have come from distributional analysis. Treating that classifier as a noisy adversary and shaping bet-sizing histograms, VPIP/PFR variance, and fold-to-3bet response curves to sit inside the population envelope while preserving EV is the central anti-detection problem at this operator.
3. PLO opponent modelling. ACR's PLO pool is softer than most networks and the format is mathematically richer (six combinations per hand, flatter equities, more meaningful blockers). Opponent-model features that work for NLH — 3-bet frequency by position, fold-to-cbet by board texture — carry less signal in PLO; the open question is what does carry signal and how fast it converges.
4. Low-buy-in MTT automation. The $5–$33 nightly schedule on ACR has enough volume and weak enough fields to be the most promising MTT-bot environment in the US-facing market. ICM is non-trivial; bubble factor varies by structure; the open question is whether MTT-specific opponent modelling pays back its compilation cost.
5. Crypto-tied withdrawal pattern shaping. ACR's crypto-friendly cashout path is a signal the operator monitors. Drip-withdrawing under jurisdictional thresholds, wallet clustering on the chain, and the relationship between withdrawal cadence and review-queue prioritisation are all open empirical questions.

If you have work in progress on any of the above, the chat at the bottom of the site is read by the team. Most useful contributions arrive as a specific question with a measurement attached, not as a general claim.