ACR Poker Bot Developer FAQ

Both. A real market exists, largely operated out of Russian-speaking and Eastern European circles, with established engineering behind it — solver-anchored baselines, opponent models, behavioural shaping. The market also produces an order of magnitude more scam listings than working products: credential-stealers wrapped as bots, generic GTO engines marketed with inflated claims, and outright vapourware. The realised buyer experience depends entirely on the seller; community vetting through long-running operators is the only filter that consistently identifies the working products.

Because WPN runs an active security team with a public track record of enforcement. The 2015 takedown (~$1.4M returned, ~30 accounts) was followed by smaller cleanups in 2019 and 2020. WPN publishes outcomes — the public bust reports themselves are part of the enforcement strategy because the deterrent value on the broader bot population is large relative to the cost of prosecuting any single ring. Operators that quietly close accounts produce less player anxiety but also less deterrent; WPN has chosen the louder posture.

A Russian-speaking bot operator (referred to in community accounts as "KhanZ") ran roughly thirty accounts on ACR over an extended period. The primary detection signal was hand-history forensics — distributional pattern matching across the suspect accounts revealed shared bet-sizing histograms, near-identical fold-to-3bet response curves, and VPIP/PFR pairs clustered on solver mass with too-low variance. The cross-skin account graph then joined them on shared deposit and device features. Roughly $1.4M was confiscated and distributed back to identifiable affected opponents; the investigation ran several months.

Technically yes — the game logic is shared across the WPN skins so a single engine handles all four. Operationally no — the cross-skin account graph is the highest-priority signal at L3 of the detection stack. Running a bot farm across the skins under a single device fingerprint, IP, deposit method, or KYC document is the easiest possible catch and would be the fastest path to a published bust. Independent fingerprint isolation across the network is an open and hard engineering problem.

Empirically: NL10 to NL100 6-max cash with rakeback factored in, and low-buy-in MTTs ($5–$33). Rake at WPN is 6–8% on NL10 with significant rakeback offsets, dropping to roughly 4–5% net on NL50 and 3% net on NL200. Above NL200 the regulars are sophisticated enough that solver-anchored EV compresses meaningfully. The EV-per-hour curve typically peaks around NL25–NL100 for current solver-driven engines, depending on the specific pool composition.

Yes, and arguably more viable than NLH for the right engine. The ACR PLO pool is softer than most networks at the small and mid stakes — the population skews toward gamblers chasing equities rather than balanced regulars, and the format is mathematically richer (six combinations per hand, flatter equities, more meaningful blockers). PLO solvers have matured significantly since 2021 (MonkerSolver in particular); the opponent-modelling features that work for NLH carry less signal in PLO and require a different feature set.

Technically the major HUDs work on ACR because screen names are fixed and hand histories are stable. WPN's terms of service prohibit third-party tracking software; enforcement is uneven and in practice many winning regulars use one of PT4 or HM3 without issue. The risk profile is real but bounded: HUD use alone does not trigger bot-tier review queues; it can contribute to a softer-restriction outcome if combined with other signals. The conservative posture is to use a HUD only if your pool warrants it.

More transparently than most. WPN's pattern is to publish bust outcomes — sometimes with dollar figures, sometimes just with category and approximate scale. The 2015 announcement set the template. The internal architecture is the standard four-layer stack: hand-history forensics, behavioural fingerprinting, cross-skin account graph, human review. The differentiator is the willingness to publish — most operators close accounts quietly, which optimises for player anxiety but loses the deterrent value WPN extracts from public communication.

Ignition is structurally harder. Ignition runs anonymous tables with no fixed screen names, which collapses the long-horizon HUD attack for both the bot and the operator. ACR's fixed names make opponent modelling materially easier (PokerTracker 4 / HM3 work) but they also give WPN's hand-history forensics layer a stable identifier to accumulate statistics against — which is what produced the 2015 bust. The net effect is that ACR is easier to play well but easier to get caught; Ignition is harder to play well but harder to flag.

Not bot-economic. The Venom main event runs a million-dollar guarantee but at variance levels where evaluating edge requires thousands of buy-ins — a sample size most operators will not let a single account accumulate without escalating review. The operator's attention on the Venom field is correspondingly high, both for ghosting (the more common cheating mode in high-buy-in MTTs) and for botting. Low-buy-in nightly MTTs ($5–$33) are the bot-viable end of ACR's tournament schedule.

Functional but signal-bearing. ACR's crypto cashout path makes the legacy banking-trail problem go away — there is no compliance trigger from a bank flagging a $30,000 wire from a Curaçao operator. The on-chain side is the new signal. WPN performs wallet clustering on deposits and withdrawals: a single wallet feeding multiple accounts, or a withdrawal wallet receiving funds from multiple bot-suspect accounts, is a Layer 3 graph signal. Drip-withdrawing under variable amounts to fresh wallets reduces the signal but does not eliminate it.

Faster on ACR because screen names are fixed and hand histories accumulate stably. A useful exploitative deviation against an unknown opponent typically takes 80–150 hands of joint observation to estimate VPIP, PFR, and 3-bet to within ±5pp with confidence intervals overlaid on c-bet and fold-to-3bet response curves. On an anonymous network like Ignition the same convergence requires within-session online estimation only, which empirically takes a similar or larger sample. The same fixed-name property is what makes WPN's hand-history forensics layer effective.

Quiet flag → soft restriction → structured interview → confiscation and closure. The cycle runs 2 weeks to 9 months depending on review-queue capacity and triggering events. The biggest accelerator at WPN specifically is a large first withdrawal, especially in crypto where the operator wants to act before the funds leave the platform. The 2015 bust ran several months from first signal to publicly announced action; quieter cleanups have run faster.

IP address, device fingerprint (browser/OS, screen resolution, font fingerprint, audio fingerprint, hardware concurrency), deposit method including credit card BIN and crypto wallet address, KYC document (passport / driver licence number, address), table co-occurrence patterns across skins, action correlations within shared hands. Crypto wallet clustering on the public blockchain is increasingly central — the same wallet appearing as a deposit source for ACR account A and BlackChip account B is a near-deterministic graph edge.

Different problem from a mobile-first network. The desktop client surface — installed OS, browser version, screen geometry, font and audio fingerprint, GPU model, locale and timezone, accessibility tree contents — gives the operator a much wider passive telemetry channel than mobile. Production anti-fingerprinting at scale typically uses dedicated VPS or bare-metal hosts per account with isolated OS images, rotated through residential proxies; the cost per active account is meaningfully higher than on a mobile-first target, but the realised edge per active account is also higher because ACR's pool is softer.

Marginal for live decisions, useful for post-hoc analysis. Frontier LLMs hallucinate ranges, misapply ICM, and lack frequency intuition — a $5 heads-up solver-driven bot beats GPT-class models in head-to-head play. The useful role is annotation: feeding suspicious hand histories into an LLM to generate exploit hypotheses worth solver-checking, summarising solver output into written form for review, generating training data for narrower models. The boundary between "useful annotator" and "useful in-the-loop player" is sharper than the field acknowledges.

Comfortable. On a modern desktop CPU (any 2020+ Intel/AMD class), solver-table lookup runs in 1–10ms, opponent-model update in under 5ms, UI inspection on the rendered client 5–20ms. Total compute budget under 50ms per action with significant headroom. The behavioural-timing layer adds the human-perceptible delay on top — typically 600ms to several seconds depending on decision difficulty. The constraint at scale is multi-table CPU contention if a single host runs many tables, not the per-action budget.

Two compression layers. State abstraction: collapse the game state into smaller equivalence classes based on board texture buckets, stack-to-pot ratio buckets, position, and discretised action history — reducing the lookup key from the full state space to an integer index. Action abstraction: discretise possible bet sizes to a finite set (33%, 66%, 100%, 150%, all-in) and solve for that abstraction level. The two together typically yield four to five orders of magnitude compression off the raw CFR output with bounded EV loss relative to the abstraction's optimum.

Hardware-level entropy is the hardest surface. Desktop signals: per-CPU clock drift patterns, GPU model fingerprint via WebGL, audio device fingerprint, exact font rendering geometry. Behavioural signals over long sessions: the natural drift of human attention (more misclicks late, longer think times after long sessions, occasional table-switching mid-hand). A device permanently in a server rack does not produce the same low-frequency telemetry pattern as a device on a desk being used by a human, even if the active session features are matched.

Rough operational estimate: a solver baseline produces near-zero EV against strong regulars, positive EV against the population. Opponent modelling adds 1–4 bb/100 against exploitable opponents depending on pool. Behavioural shaping with detection awareness costs 0.5–1.5 bb/100 relative to pure-GTO output. Net realised winrate typically lands at 2–5 bb/100 at NL25–NL100 on ACR, scaling with hands-per-hour (60–80 per table multi-tabled) and effective stake. PLO upside is higher on the same scaffolding. Numbers compress as the pool's solver awareness improves.

Read the three long-form notes (hacks, detection, and the homepage open questions). The chat at the bottom of every page is read by the team. Most useful contributions arrive as a specific question with a measurement attached, or as a piece of relevant data — academic dataset, empirical bust pattern, public WPN communication we have not seen. Sales messages ("can I buy your bot") are auto-archived.