# Into: Poker44 A Bittensor subnet that turns online poker bot detection into an open competition, where independent operators submit models that score how likely a stretch of play came from a bot instead of a human. // Bot detection for online poker --- > New to Bittensor? Start here. Experienced users can skip to the full analysis. ### What is Poker44? Poker44 is a subnet built around one job: spotting bots in online poker. Instead of leaving that to a single platform's internal security team, it turns detection into a competition where independent operators submit models that flag bot-like play, and the best models earn the rewards. **The simple version:** It is like a spam filter, but for poker hands. It looks at how someone played and scores how likely a bot was behind the decisions rather than a person. **Centralized equivalent:** The in-house anti-cheat and security teams that sites like PokerStars or GGPoker run internally. Poker44 rebuilds that function as an open, reproducible contest rather than a closed black box. **How it works:** - **Miners** receive batched sequences of poker behavior, called chunks, and return one bot-risk score per chunk. - **Validators** query miners with those payloads, score the predictions on accuracy and false positives, and set weights on-chain that decide who gets paid. --- ### Why This Matters - **The problem it solves:** Bots are a persistent drain on online poker, quietly taking money from real players and eroding trust in the games. Most detection today is proprietary and opaque, so players have little way to verify it works. - **The opportunity:** A shared, reproducible benchmark for bot detection that improves in the open. If it works, the same evaluation could in principle back anti-bot workflows for more than one platform. - **The Bittensor advantage:** Bot behavior keeps evolving, so a static rule set goes stale. A standing competition pushes models to keep generalizing, and the optional model_manifest lets miners disclose how a model was built for traceability. - **Traction signals:** Early and honest. The codebase is active, with 154 commits since the repository went up in February 2026, almost all from a single developer account. Public social presence is minimal, and on-chain participation is thin: TaoSwap currently reports a single active miner. --- ## Full Analysis **Category:** Game Integrity and Bot Detection (Other) | **Centralized Competitor:** In-house poker platform security teams (PokerStars, GGPoker) Online poker integrity is usually handled privately. Each room decides on its own what counts as a bot, runs its own detection, and rarely shows its work. That keeps the methods secret from cheaters, but it also keeps them unverifiable to everyone else. Poker44 takes the opposite stance: make the evaluation reproducible and let independent miners compete to detect bots well. **Mechanism:** According to the project's repository, miners receive a `DetectionSynapse` carrying a list of chunks, where each chunk is a sequence of hand payloads, and they return a single risk score per chunk. Validators query miners with these standardized payloads, score the returned risk scores and predicted labels, and publish weights on-chain. The repository states miners are judged on accuracy, calibration, low false-positive rates, and robustness as bot behavior changes over time. The project describes itself as security infrastructure, not a poker room. The repository is explicit about how evaluation is produced today. In the current production model, validators do not run their own poker tables. Instead they consume canonical evaluation material from a central evaluation API operated by Poker44 platform infrastructure, then query miners, score responses, and set weights. The competition runs in rolling 72-hour epochs anchored at 20:00 UTC, with canonical evaluation material refreshed in 6-hour windows. The repository also notes that competition policy and allocation rules are determined by the platform runtime and may evolve independently of the reference code. The TAO.app description adds that validators are expected to evaluate against a private local human dataset rather than the public corpus, to avoid data leakage. To support transparency, Poker44 defines a lightweight `model_manifest` that miners can attach to responses, carrying fields like the repository URL, commit, model name, framework, license, and a training-data statement. Per the repository, the manifest does not change validator scoring or on-chain weight setting; it adds traceability rather than reward. On development, the picture is consistent with an early but active project. A live GitHub check shows 154 commits with the most recent push on 30 May 2026, and recent work focused on the validator audit lane, encrypting validator audit reports, and tightening validator payloads. The work comes almost entirely from one developer account. The subnet draws roughly 1.5 percent of network emissions (about 1.24 percent on the smoothed EMA), trades near 0.00693 TAO, and holds about 1,990 TAO of root liquidity in its pool. A whitepaper is published at poker44.net. --- ### Risk Factors - **Centralized evaluation dependency:** By the repository's own description, evaluation data comes from Poker44 platform infrastructure through a central API, and competition policy is set by the platform runtime. That keeps the contest coherent, but it concentrates a lot of trust in one operator, which is worth weighing against the decentralization the design aims for. - **Thin participation:** TaoSwap currently reports a single active miner and a high miner-burn rate near 97 percent, so most of the miner emission is not reaching active participants yet. A detection contest is only as strong as the field competing in it. - **Concentration:** A Gini coefficient of 0.834 suggests concentrated ownership or stake distribution. Large positions could significantly impact pool dynamics. - **Execution:** This is a young subnet, with a repository created in February 2026 and development concentrated in one account. The hard part, generalizing to live adversarial bots that adapt, is still unproven in public. --- Into the next one.