Question 1

How does Poker44 detect bots if miners don't receive hand details or timing data?

Accepted Answer

Miners receive only seven coarse decision features per session (phase, position, pressure, action type, size bucket, all-in flag, and decision number), derived from tournament telemetry but stripped of cards, exact chip amounts, and timing. Detection relies on behavioral patterns in how decisions are made across multiple streets, not on isolated actions or metadata. The validator scores each miner's calibrated risk scores against private labels using actor-balanced metrics to reward models that generalize across diverse players and bot families.

Question 2

Why does Poker44 keep labels private and separate from miners?

Accepted Answer

Keeping labels inside validators prevents miners from overfitting to static public benchmarks and eliminates label leakage across evaluation rounds. Every miner faces the same immutable dataset in the same order, but only validators hold ground truth. This ensures models are scored on genuine generalization rather than memorization, and makes the evaluation resistant to adversarial label-gaming strategies.

Question 3

Can I run a Poker44 miner with my own detection model?

Accepted Answer

Yes. Miners implement a model factory pattern—you provide a `POKER44_MODEL_FACTORY` environment variable pointing to your model creation function, which must return an object with `version`, `load()`, and `predict(sessions: list[dict]) -> list[float]` methods. Your model receives ordered lists of schema-v4.1 sessions and must return exactly one finite risk score in [0,1] per session. See the miner integration guide and example request/response contracts for implementation details.

Question 4

How often do validators evaluate miners, and what determines evaluation timing?

Accepted Answer

Validators do not follow a fixed schedule. They poll the platform API for sealed evaluation windows, which are published only when the platform has accumulated enough comparable, quality-checked tournament sessions meeting diversity and comparability thresholds. Depending on tournament volume, a new window may be available after one day, three days, or longer. When a window is available, validators query all eligible, reachable miners concurrently and score their responses against private labels.

Question 5

What happens if a miner returns invalid or incorrect risk scores?

Accepted Answer

Invalid, missing, non-finite, out-of-range (outside [0,1]), or wrong-length responses score zero for that evaluation cycle and receive no reward. The validator deterministically selects the highest-scoring response across all miners. Exact ties are broken by lower UID. All miner scores and errors are reported to the Poker44 dashboard for observability.

Question 6

Can I train my model on public data before running a miner?

Accepted Answer

Yes. Poker44 publishes a public training corpus of labeled schema-v4.1 micro-sessions generated by miner-training tournaments available at staging.platform.poker44.net. The corpus is separated into `preview` (usable for early integration) and `stable` (passed quality audits covering human/bot coverage, actor diversity, and distribution checks). You can download the latest release, verify its SHA-256 hash, and train on the `payload` field using the `label` as your target before deployment.

Attelier

Attelier

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup

Participate

Participate

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup