Independent AI poker research collective

RhythmLag

AI poker and decision systems under uncertainty

Public research portfolio Edition 02
Beijing · 2026
01 Project
English · EN

Poker as a laboratory for intelligence.

RhythmLag is an independent AI poker research system for multi-agent decision-making under uncertainty. Poker provides a rigorous, measurable environment for studying imperfect-information reasoning, strategic adaptation, and long-horizon decisions. The system combines persistent self-play, versioned policy milestones, evidence-led Champion selection, and bounded real-time search in one reproducible loop. Its methods are platform-neutral and designed to be evaluated across poker environments. The aim is to develop poker agents that can be measured, challenged, and improved over time, while extracting transferable insights for intelligent decision systems.

02 Research modules
A portfolio, not a single model

Five connected modules turn training into a research program.

RhythmLag is organized as a chain of distinct questions: how a policy learns over time, where a bounded experiment should concentrate effort, what evidence is sufficient to compare two states, how a control layer may recommend a transition without silently acquiring authority, and how a verified policy becomes the starting point for decision-time search. The modules share provenance, but each keeps its own measurement and permission boundary.

237 / 237Focused contract checks passed
101Enhance Training checks · 24.51 s
136Takeover checks · 4.77 s
19Focused test files across both systems
Learning linePersistent

Policy learning and milestone lineage

Long-horizon self-play advances a versioned policy lineage. A milestone is treated as an identifiable research object: its parentage, accumulated evidence and evaluation context remain bound together so that later comparisons refer to the same object.

Boundary · a newer artifact is not automatically a stronger or publishable policy
Enhance TrainingBounded experiment

Directional learning without rewriting the baseline

Enhance Training isolates a declared strategic question and allocates a bounded learning campaign around it. The source policy remains identifiable, accepted work is recorded, and the resulting candidate returns to comparison rather than replacing the main line by assumption.

Boundary · targeted learning is an experiment, not Champion or production authority
TakeoverGoverned control research

Observe, diagnose, replay and recommend

Takeover studies how a control layer can assemble immutable observations, diagnose an incident, replay a proposed state transition and issue a bounded recommendation. Its present public meaning is governance: evidence can move forward while execution permission stays with the operator.

Boundary · local evidence only; no implied authority over live training, Champion or production
EvaluationEvidence classes

Comparison before selection

Paired self-play, coverage measurements and external decision replays answer different questions. RhythmLag keeps those measurements separate, records uncertainty, and asks whether evidence supports a research decision before any candidate is described as selected.

Boundary · coverage, agreement and strategic value are not interchangeable metrics
Decision layerVerified input

Bounded real-time search from an exact policy

Decision-time search is downstream of policy learning, not a substitute for it. It starts from an exact verified policy, reasons within a bounded decision budget and preserves the separation between offline learning evidence, online search behavior and any platform-facing integration.

Boundary · a search result does not alter the training lineage or publication state
03 Collective
English · EN

Different disciplines, one standard of inquiry.

RhythmLag is built by a small circle of like-minded friends and alumni of Peking University, Peking University Health Science Center, Tsinghua University, Stanford School of Medicine, and KU Leuven. Our work spans computer science, artificial intelligence, immunology and medicine, nuclear physics, and neighboring fields. We work independently, combining engineering practice, experiment, and scientific criticism.

Computer ScienceArtificial IntelligenceImmunology & MedicineNuclear Physics
04 Research snapshot
Research snapshot · updated 30 August 2026

Historical training evidence and current systems verification.

Figure 01 presents the current verification surface for Enhance Training and Takeover, grouped by the public capability each check exercises. Figures 02–04 retain the receipt-backed M1–M20 record: training chronology, indexed policy growth, paired diagnostics and action-reproduction studies.

M1–M20 evidence + current systems SNAPSHOT · 2026-08-30 · PUBLIC RESEARCH EDITION
8.21K hCumulative worker time at M10
310.4MIndexed postflop rows at M10
650,016Mirrored pairs per stage comparison
15,169Public 10K decisions replayed at M20
11,000Independent benchmark inputs
FIGURE 02 · M1—M10

Training chronology and indexed policy growth

Cumulative CPU-worker hours Indexed postflop rows
100500 247.5 h32.4M8.21K h · 310.4M M1M2M3M4M5M6M7M8M9M10
Both series are normalized independently to their M10 endpoint; labels preserve the observed values. Indexed rows grew 9.6× from M1 to M10. Row count measures indexed policy coverage; playing strength is evaluated separately.
FIGURE 03 · 95% CI

Adjacent milestone paired deltas

M2 : M1+103.65 M3 : M2+65.22 M4 : M3+55.81 M5 : M4+46.71 M6 : M5+26.36 050100125 bb/100
Paired self-play estimates under the reported coverage and fallback protocol; 650,016 mirrored pairs per comparison. Whiskers show 95% confidence intervals.
FIGURE 04 · M20

Action-reproduction studies

Public 10K reference · M2015,169 decisions
Exact action65.76%
Action family73.87%
PokerBench · M2011,000 / 11,000
Exact action51.29%
Action family62.15%
Representable exact65.94%
Agreement measures action reproduction on fixed records and forms a separate measurement class from win rate, EV, or exploitability. PokerBench: 10,950 converted inputs, 50 conversion failures, 6,056 scored decisions, and 4,710 representable scored labels.
FIGURE 01 · CURRENT SYSTEMS · 2026-08-30

Verification coverage by capability domain

Enhance TrainingTakeover
101 / 1019 TEST FILES · 24.51 S
Orchestration & distributed41
Learning modes28
Transition & identity9
Checkpoint reliability8
Public status surface15
136 / 13610 TEST FILES · 4.77 S
Governance core47
Incident & diagnosis49
Proposal replay4
Operator flow34
Public status surface2
The 237 collected checks are grouped by the public capability they exercise. Bar lengths use a common 0–50 scale. This figure measures contract-verification coverage, not gameplay strength or authority over live systems.
05 Publications
Working papers and research notes
Systems Note 02 · RL-SN02-2026

Enhance Training and Governed Takeover

A structured account of persistent policy learning, bounded Enhance Training, paired evidence, effective-work accounting and a Takeover layer whose verification surface now spans governance, diagnosis, replay and operator flow.

7 pagesEN · CN · DESystems note30 Aug 2026 Read new note Research Note 01 · RL-RN01-2026

AI Poker as a Laboratory for Decision Intelligence

An introduction to imperfect-information decision research, milestone-based learning, and the distinction between coverage, action agreement, and strategic performance.

7 pagesEN · CN · DEResearch edition Read note Systems Note 01 · RL-SN01-2026

Computational Environment and Measurement Scope

A systems note on aggregate computing capacity, cumulative worker time, indexed policy coverage, and the measurement units used across M1–M10.

7 pagesEN · CN · DESystems note Read note Working Paper 01 · RL-WP01-2026

Mathematical Framing and Computational Context in AI Poker Research

A compact treatment of counterfactual regret, indexed-state coverage, paired estimation, and the computational context of milestone-based AI poker research.

7 pagesEN · CN · DEWorking paper Read paper

Each online edition includes a stable document identifier, publication date, page index, and direct navigation between its English, Chinese, and German sections.

06 Alumni initiative
Alumni-led and independent

This project was initiated by alumni and is conducted independently. It is not an official research project of any university shown below.

Peking University mark Alumni background Peking University Peking University Health Science Center mark Alumni background Peking University Health Science Center Tsinghua University mark Alumni background Tsinghua University Stanford School of Medicine mark Alumni background Stanford School of Medicine KU Leuven mark Alumni background KU Leuven