Research Lab
PokerAlpha as an interactive laboratory: every page below leads with a question, answers it with real engine output, and ends with its method, limitation and reproduction command.
Research Phase II
What did controlled experiments add — and what failed?
In exact river subgames: error-guided abstraction refinement beats equal-capacity uniform, safe decision-time resolving, translation costs. Plus 1.96× bit-identical native training speedup and five negative results. Research only; the release is unchanged.
Watch a Solver LearnCan you watch strategy learning happen?
Live CFR in your browser, validated against the Python reference; exploitability falls from 0.33 to 0.0015 in 10k iterations.
ArenaCan you exploit a trained strategy?
Play Kuhn vs the CFR policy; afterwards see the exact exploitability of YOUR observed strategy.
Decision SurfaceWhere does the recommended strategy change?
81 real engine analyses over bet size × stack; watch call become fold and the gate degrade.
Solver Trust MapWhere does the solver trust itself?
169 hands × 3 nodes with real gate verdicts, visits and cross-seed disagreement — mostly rejected preflop, honestly.
Solver MicroscopeHow did one policy evolve across training and seeds?
Nine real artifacts (300k/1M/2M × seeds 0-2); flip between low- and high-disagreement infosets.
Policy AtlasWhat kinds of policies did the solver learn?
6,000 real infosets, PCA-projected; 37% of the table is highly mixed strategy.
Break It / Red TeamWhere is PokerAlpha fragile?
Automated adversarial search over legal states; failure clusters and the safeguards that fire.
Hand Replay + Range EvolutionWhat did the system believe during a hand?
A real decision with the modelled opponent range evolving street by street.
Agent ArenaWho beats whom, with what uncertainty?
Seat-swapped, seed-paired tournaments; the oracle's measured edge matches computed exploitability.
Game LabDoes this extend beyond poker?
Kuhn, Leduc, and a sealed-bid auction duel — one solver stack, three games, exact exploitability each.
EvidenceCan every headline number be traced?
A machine-checked claims registry: value → method → source file, plus the nine-rung validation ladder.
Everything on these pages works without a local engine — the data is PRECOMPUTED real output, regenerated by one script (scripts/build_web_research_data.py) and stamped with the commit and strategy SHA that produced it.