A counterfactual regret minimization solver for heads-up no-limit Texas hold'em, built from scratch in Rust. It covers the same pieces as Libratus and Pluribus at laptop scale: an abstracted blueprint trained by Monte Carlo CFR, an exact range-vs-range postflop solver, and a measurement stack that reports how exploitable the result is.
Every layer is checked against something it cannot fake: closed-form equilibria, published game values, brute-force enumeration, or an independent implementation.
Each milestone shipped with a walkthrough PDF and only counted as done when its gate passed in CI.
| Milestone | What it adds | Gate |
|---|---|---|
| M1 | Generic Game trait, Kuhn poker, vanilla CFR, exact best response | game value −1/18; strategy inside the analytic equilibrium family |
| M2 | Leduc hold'em, CFR+, Linear CFR, Discounted CFR | 936 information sets; value −0.0856 matches the literature; CFR+ 280× better than CFR at 10k iterations |
| M3 | Perfect-hash 5/6/7-card evaluator (colex rank hashing, direct flush table, 125 KB) | matches a brute-force reference on every 7-card hand; ~62M hands/s |
| M4 | No-limit range-vs-range river and turn solver: vectorized DCFR, O(n) showdowns with card removal | polarized river matches the closed form P/(P+B); agrees with the scalar engine; catches 5 injected bugs |
| M5 | Flop solving: rayon parallelism with safe split_at_mut, exact suit isomorphism, memory budget | parallel results bit-identical to serial; isomorphic solve equals the full solve |
| M6.1 | Perfect hand-isomorphism indexer and full no-limit rules | flop indexer proven a bijection exhaustively; 169 / 1.29M / 13.96M / 123.16M classes |
| M6.2 | Exact hand-strength tables for every class; E[HS²] percentile buckets (169/50/50/50) | preflop equities match published values (AA 85.20%, 72o 34.58%) |
| M6.3 | Abstract no-limit blueprint (13,704 nodes) trained by multithreaded external-sampling MCCFR | MCCFR converges on Leduc against the exact engine; checkpoints round-trip |
| M6.4 | Duplicate-match arena; f64 accumulators; blueprint exported to a browser table | self-play scores exactly 0; browser bucketing bit-identical to Rust on 660 spots |
| M6.5 | Local best response: exploitability lower bound with exact Bayesian range tracking | equities match brute force; ranges equal Bayes' rule to 10⁻¹² |
Small games are solved exactly and verified with a full best response. The full game is too large for that, so the blueprint is measured with local best response, which finds a lower bound on how much a counter-strategy can win.
The postflop solver was checked against an independent scalar implementation, and deliberately injected bugs were used to test the check itself. It first caught only one of three: the test ranges had no tied hands and no shared combos, so those code paths never ran. Adding them made all five injected bugs fail the test.
A checkpoint trained 12× longer lost to a shorter one by a statistically clear margin. The cause was 32-bit accumulators: past about 8 million, adding a probability like 0.3 rounds to nothing, so 3.4% of the strategy table had silently stopped averaging. Switching to 64-bit fixed it at no speed cost.
Matches between checkpoints became intransitive: A beat B, B tied C, C beat A. Near equilibrium, a match measures how leaks line up, not quality. Local best response ranked them instead, and showed that more training was no longer the bottleneck.
Limits, stated plainly: the blueprint plays an abstracted game with a fixed bet menu and no real-time search, and local best response gives a lower bound, not the true exploitability. Real-time re-solving of the turn and river with the exact solver is the next milestone.