Benchmarks

Committed, reproducible measurements — the figures below are generated by scripts in the repository from recorded results, not illustrations.

Native C++ backend speedup vs Python reference
Native C++17 training backend vs the Python reference: 74.2× warm training throughput on the release configuration, with bit-exact regret/strategy tapes.
Confidence gate risk/coverage
Solver confidence gate risk/coverage: tightening the gate trades coverage for lower disagreement risk on held-out checkpoints.

Local product performance

API cold start → /health~1.6 s
Warm /analyze (2000 equity / 1000 rollout sims)~0.10 s
Full 169-hand preflop matrix (169 artifact lookups)~1.0 s

Measured on an Apple-silicon laptop against the committed 2M release artifact; your numbers will vary with hardware and settings.

Reproduce everything: benchmark methodology · recorded results