Releases
Every strategy the site ships, with its grade.
Each release is a frozen strategy published as a GitHub release asset and fetched by its SHA-256. Its grade is the exact best response's verdict, read from the committed file the catalogue names.
| Release | Published | Tag | Exploitability | Shipped to | Exported by |
|---|---|---|---|---|---|
deep-pot-piDeep CFR, 4 × 256 net, the policy net Π at iteration 200, regret targets ÷ √potdeep-pot0.5-256x4-s16000-lr5e-4-t200-pi.npz | 2026-10-28 | rung4-deep-pot0.5-t200 | 0.116chips per hand at 2 million hands per seat | no page; published only | — |
deep-pot-exactDeep CFR, 4 × 256 net, the exact average at iteration 200, regret targets ÷ √potdeep-pot0.5-256x4-s16000-lr5e-4-t200-exact.npz | 2026-10-28 | rung4-deep-pot0.5-t200 | 0.086chips per hand at 2 million hands per seat | Play | uv run minidraw-deals |
deep-piDeep CFR, 4 × 256 net, the policy net Π at iteration 500deep-256x4-s16000-lr5e-4-t500-pi.npz | 2026-10-26 | rung4-deep-t500 | 0.136chips per hand at 5 million hands per seat | Range explorerLearn | uv run minidraw-ranges --release deep-pi |
deep-exactDeep CFR, 4 × 256 net, the exact average at iteration 500deep-256x4-s16000-lr5e-4-t500-exact.npz | 2026-10-26 | rung4-deep-t500 | 0.108chips per hand at 5 million hands per seat | Range explorerLearn | uv run minidraw-ranges --release deep-exact |
tabularTabular LCFR, the linear average at 2,000,000 iterationsstrategy-lcfr-2m.npz | 2026-10-02 | rung3-lcfr-2m | 0.061chips per hand at 20 million hands per seat | Range explorerLearn | uv run minidraw-ranges --release tabular |
The catalogue is RELEASES in src/drawmaha_solver/minidrawmaha/answer/releases.py; each grade is read from the file and key it names (figures/rung3/answer_sheet.json, figures/rung4/). An export copies the grade into its manifest, and the tests hold the copy to the file.
Π is the original Deep CFR's policy net: a second net fit to the strategies a run sampled, standing in for their average. The exact average sums those strategies with no second fit, and it is less exploitable than the Π of the same run; the research page traces the gap to Π's fit.