Play it now: quantplay.io — no account, no server, nothing to install. The whole engine (Python, ~1.5 MB) is downloaded once and runs on your machine via WebAssembly. Your hands never leave your browser.
QuantPlay grew out of a three-month attempt to build a heads-up bot that could rank in the top 5 of the
GTO Wizard AI leaderboard. That goal was not reached (the honest balance sheet is
!_PROJEKT_BILANZ_2026-09-10.md). What came out of it, and what this
repository is about, is a 6-max trainer you cannot buy anywhere, a bot with every number measured and
sourced, and a measurement discipline that refuted more ideas than it confirmed.
Six modes, all against a league of profiled bots (TAG, LAG, nit, station, maniac, whale …) whose play was tuned in paired self-play and gated against regressions:
Plus: pre-fold while others still act (the hand is played out in the background and chips move correctly),
hand replay with per-decision grades, an opponent panel showing what the bots have learned about you, a
session analysis after ~100 hands, and a 61-entry poker glossary whose formulas are the audited ones in
knowledge_base/math/formulas.py. The UI is in German.
pip install -r requirements.txt
python -m pokerbot.web.six_server --open --trainerWindows / PowerShell, Python 3.12, run from the repo root. The trainer is at http://127.0.0.1:8000/training.
Bot decisions are local, instant and free; no LLM is involved in play.
web/ turns the same trainer into a static site. web/build.py zips the Python
package plus the knowledge files it reads at runtime; web/src/worker.js boots
Pyodide in a Web Worker, installs five pure-Python wheels and imports the trainer;
web/src/bridge.js replaces fetch('/api/…') so the unchanged front-end talks to
the worker instead of a server. The route dispatcher is
pokerbot/web/browser_bridge.py, which calls the FastAPI endpoints
directly (Pyodide has no threads for the ASGI threadpool). Tested under CPython in
tests/test_browser_bridge.py. Measured in Node + Pyodide: import 0.9 s,
five full hands with grading in 0.32 s, slowest single request 0.11 s (tournament mode 0.35 s).
python web/build.py # -> web/dist (deterministic, content-hashed)
python -m http.server 8765 --directory web/distThe site is deployed from web/ to Vercel. Prince (the heads-up bot) does not take over heads-up pots in
the trainer: measured in 2,992 paired decks it lost against the league core, so it is off by default and
the browser build ships without its neural advisors.
A heuristic engine (preflop blueprint from CFR push/fold + solver-distilled tables, postflop equity /
pot odds / MDF with solver-frequency advisors) wrapped in a chain of measured guards, a bounded exploit
overlay, and optional real-time re-solving (TexasSolver) at river/turn nodes. Every number below has a
source in docs/MESSKATALOG.md; the channel decides what a number means.
What was learned the hard way is in the balance sheet: guards around a heuristic engine hit an asymptote
(≈ −10 predicted, never reached); self-play gains do not transfer to a re-solver; the per-hand standard
deviation is 294 bb, so a ±4 bb/100 answer costs ~5,400 hands; and a well-prompted frontier LLM plays
heads-up better than this bot. Ideas that were refuted by measurement are listed explicitly in
docs/MODULKATALOG.md — for anyone rebuilding, that is the most valuable part.
CLAUDE.md holds the working conventions for AI-assisted sessions, including the measurement doctrine
(paired decks, A/A null test must be exactly 0, pre-registered expectations, three-run rule).
PolyForm Noncommercial 1.0.0. You may use, copy, modify and share everything here for
noncommercial purposes: personal study, research, teaching, hobby projects, and use by noncommercial
organizations. Any commercial use — selling, running as a paid service, using the bot, trainer, ranges,
knowledge base or measurements inside a commercial product or to make money at the tables for a business —
requires written permission. Ask via a GitHub issue or the contact on the organization page. Keep the
NOTICE file (the Required Notice: line) with every copy.
Closed on 2026-09-10, reopened only to publish. Large artifacts (solver caches, trained nets, LLM checkpoints, hand histories) are not in the repository. The books the knowledge base was extracted from are not included either.