Inspiration
We wanted a bot that plays like a strong human: sound theory by default, exploiting opponents’ mistakes when it sees them. Jonathan Little’s teaching on pot odds, stack-to-pot ratio (SPR) and exploitative play shaped our approach.
How we built it Audited the short-deck hand rankings against the official evaluator before writing any strategy. Built a Monte Carlo equity engine against a range model that narrows with each opponent action. Tracked each opponent’s fold, call and raise rates by situation, and bluffed only when the expected value of betting beat checking. Tested every change with paired, duplicate-deal matches against a panel of baseline bots, “student” bots, an exploiter bot and our own older versions. We kept a change only if it won on fresh seeds. What we learned
Measurement beats intuition. Many ideas that sounded smart lost in testing. Fixing one real bug mattered more than most strategy tweaks.
Challenges All-in pots create a lot of noise, which hides small edges. Improvements against one opponent type often hurt us against another. Test runs were slow, so we had to choose experiments carefully.
Built With
- claude
- python
