01
Research log · Heads-up no-limit hold'em
A poker bot, measured
against GTO Wizard
Every version is played against GTO Wizard's benchmark bot and scored with AIVAT, a variance-reduced estimate of win rate. This page tracks what was tried, what worked, and by how much.
Progress
AIVAT by version
bb/100 against GTO Wizard, higher is better. Whiskers are 95% intervals; dashed lines are public leaderboard references.
Dot-and-whisker chart of AIVAT per bot version. The same numbers are in the table below.
Architecture
How the bot decides
02
Preflop guards
03
Postflop range solver
04
Measurement
Experiments
Every version, in order
| Version | What changed | Hands | AIVAT (bb/100) | Verdict |
|---|
Findings
What the data said
Solve the spot, don't look it up
The biggest single gain came from solving postflop streets exactly against the actual ranges instead of reading a history-blind table: +16 bb/100 over the best table-only bot.
Find the leak before fixing it
Loss attribution pointed at calling 4-bets with hands like 98o. A targeted guard fixed it; retraining a larger blueprint locally would have needed months of compute.
Measure what you can, replay the rest
10,000 hands only resolves about ±8 bb/100. Smaller changes were validated by replaying recorded hands with held-out data, which caught a false positive from 200 spots.
Live
Running now
Contact