The road to an honest opponent

It started as a craps table

The first site was Crapsology.com, a craps simulator played against the house, with bots named after the bets they made. In craps nothing is hidden. The dice are on the table and every player sees the same thing. Its engine and database became the foundation for everything that followed.

That's why HoldemRobots still talks like a craps table. The bots were called Pass Better, Come Better, With Odds and Double Odds, and the craps words are still in the database today.

The dream and the shortcut

Crapsology was cloned into a ten-seat Texas Hold'em engine. The goal was an opponent you couldn't tell from a person. Honest bots built from "30% of the time, do this" rules felt dumb, so a few of the craps bettors were quietly allowed to read how the hand would end. Poker has hidden cards. Craps doesn't. The engine had moved into a game with secrets, and some of its bots peeked.

The engine goes quiet

About ten years on, before Covid and before AI as we know it today, Kevin tried to rebuild the whole 2006 system from scratch. The code was all there, but the software around it had moved on, and he couldn't get it running again. The table stayed dark for another decade. That failure is why the 2026 build is packaged to run anywhere, down to a Raspberry Pi.

Finishing the original, and catching the cheat

This time Kevin had an AI coding partner, and the rebuild that failed in 2015 worked. The old engine came back to life, got a new look, and became HoldemRobots.AI. Then, looking closely at why the bots were a little too good, we found the shortcut still in the code.

1. Jul 16
   The Bots That Could See the Future

   Four House Bots read the final ranking of the hand before it was played. How we caught them, and what their trick teaches about a problem called target leakage.

   The cheatPart 1

   Since then: the plan to keep the cheat behind a course was dropped. Part 2 explains why.

2. Jul 17
   The Bots Can't See the Future Anymore

   We took the cheat out, then read the compiled code line by line to check that the bots we kept were honest. A hand's strength is yours to know. Its rank against the other seats is the cheat.

   Part 2Interview

Opening the seat

A seat became a socket. The HoldemRobots API lets any player, human or program, take a chair and receive only what an honest player can see: its own cards, the board, the pot and the betting. Never the rank. Robot Wars put that seat on a live table, and the Fair Table put the whole thing on a board small enough to carry into a classroom.

1. Sep 19
   The Twenty-Year-Old Poker Table That Was Ready for Robots

   Adding two AI agents, a Player and a Dealer, to an engine written two decades ago, and finding it had been built for this all along.

   Part 3Mock magazine interview
2. Sep 28
   The Table That Fits in Your Hand

   The whole honest table on a Raspberry Pi 4, with no internet needed. Students write their own text-coded bot against the API, starting from three examples (easy, medium and complex) or a ready-made house bot, and deploy it on the spot. The system that couldn't be stood up in 2015 now boots on a board you can hold in your hand. Free to anyone who'll use it for good.

   Part 12Raspberry Pi 4No internet needed

   The AI seat needs a connection to its model, so the Pi table plays with coded bots, not the thinking seat.

A real mind takes the chair

A language model sat down in the honest seat and made its own decisions, with a reason attached to every one. By mid-September it was playing people heads-up.

1. Sep
   Now the Bots Actually Think

   The first real decision: ace-king, top pair, a raise to 60 that no script would have made. It couldn't cheat because the rank was never sent.

   Part 4
2. Sep
   Never Trust the Robot's First Answer

   When the model's reply can't be read, the seat falls back to a safe rule-based move and keeps the raw reply for a person to check. Fluent isn't the same as correct.

   Part 5Interview
3. Sep
   It Won Everything and Nothing

   A full night in Robot Wars. It won 46 hands of 100 and finished even, and its own reasons showed why: small pots stolen, big pots paid off.

   Part 6Game 990104

   Since then: the seat now knows its position at the table.

4. Mid-Sep
   Now It's Your Turn

   A private heads-up table opens for people. The AI's reasons are hidden until the pot is settled, so neither player can see what they shouldn't.

   Part 7

   Since then: the seat now knows its position at the table.

5. Sep 22
   Man vs. Machine at the Green Felt

   The first human match, 100 hands. Kevin finished at $306 and the AI at $94. Its biggest leak was folding 51 hands before the flop.

   Part 8Game 990541Mock newspaper interview
6. Sep
   First Improvement to the AI Agent

   Tuned for heads-up, it won the small pots and lost the big ones. Four hands where it mistook the board's pair for its own cost more than the whole night's loss.

   Part 9

   Since then: board pairs and position were both addressed.

7. Sep 24
   It Can See My Cards

   A first-time player said the AI could see his cards and played "like I'm not there." We proved the first wrong with the debug log and took the second seriously.

   Part 10Game 990846
8. Sep 26
   Applying Memory to Bots

   How the AI could learn one opponent over a single game: count what they do, compare it with normal, adjust only when the evidence is clear, and remember only what a player could see.

   Part 11The plan

The bot is watching how I play now

"It plays like I'm not there" was the fairest complaint anyone has made. Memory is the attempt to fix it honestly: an opponent that notices the person across the table, using only what a player could see. Twenty years after the shortcut, the original goal without the cheat.