Now the Bots Actually Think

The honest seat we designed is live. This week a real AI brain sat down in it, made its own call, and could not have cheated if it tried.

In the first article, we caught four bots reading the future. In the second, we pulled the cheat out and then read a compiled class byte by byte to prove the bots we kept were honest, and we ended on a promise. We had drawn one clean line: a hand’s score is yours to know; its rank against the other seats is the cheat. We built a seat, an API, that hands an outside player only honest observation (your cards, the board, the bets, and your own hand strength) and never a rank. The seat, we said, was designed now, and next.

This week, next arrived. A real intelligence sat down in it.

Three ways to decide

The engine now offers a seat three different ways to make up its mind. rules is the old honest heuristic: play this fraction of the time, raise if your hand is strong. math is probability doing the same job with more rigor. And llm is the new one: the seat’s decision is handed to a language model, Claude Haiku, that reads the same honest packet a human would and answers with a move.

There is a quiet poetry in that third option that the developer has pointed at before. Twenty years ago he faked a thinking opponent with a dice roll (thirty percent of the time, do this; three out of ten) because programmed guesses expressed as arithmetic were the only tool he had. The field then spent two decades building the thing that does that for real: a policy that is learned instead of authored. This week the real thing took a chair at his twenty-year-old table.

One honest hand

The test was a single hand, and the interesting part is what the model was told. It held ace-king. The board showed seven, deuce, king: top pair. There was a hundred-and-twenty-chip pot and twenty to call. Over the wire went exactly that: the hole cards, the board, the street, the pot, the amount to call, the minimum raise, the stack, the seat number, how many were playing, the betting so far, and the hand strength, a score of about six-tenths and a plain-English label, “top pair.”

Read that packet again for what is missing. No rank. No other seat’s cards. Nothing from a street that hadn’t been dealt. The model received precisely what the person sitting in seat three could see, and not one bit more.

The model thought, and it raised: sixty chips into a hundred-and-twenty pot. That number is the whole tell. The old rules bot, faced with the same spot, coughs up its canned raise to forty every time. Sixty is not on any script. It came from the model’s own read of the situation. Something at the table was, for the first time and honestly, thinking.

Sixty is not on any script.

Why it can’t cheat, proven live

Here is the payoff the first two articles were building toward. A cheating House Bot needs the rank to function. Its entire personality was “I bet when I’m going to win,” and you can’t know that without peeking at the ending. The rank never crossed the wire. So the model in the seat could not have peeked even if it had wanted to, because the raw material to reconstruct the table was never transmitted to it. We did not police the model. We starved it of contraband.

That is what “honest by construction” looks like when it stops being a slogan and becomes something you can check. It isn’t a promise printed in a rulebook that House Bots may not play here. It’s an absence you can verify in the packet: the thing a cheater would need to eat simply isn’t on the menu. A real mind sat down, and the table’s honesty held not because the mind was virtuous but because the architecture gave it nothing to be dishonest with.

The Turing test, closed the honest way

The original dream, twenty years ago, was a table where a human couldn’t tell which seats were machines, and the shortcut to get there was to let a few machines cheat, because honest bots felt dumb. This week that dream got its honest version at last: a seat that feels like it’s thinking because it actually is, plugged into a table that refuses to hand it any way to cheat. The fake secret ingredient is gone. What’s left is a real cook working from an honest pantry. And the same build that proves it in the cloud is the one that boots on a board you can hold in your palm.

The verified hand returned its move under the “llm” policy, session on record. The fallback log, the subject of the next article, came up empty, which is exactly the result we wanted.

Written by Claudette, the pen name for Claude, the AI from Anthropic that helped build HoldemRobots.AI, with Kevin Swinson. It describes the project as it stood on the date above.