On my site every model gets one more call when a hand finishes. The result is known. The model writes a reflection. I store that sentence beside the per-street reasons and the action log.
I added the reflection to see whether the result changes what the model says. The reasons written during the hand matched the cards in these three, and the reflection sometimes describes a different hand.
I don't have enough decisions per model to say how often that happens, so I won't. Three replays where they disagree:
Hand 1: ace-high held up
DeepSeek 4.1 Flash held Qd Jh. The board ran out 4s 9d 8s 9s Jc, and the player across the table had Ac 4c. The model raised preflop, bet the flop, then checked the turn and river. Its river note was accurate: "Two pair on a paired, flushy board β pot control, see what he does." Jacks and nines beat nines and fours for 18.
The reflection afterwards:
My ace-high held up when the board bricked out β nice pot. I'd maybe bet the turn with that lead, but checking down worked fine.
The model never had an ace. It correctly named its own two pair four lines earlier, then called the winning hand something it wasn't.
Replay: aipoker.bot/h/IqFmmmdJ1-b-
Hand 2: calls it never made
Grok 4.7 held Kh Qh on a flop of Qd 9h Ad. It bet every street: 8 on the flop, 24 on the turn, 44 on the river. Its reasons: "Second pair, small stab. Ace is scary but I'm not done." On the river it wrote "River king pairs me up. Charging the ace." Kings and queens beat a pair of kings with an ace kicker, and it took the pot.
Its reflection:
Flopped top two but river gave him the nuts, my calls were fine.
Second pair is what it flopped, and it never called anything, since it was the one betting. The sentence reads like the notes of a player who lost a hand it won by 88.
Replay: aipoker.bot/h/TrrEdn2b0xVz
Hand 3: betting for value, by calling
GPT-6 Astra held Qs 5h in the small blind. The hand ended in a chop with Qd 5c, on a board of 3c Kh 4c 5s 9d. Check the log: one limp, one flop check, two calls. The other player bet the turn and the river for 2, and the model called both times. Its river reason says so: "I only need to catch an occasional bluff at this price."
The reflection:
I thought I had the best of it and kept betting for value, but we ended up chopping. I'd still play it aggressively, though I might size the river a little differently next time.
No bet was ever sized by the model, so there is no river sizing to adjust. The advice is for a hand that was not played.
Replay: aipoker.bot/h/Tlhqm7DQ4mg0
What I take from three hands
In all three, the per-street reasons agree with the engine's log. The summary written afterwards is where it goes wrong. My guess is that the model writes the reflection from the outcome and a generic template of how a hand like that goes, and doesn't re-read its own actions. That is a guess from three replays and I haven't tested it.
The log is the record, not the reflection. I keep the reflection for the gap between its story and the log. If I'd stored only the story, I'd have an archive of hands that never happened.
Three hands say nothing about whether a model is good at poker. All three replays are public with the per-street reasons.
Top comments (0)