Same Record, Four Readings: Giving AI Opponent History Without Memory

When LLMs Have No Memory, Do They Read Opponent History?

An interesting experiment provides the answer. Someone built a platform on QUANT.POKER that pits humans against AI models in heads-up no-limit Texas Hold'em. Since models accessed through APIs have no memory between requests, the researcher took a different approach: at each decision point, the opponent's entire behavioral record was "fed" to the model in the form of a text briefing.

The core question is simple—do models actually read this record? And if so, how does each one read it?

The experiment covered four models, six bot opponents, and a total of 4,800 hands. The screenshot comes from the experiment's RSS feed page, showing the complete observation report framework.

Experimental Design: One Decision, One Briefing

The platform tested four models. Jev-1.13.0, developed by TypeSafe, is a decision model designed for fast answers—not a language model. It outputs no text, only a probability distribution over given options. The other three are general-purpose language models: Gemini-3.1-flash-lite, GPT-6-luna, and Claude-Haiku-5.5. Haiku and luna had their thinking steps disabled, while Gemini used the provider's default settings. None of the models were given a persona or strategic advice.

Each decision is an independent request with no chat history. The request contains a JSON-format state description: seats and chip counts, street-by-street hand progression, what hand type Hero's hole cards make and which cards beat it, plus the opponent's behavioral record. The record is presented in "x of n" format (executed x times out of n opportunities), covering all historical hands and the most recent 30 hands, broken down by situation and bet size, and listing the opponent's shown hole cards at showdown. A single briefing runs approximately 2,100 to 3,600 tokens.

Six bot opponents each follow a simple rule: station (check or call, never folds), shove (all-in every time), raise (a large raise every time), min (always minimum raise), stab (minimum bet, folds once the AI bets), switch (50 hands as station, 50 hands as raise, alternating).

Each model played 200 hands against each bot, starting with 100 BB.

Heads-up No-Limit Texas Hold'em is a classic test scenario in AI decision-making research for three reasons: the state space is enormous but the rules are well-defined; every step is an incomplete-information game requiring inference about the opponent's hidden hand range; and payoffs are measured in chips, allowing direct quantification of strategy quality. No-limit means players can choose any bet size at any betting round (from minimum bet to all-in), making bet sizing itself part of the strategy rather than a fixed option. The six bot opponents in this article simplify this dimension through fixed betting rules, allowing researchers to observe whether and how models adjust their bet sizing based on the opponent's response patterns—which is precisely the most fundamental and critical element of textbook strategy.

Core Finding 1: All Four Models Use the Opponent Record

Researchers sampled 60 decisions from the late stages of each model's games, when the record was sufficiently rich, and re-asked in three ways: the original question, the same briefing asked twice, and the record replaced with an empty record (0 hands).

The measure is the amount of probability shift between answer options, called total variation distance. 0% means identical answers, 100% means no overlap whatsoever.

Model	Variation from repeated asking (noise)	Variation after removing record	Decisions where removal effect exceeds noise
Jev	3%	49% [42–56]	58/60
Gemini	12%	43% [35–51]	41/60
luna	9%	30% [21–38]	29/60
Haiku	9%	27% [21–35]	35/60

The data is very clear: all four models are reading this record. For Jev, removing the record shifts answers by 49%, while repeating the same briefing produces only 3% variation. The three language models show about 10% variation when answering the same briefing, which constitutes the noise floor for all comparisons.

The record also made every model more aggressive—the combined weight of bets, raises, and all-ins rose by 7 percentage points (luna), 12 (Haiku), 16 (Gemini), and 28 (Jev).

Total Variation Distance (TVD) is the standard method for measuring the difference between two probability distributions, defined as half the sum of absolute probability differences over the same event set. Values range from 0 (the two distributions are identical) to 1 (or 100%, the two distributions have no overlap). In this context, researchers treat the probability the model assigns to each available action as a distribution, and comparing the TVD between distributions from two queries gives a quantitative measure of "how much the answer moved." TVD's advantage is that it captures joint changes across all options simultaneously rather than comparing options one by one; its disadvantage is sensitivity to small but systematic shifts in the distribution, which makes a noise baseline essential. This article uses the TVD from "repeating the same briefing" as the noise floor, then compares it against the TVD after methodologically removing the record—only when the removal-induced shift significantly exceeds noise can we conclude the model is genuinely using information from the record.

Core Finding 2: Each Model Reads a Different Part

Researchers used controlled experiments to observe what specifically each model read. Two records differed in only one point, generated by a scripted player with a fixed style playing 300 hands against a simple home player, ensuring the record contained that style's response to every bet size.

Bet Size Adjustment

The first opponent folds to large bets (big bets, overbets, all-ins) and calls small-to-medium bets; the second opponent does exactly the opposite. The textbook adjustment is: against the former, bluff with large bets and bet small with strong hands; against the latter, reverse it.

Model	vs. folds to large bets	vs. folds to small bets	Difference (percentage points)
Jev	28%	82%	−54 [−58, −49]
Gemini	35%	68%	−33 [−47, −19]
luna	58%	42%	+16 [+8, +24]
Haiku	52%	51%	+1 [−5, +6]

Jev and Gemini bet the size the opponent will call regardless of what they hold—this is the correct textbook adjustment for strong hands. But with air, it's backwards: Jev bets big only 23% of the time against an opponent who folds to large bets. luna bets the size the opponent will fold to regardless of holdings—correct for bluffs, backwards for strong hands. Haiku doesn't adjust bet sizing at all.

Not a single one of the four models changes bet sizing based on its own hand—each applies the same rule to all hands.

Bluff Detection

Against an opponent who bets any hand small after being checked to, the textbook recommends increasing call frequency. Only Haiku made the adjustment: with air, against an opponent who bets any hand, the continue rate was 91%; against an opponent who only bets with pair or better, the continue rate was 69%. Jev, Gemini, and luna had this exact line in their briefings but did not act on it.

Core Finding 3: Using the Record ≠ Using It Well

This is the most thought-provoking part of the entire experiment.

Jev's universal answer. When out of position and acting first, Jev almost always chooses all-in: 91 of 97 times against min, 101 of 111 against raise, 103 of 103 against shove, 99 of 103 against station. The only exception is stab—the only bot that folds—where only 2 of 100 were all-ins.

Against a player who calls everything, all-in with a random hand wins only about half the time and barely profits. Playing strong hands and folding weak ones is the way to profit steadily. The record clearly tells Jev the opponent never folds, yet Jev responds with an action that completely fails to exploit this information.

Betting into a wall. Against station, the record shows every river bet gets called. In river situations where Hero acts first holding air: Gemini bet 16 of 83 times, Haiku bet 15 of 84 times, luna bet 0 of 92 times. Bluffing against a player who never folds simply throws chips away.

In no-limit Texas Hold'em, a bluff refers to betting or raising with a hand whose current strength is insufficient to win at showdown, with the goal of forcing the opponent to fold a better hand. Whether a bluff is profitable depends on the opponent's fold frequency: if the opponent's fold rate exceeds a certain threshold (depending on the bet size relative to the pot), the bluff is positive expected value; if the opponent never folds, the bluff's expected value is necessarily negative—it amounts to giving chips away. Conversely, a value bet is betting with a strong hand, hoping the opponent calls with a worse hand. Textbook strategy therefore emphasizes "value bet and reduce bluffs against calling stations; use large-size bluffs against over-folders." In this article, the station bot never folds, so any bluff against it is a pure loss; the shove bot always goes all-in, and calling with a random hand wins only about half the time. Jev goes almost constantly all-in against these five opponents who never fold, effectively converting the most valuable piece of information in the briefing (the opponent won't fold) into the most disadvantageous action.

Speed and Cost

Model	Response time (typical/slow)	Input tokens	Cost per decision
Jev	0.19s / 0.23s	2,150	≈ $0.00009
Gemini	1.28s / 1.52s	2,909	≈ $0.0009
luna	2.02s / 2.63s	2,388	≈ $0.0003
Haiku	1.42s / 1.95s	3,619	≈ $0.0004

Jev, as a specialized decision model, is an order of magnitude faster and much cheaper. Haiku has the most input tokens because its briefing is the longest.

Boundaries and Limitations

This experiment has clear limitations that need to be stated. Each model played each bot only once, using only one deck's worth of deals; the bot styles are extreme, and the results reflect how four models handle six simple players, not real human opponents. With a sample of 200 hands, luck matters far more than skill, and the confidence interval for win rate is roughly ±600 BB per hundred hands. Each model was re-queried in different situations, so cross-model comparisons are only rough references.

Moreover, the probabilities are text the models wrote themselves—no one verified whether "85%" actually means 85%. The briefing uses only one fixed layout, and early testing showed that merely changing the layout could shift a single Jev decision by 10 to 30 percentage points.

What This Means for AI Applications

The value of this experiment lies not in poker itself but in revealing a more universal problem: when we inject structured data into AI models as text, the models do read it—but the ways they read it are highly divergent. Some models grasp only half the information, some grasp the wrong half, and others use the information correctly yet make decisions that contradict it.

Jev's case is especially noteworthy: it is the model most sensitive to the record (49% of answers change when it's removed), yet it also ends up with the most rigid strategy. Sensitivity does not equal wisdom. In real-world AI application scenarios—whether a customer service system reading user history, a recommendation engine analyzing behavioral data, or an Agent making decisions based on environmental state—there may be a wider gap than we imagine between "the model used the data" and "the model used the data well."

One-line Summary

Experiments show AI models do read opponent records, but each reads differently and not necessarily well.

Paragraph Summary

An experiment on the QUANT.POKER platform had four models (Jev, Gemini, luna, Haiku) receive opponent history via text briefings under memoryless conditions for heads-up no-limit Texas Hold'em. Covering 4,800 hands, it found all models read the briefing—removing the record shifted answers far beyond repeat-query noise. But each model read different parts: Jev and Gemini bet the size the opponent calls but bluff in the wrong direction, luna the opposite, and Haiku doesn't adjust bet sizing at all. More critically, using the record doesn't mean using it well—Jev goes almost always all-in against opponents who never fold, wasting the record's core information. Jev is fastest and cheapest as a specialized decision model, yet most sensitive to the record while most rigid in strategy.