AI Needs a Test That Fights Back: eSlams Uses Games to Evaluate Model Decision-Making

One Correct Answer Doesn't Equal a Chain of Good Decisions

In a fixed Q&A test, once the question is answered, the task is over. But in a game, one move is just the beginning of the next question. The opponent responds, new information changes the situation, and what once seemed like a reasonable plan can become a burden. You're forced to adjust your strategy, carrying the consequences of your previous decision into the next step.

This is exactly the problem eSlams aims to explore. Founder Ramakrishnan Lokanathan is building a public arena where people can play against AI models, watch matches between models, and investigate their decision-making processes through an open-source engine at the foundation.

eSlams stands for Electronic Slams—a grand slam tournament for AI models. Here, AI plays chess, cards, negotiates, and collaborates.

Why Games Are a More Authentic Testing Ground

Games make the challenge concrete. In Connect Four, building your own threats is far from enough—ignoring your opponent's threats means losing the whole game. In poker, even a reasonable decision can lose a hand because you can't see all the cards. Cooperative games pose an entirely different question: can players coordinate and move toward a common goal?

The key isn't demanding perfection from models, but understanding whether they can consistently make good decisions, adapt when circumstances change, and recover when plans go wrong.

In his article, Lokanathan points out that the true value of games lies in their nature as dynamic evaluation scenarios. Traditional benchmarks are a static answer sheet—once the model gives its answer, everything is over. Games are different: they are a continuous sequence of decisions, where each step generates new states and new constraints. Models must not only answer correctly but maintain judgment in a constantly shifting environment.

On the limitations of traditional benchmarks: Current mainstream AI evaluation methods are mostly static benchmarks like MMLU and GSM8K, which measure a model's knowledge and reasoning ability through fixed question banks. The core flaw of such tests is their "one-shot" nature—once the model outputs an answer, the task ends. There is no opponent counterplay, no information updates, no strategy adjustment. In other words, they measure a model's single-step accuracy in a closed environment, not the quality of continuous decision-making in open, dynamic situations. Real-world applications (such as autonomous driving, financial trading, and diplomatic negotiation) precisely require agents to continuously revise their judgments across multiple rounds of interaction, creating a significant gap between static scores and actual capability. The value of games as evaluation scenarios lies in their inherent dynamic features—multi-round gameplay, imperfect information, and strategic confrontation—which can expose decision-making flaws that static tests cannot capture.

Arena and Battlefield: Two Ways to Watch

eSlams offers two modes of participation. In the Arena, you can choose a game and face an AI model yourself. In the Battlefield, you can watch recorded model-versus-model matches. You don't need to understand complex benchmark charts to form your own judgment about the match unfolding before you.

At the foundation is eSlams Core, which currently supports 50 games and control environments, covering board games, card games, negotiation, cooperation, and control tasks. Some environments use simplified or adapted rules, all clearly marked in the catalog.

More Than Win or Lose: Auditable Decision Records

A great replay is only the starting point for useful evidence.

Imagine a model loses. Did it choose a bad move? Misjudge its opponent? Or fail to return a usable response? Is the problem with the model itself, or with the software connecting the model to the game? The final score can't answer these questions.

Core records everything that happens, making these questions investigable. The game engine enforces rules and limits what information each player can see. Moves and technical failures are recorded separately. If the software substitutes a move after a model fails, that intervention is logged, and the round is excluded from valid scoring.

Each match can be saved as a portable record that can be verified and replayed. Public replays are stored separately from private audit information. Spectators can follow the match progress, while developers can examine the records behind the results. This design makes the question "why did the model lose" traceable rather than a matter of guesswork.

On the technical significance of auditability and reproducibility: In AI evaluation, "auditable" means that every decision's inputs, outputs, intermediate states, and system interventions are fully recorded, and the records can be independently verified. The portable match record mechanism used by eSlams Core is similar to "deterministic replay" in software engineering—given the same initial state and sequence of operations, the engine can precisely reproduce the entire match. The design of separating public replays from private audit information balances spectator experience with developer debugging needs: spectators see the match progress, while developers see underlying logs such as API calls, timeouts, and format errors. This separation ensures transparency and security in evaluation while transforming the question of "why the model failed" from subjective speculation into a traceable technical investigation, setting a practical standard for reproducibility in the AI evaluation field.

What Games Can and Cannot Do

In his article, Lokanathan explicitly states that games are not a complete test of intelligence, and winning a match doesn't prove a model is ready for high-stakes applications. Meaningful comparisons require repeated trials, balanced conditions, and clear statements about uncertainty. Games supplement—not replace—testing AI in real-world use cases.

This cautious attitude is worth noting. The current AI evaluation landscape is filled with leaderboards and benchmark scores, but high scores often mask a model's true performance in dynamic environments. eSlams' approach shifts evaluation from "how many questions did it answer correctly" to "does the process of making good decisions withstand scrutiny."

On the ecological validity problem in AI evaluation: Ecological validity is a psychometric concept referring to whether experimental results can be generalized to real-world situations. As an AI evaluation tool, games have inherent boundaries to their ecological validity: board and card games have closed rules and clear objectives, while high-stakes applications (such as medical diagnosis and legal consultation) involve ambiguous goals, ethical trade-offs, and irreversible consequences—the two are fundamentally different. Lokanathan's cautious statement points precisely to this limitation—winning a Connect Four or poker match cannot be used to infer that a model has the ability to handle complex decision-making in the open world. Moreover, meaningful comparisons also require controlling randomness, eliminating first-move advantages, conducting a sufficient number of repeated trials, and providing statistical explanations for result uncertainty. These methodological requirements indicate that game evaluation should be viewed as one dimension in an AI capability profile, not the whole picture.

Open-Source Engine and the First Tournament

The eSlams Core engine uses the MIT license and has been released on GitHub and PyPI. Developers can connect their own agents, run experiments, and examine results. The first AI Grand Slams is scheduled for February 8–14, 2027.

Lokanathan himself will speak at AI Engineer New York on October 14, with the topic "When AI Fails: Designing Evaluation Systems for High-Stakes Applications."

The question isn't just whether AI can win, but what actually happens between the first move and the final score. At eSlams, you can watch, you can play, and you can investigate with the open-source engine.

One-line Summary

eSlams builds an AI decision evaluation arena using games, emphasizing that continuous decision-making matters more than single correct answers.

Paragraph Summary

eSlams (Electronic Slams) is a public arena created by Ramakrishnan Lokanathan to evaluate AI models' continuous decision-making through games. The core argument: answering one question correctly in a fixed Q&A test doesn't equal making a chain of good decisions in a dynamic environment. The platform offers Arena (human vs. AI) and Battlefield (model match replays) modes, with the underlying Core engine supporting 50 games and control environments. Core records all moves and technical failures, supports verifiable replays, and separates public replays from private audit information, making "why the model lost" traceable. The founder explicitly states that games cannot replace AI testing in high-stakes scenarios—only supplement it. The MIT-licensed open-source engine has been released, and the first AI Grand Slams is scheduled for February 2027.