Try a two-armed bandit. Each choice reveals a reward. Compare your observed choices with simulated choice policies, then change their learning rate and exploration. The experiment illustrates a method; it does not run CORE or query a real LLM.
Which arm pays more, and when the reversal happens, stay hidden until you finish all 24 rounds, so you learn from rewards just like the simulated policy. Each run draws a fresh reward table, shared by you and the simulator, from a random seed (recorded in the CSV for reproducibility), so neither earlier runs nor a page reload reveal the next.
Choose an option to begin.
Complete several trials to see how choice patterns diverge.
The simulator uses a basic Rescorla–Wagner value update and a softmax policy, not the published CORE architecture. Greedy and random baselines provide contrasts. The simulated agent's actions are sampled independently from yours, and both receive the same potential reward for a given arm and round.
For a real study: administer matched tasks to humans and language/decision models, fit the actual CORE implementation to each sequence, evaluate held-out predictive likelihood and behavioral signatures, and report uncertainty. Choice agreement alone cannot identify a cognitive mechanism.