The jev model from TypeSafe was tested in the game 2048

Developers put jev, a new model of the System One class from TypeSafe, through a series of tests in the logical game 2048, with the goal of determining the neural network's ability to make optimal decisions in a limited game field. During the experiment, the model showed high efficiency when provided with complete data on the state of the board and available moves, in such conditions jev successfully followed the given game rules, demonstrating logical consistency of actions, however, when trying to analyze the game based solely on the visual representation of the board without an explicit list of possible moves, the model's performance decreased significantly. The tests confirmed that the System One architecture requires structured input data to achieve predictable results, the model demonstrates dependence on contextual cues, which limits its autonomy in tasks with high dynamics of changes, where independent interpretation of visual patterns without auxiliary metadata is required. The analysis results, published in an open repository, highlight the difference between the model's ability to process formalized instructions and its capabilities in tasks with implicit rules, the obtained data allow developers to better adjust prompts for integrating the model into applied systems that require high accuracy of decision-making.