Cognitive scientist Donald Hoffman’s Interface Theory of Perception (ITP) proposes that natural selection favours perceptual systems shaped by fitness rather than by the world’s true structure, much as the icons of a computer hide the underlying hardware yet enable the user to interact with it. In line with this, Mark, Marion and Hoffman (2010) introduce the Interface Game, a one-dimensional resource-selection task in which competing perceptual strategies are tested across many simulated generations. Their result, that Interface strategies dominate Truth strategies when high resource value does not always mean high fitness, is a central computational support for ITP. Recent computational work has begun to challenge this conclusion. Berke et al. (2022) show that Interface dominance collapses under multiple independent fitness goals, while Charan et al. (2021) show that an intermediate partial-truth strategy outperforms strict Interface under environmental change. Both critiques use hand-coded perceptual strategies, leaving open the question of what happens when perception is instead shaped by evolution. This project addresses that question by replicating the Mark, Marion and Hoffman (2010) Interface Game and extending it along three dimensions (category granularity, fitness multimodality, and resource dimensionality), then building Artificial Life (A-Life) simulations, in which populations of neural agents evolve via neuroevolution. A-Life simulations are run in both 2D and 3D, with agents perceiving the world through an information bottleneck of variable width that forces them to compress their input into a small number of internal signals. The environment includes a fitness-irrelevant distractor as an experimental control. Their internal representations are then analysed using three complementary methods (linear probes, mutual information, and representational similarity analysis). The Hoffman replication confirms the original predictions; however, the Artificial Life simulations show bottlenecked agents that do not preferentially encode fitness-relevant information. Contrary to ITP’s prediction, no preferential encoding emerges at any bottleneck width, and the variable that contributes most to the agent’s fitness is encoded less strongly than the fitness-irrelevant distractor across every condition tested. These findings suggest that the strict Interface result depends on assumptions that may not hold when perceptual mappings are learned rather than hand-coded. Where Berke et al. (2022) showed that Interface dominance collapses when agents must serve multiple goals, this project shows that it can fail even when there is one. The project contributes computational evidence to an interdisciplinary debate spanning artificial intelligence, artificial life, cognitive science, and perception theory.