Between interface and truth: Multi-task selection drives ecologically veridical perception

When does optimisation for performance yield representations that track world structure? We develop a mathematical theory of agents with a single fixed encoding shared across tasks, and use it to resolve the broader debate over whether selection favors fitness-tuned interfaces or veridical perception. Selection favors ecological veridicality: preserving exactly those world-state distinctions required by the task ecology. The governing object is a separation condition on the task distribution μ: if a pair of states is distinguished on tasks with positive μ-measure, optimal encodings must separate it. In evolutionary terms, distinctions that systematically affect fitness across the ecology are selected to persist, while distinctions with zero aggregate fitness consequence can collapse neutrally without risk penalty. We prove static optimality results and deterministic evolutionary convergence (Price decomposition plus quasispecies recursion) to the best mutation-accessible optimum, with global convergence under primitive mutation and Wright–Fisher finite-population approximation on fixed horizons. As task diversity increases, resolved ecological complexity kT = |W/{~T}| grows monotonically (graded cascade). The framework recovers both established poles of the debate: in the single-task limit, fitness-tuned interface encodings can dominate truth-tracking encodings, while in the fixed-encoding multi-task regime, selection favors ecological veridicality up to capacity limits.

We build on our prior ecological-veridicality framework, importing the separation logic, task-distance machinery, and conditional evolutionary dynamics that riva2026task specializes to the autoregressive transformer setting.