Environmental Pressure as a Driver of Causal Representation in Learning Systems

We investigate the conditions under which learning agents develop causal rather than merely associative representations of the world. Through eleven controlled computational experiments, we show that hidden-state flip frequency is the primary environmental determinant of representation quality. A sharp phase transition separates low-structure from high-structure equilibria between flip_mean=80 and flip_mean=200, largely independent of penalty severity. This result holds across evolutionary simulations (five-seed replicated) and a neural GRU belief model, both stabilising at causal capacity ||Δz|| ≈ 0.55 under identical high-pressure conditions. Two validation experiments establish the honest limits: an intervention test shows behavioral adaptation (89.8% recovery) without Level 2 causal inference; a baseline ablation separates architecture from training pressure. A standardised CLW Benchmark Suite evaluates causal reasoning across three environments grounded in Pearl's Ladder of Causation. All code is open source under Apache 2.0.