Structure Without Generation

A small, fully interpretable, gradient-free, non-generative learning system recovers linguistic organization — frequency, class structure, grammar adjacency, distributional similarity, and affective markup — from symbol streams by watching alone, with no generative objective and no backpropagation. A companion experiment shows the same instrument causally triages its own map: a planted causal chain, invisible to passive watching, is recovered under actuation, and a hidden-driver phantom — the confounder that correlation cannot distinguish from cause — is convicted wherever the map can name it. Positioned as an exhibit in the world-models debate (auto-regressive training vs. joint-embedding predictive architectures): generative systems are fluent but unaccountable; this system is accountable but not fluent. Every claim carries its measured boundary.