Formalizing World-Model Non-Identifiability in Inference

A new paper on arXiv (2605.12255) argues that divergent conclusions from the same observations stem from non-identifiability in inference and learning, not cognitive defects. The authors propose two levels: θ-level non-identifiability, where conclusions diverge under the same world model due to different inference settings; and W-level non-identifiability, where repeated use of an inference setting biases data exposure, causing learned world models to diverge. They introduce an inference profile θ = (R, E, S, D) comprising Reference, Exploration, Stabilization, and Horizon, showing how outputs split even for identical observations.