This paper formulates AI evaluation as a reduced observability problem. Behavioral measurements such as logit margins, refusal rates, calibration errors, robustness scores, and embedding projections define a reduced behavioral map from intervention space to observable behavior. The paper distinguishes structural behavioral walls, where the behavioral Jacobian loses rank, from behavioral horizons, where differences remain structurally visible but fall below detectability under evaluation covariance. It also gives a rank-lifting criterion for probe design: additional benchmarks improve observable rank only when they vary along previously hidden behavioral fibers. The framework clarifies benchmark redundancy, noise-limited invisibility, and the design of probes for AI model comparison.
Hiroyuki Shioiri (Tue,) studied this question.