This paper reports a preregistered computational measurement-validation study of whether eight proposed sitesof precision misallocation in a partially observable perception-inference-control loop can be recovered as aclean behavioral factor structure. Agents were simulated in two environments - a delayed control task and aconfounded causal sandbox - and evaluated with a 32-indicator behavioral battery. The validation programcombined three tests: lesion-based signature validation, factor recovery under full latent variation, andadversarial collapse checks under reduced-dimensional generative truths.The framework recovered nontrivial structure. In Experiment 0, all eight single-joint lesions showed diagonaldominance, and the global lesion produced broad behavioral reactivity. In Experiment 1a, the eight-factorstructural candidate improved over the one-factor model on CFI (0.313 to 0.603) and RMSEA (0.202 to0.159), and it satisfied intended-highest and cross-loading checks under a constrained CFA-style candidate. InExperiment 1b, the apparatus did not falsely require eight factors under one-factor truth (Population A), and itrecovered a preregistered J3/J4 collapse within backend-comparable conditions (Population C).The primary preregistered claim, however, was not supported. Experiment 0 failed because J0 retainedsubstantial off-target leakage. Experiment 1a failed because the eight-factor candidate still showed poorabsolute fit, with RMSEA far above the preregistered threshold of 0.08. Experiment 1b failed its clusteredtruth adversarial test, Population B. The most defensible conclusion is not that the theory is wholly false, butthat the current operationalization does not justify a strong claim of clean eight-joint behavioral separability.The contribution is best understood as a computational stress test that clarifies where the theory-measurementpackage recovers structure and where it remains entangled.
Seth Frame (Mon,) studied this question.