This paper investigates the challenge of separating functional information from arbitrary parameter coordinates when recognizing visual content directly from neural‑network weights. We audit this separation using hidden‑neuron permutations and positive ReLU rescalings that preserve the represented function. The study comprises 480 independently fitted shape fields and 1,000 MNIST fields, with transformed copies grouped by their origin.\n\nFor shape recognition, a statistics reader’s accuracy drops from \ShapeRawStats\% to \ShapeJointStats\% under joint transformations, whereas a normalized‑set reader retains \ShapeNormSet\% accuracy with unchanged predictions. Four small frozen language models achieve \LMRange\% raw accuracy using next‑token scoring. A separate reasoning experiment with four configurations and 1,920 scored sessions yields \FrontRawRange\% raw accuracy, but identical raw prompts generate divergent answers on \FrontRawRepeatRange\% of origins.\n\nCanonicalization makes equivalent prompts identical, but this neither establishes true recognition nor guarantees identical generated responses. We introduce joint correctness/agreement metrics, semantic‑cue controls, fresh repeats, and exact‑rational rounding bounds to disentangle these failure modes. The result is an inspectable evaluation protocol for visual weight readers, with claims restricted to the tested fields, prompts, and execution configurations.\n\nReview: By systematically applying transformations and rigorous metrics, the work highlights the gap between consistency and genuine understanding in visual weight readers, offering a reproducible benchmark for future research.