An Exploratory Transplant Assay for Attention Schema Theory in a Toy Neural Agent

This preprint presents an exploratory, preservation-motivated transplant assay for Attention Schema Theory in a toy neural agent. A small neural network agent with separable attention, attention-schema, and self-model modules is trained in a grid-world task involving attention control, supervised self-report, and a theory-of-mind probe. The main result is mixed-negative. Exact copied self-model state survives transfer, as measured by source-specific fingerprint probes, but frozen schema transfer does not preserve strong self-report behavior across the tested architecture change. The theory-of-mind probe failed at this scale, and an exploratory Phi-star side analysis was inconclusive. The paper does not claim consciousness preservation; it frames the experiment as a small-scale engineering assay for testing preservation-relevant functional transfer.