Learning to Feel: Perception-Action Learning, Phase-Compatible Recruitment, and Conscious Sensory Access in Self-Aware Networks

A physical signal can affect neural tissue without becoming a consciously accessible sensation. I propose that perception-action learning can change this relationship by altering which learned relations a sensory event recruits and how those relations influence receiving populations. In Self-Aware Networks (SAN), retained learning and present coordination jointly shape a recurrent sensory construction that guides further inference and action. The proposed contribution is a testable connection among established operations, not their invention or a completed causal demonstration assembled from separate studies. The account distinguishes stimulus magnitude, learned significance, affective value, neural activity, receiver-relative influence and experienced intensity. It predicts acquired changes in consequential recruitment and asks whether a specified timing-dependent part contributes to conscious access. Conditional mathematics separates retuning from gain and identifies ambiguities in pooled neural measurements and sensory reports. Five constructed applications implement design checks, receiver-level inference, learned temporal selection, reciprocal statistical reconstruction and joint receiver/behavior identification. In the joint test, a receiver observation separates constrained explanations that predict identical behavior, while preserving a nonoscillatory implementation with equivalent predictions. Strong alternatives delimit their results: informative replay reproduces active learning; a nonoscillatory receiver performs at least as well as the phase-feature learner; and direct inference reproduces recurrent reconstruction. Targeted observation improves consequence prediction, not reconstruction of every represented variable. A prospective three-arm experiment combines matched tactile probes, receiver-specific intervention and complementary access measures. State-resolved findings motivate separate measurements of encoding, occupancy and receiving influence; the added analysis constrains their interpretation without claiming new biological data. Sensory discrimination, felt magnitude and decision strategy are not interchangeable outcomes. The broader SAN claim that an appropriately organized self-world learning process constitutes experience remains explicit and requires further explanation and evidence. No biological experiment or conscious artificial system is reported. Reviewed preprint version 1.1. The Hesse/Tsao comparison adds state-resolved representation and receiver-use constraints. The release clarifies prospective testing, nested randomization and post-treatment selection, preserves negative comparisons, and corrects version references and artifact-availability labels. No new biological data or independent peer review is claimed. Version 1.0 remains accessible unchanged. The selected companion contains updated sources, figures and small arithmetic checks. Existing application results remain in the preserved v1.0 companion; Learning to Feel also includes its previously public selected code/results. Publisher-derived MATLAB code and full private replay environments are not redistributed. CC BY 4.0 covers original prose, figures, documentation and results; original software is supplied for inspection without an added general software reuse license.