Selective Dissipation for Continual Learning: Receiver-Relative Residual Trajectories and Self-Regulated Plasticity

A deployed learner should resolve familiar encounters without unnecessary rewriting, correct familiar errors, and remain able to acquire new relationships. I propose selective dissipation: diagnosis in a history-shaped receiver with durable weights held fixed, followed by a separate proposal and verified plasticity transaction. The hypothesis concerns whether measured residual trajectories help a finite controller allocate useful retained changes, not whether decay deletes knowledge or creates information. Conditional-information results show that deterministic trajectories add no Shannon information beyond their complete generating state; fixed-set inequalities bound measured retention, not generalization or future plasticity. Earlier synthetic studies found a restricted trajectory lead that reversed against stronger history controls, endpoint tradeoffs under persistent writes, and improved expression without stored expert changes. I now construct a trained recurrent receiver with learned proposal gates, bounded protected transactions, and later-task evaluation. Eight frozen methods receive the same 220-event streams in four new polynomial and four shifted-rule cases. Ordered-trajectory pre-feedback mean squared error is 0.40837 versus 0.41427 for a complete-history gate on polynomial streams, but 0.49686 versus 0.49426 under shift. An A-GEM-style protected control performs better on average in both families, at 0.39361 and 0.47014. Permuting trajectory order does not consistently impair performance. State removal and restoration locate a modest retained parameter contribution, while direct recent-context effects are small and prior learning slightly worsens the trajectory arm's later-acquisition endpoint. Thus the proposed operations are now joined in a bounded synthetic construction, but a general advantage of trajectory control, strong carried-state use, and preserved future learning are not demonstrated. Learned update location/timescale, broader architectures, original SAN-Cycle source recovery and biological interpretation remain open. Preprint v1.0, 23 September 2026. The companion contains source, applications, frozen results, mathematical material, figures and bounded reproduction instructions. Negative and tied comparisons are preserved. Automated and AI-assisted checks are not independent scientific peer review. External raw inputs and private correspondence are not redistributed; consult the source-access and rights notices. CC BY 4.0 covers original prose, figures, documentation and results. Original software is provided for review without an added general reuse license; explicitly attributed third-party code retains its own license.