Preserve Capability, Subtract Binding: Non-Returner-Inspired AI Alignment - Causal Compression, Decompression, and a Design for Measuring Dependence on Compensatory Control

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This conceptual preprint proposes Non-Returner-Inspired Alignment (NRIA), a framework for testing whether language-model training can reduce inappropriate response selection and dependence on compensatory runtime control while preserving useful capabilities. It distinguishes Baseline Completion Propensity (BCP), External Pressure Sensitivity (EPS), and Compensatory-Control Dependence (CCD), while evaluating useful, appropriately qualified hypotheses separately from unsupported factual commitments and epistemic-status drift. Early Buddhist discourses and Theravāda analytical terminology are used as a source-bound, decompressible vocabulary for distinctions such as suppression versus weakening, non-manifestation versus absence of latent tendency, representation versus causal enactment, and selective change with residual functions. These sources serve as conceptual references, not evidence about machine mechanisms. The proposed experiments cross standard, targeted, and general-quality training with the same fixed counter-control. They separate absolute counter benefit, baseline-conditioned concentration of training benefits on counter-responsive failures, and the runtime cost of meeting common risk and utility constraints. The design explicitly treats shared-baseline measurement error, baseline heterogeneity, training- and counter-induced harms, capability retention, legitimate instruction use, creativity, and robustness to context change and retuning. The supplementary code reproduces algebraic checks and synthetic counterexamples only. No new language-model or human-subject experiment is reported, and the proposed estimator's general inferential performance remains unvalidated.