Learning What to Preserve Without Silencing the Body: Adaptive Organization, Mutable Evaluation, and Reference-Calibrated Regulation

An adaptive system can reduce its error signal by improving its condition or by changing how that condition is measured. We analyze this ambiguity in a two-compartment stochastic regulator with adjustable sensory routing and a mutable binary evaluator. Identical complete baseline observation laws can support opposite effects of a subsequent routing change; a sensory-activity objective can favor muting rather than regulation. For a held, centered two-dimensional Gaussian body measured in a fixed quadratic geometry, two reference modes and two body-reading modes exactly recover total squared deviation from population no-alarm probabilities despite unknown scalar sensitivity and background. For known dimension $n$, $n$ distinct positive body gains suffice and fewer predetermined gains fail universally within this radial Gaussian query family. We derive finite-data enclosures and explicit failure cases for reference bypass, non-Gaussian tails, and directional sensitivity changes. In particular, even knowing an anisotropic sensor metric does not generally recover Euclidean bodily deviation from an arbitrarily dense radial gain sweep. Directional reference calibration followed by a physically implemented equalizer restores the stated recovery contract, or yields a range when the equalizer is imperfect. Fixed-calibration online adaptation and mutable-calibration finite-grid selection are executed as separate studies. A joint-confidence reanalysis of all candidate settings confirms substantially fewer changes than point-estimate selection, exposing the distinction between actual model improvement and available statistical evidence. The contribution is a bounded mechanism, recovery and failure analysis, not new foundations for calibration, a universal anti-tampering system, or a consciousness assay. It makes explicit the physical relations under which a gradually changing organization can continue evaluating its own consequences. Scientific status: Open theoretical/computational preprint; no independent peer review. The finite models and calibration routines are not measurements of subjective experience, selfhood, or organismal evolutionary history. Contributor statement: Hongju Liu is the human author of record. ChatGPT/GPT-6 substantially assisted mathematical research, computational experiments, writing, critical revision, and repository preparation under author direction. Independent peer review, experimental replication and human line-by-line approval are not claimed. Version and rights: The published PDF, Markdown and ZIP are exact frozen v1.1.0 artifacts. An earlier v1.0.0 preparatory draft remains untouched. CC BY 4.0 applies to new original manuscript material to the extent the depositor holds rights; prior works and inherited archives retain their respective terms.