Reference Inversion and the Loss of Corrigibility

Adaptive systems rely on models to compress complexity, anticipate outcomes, and guide action. These models remain adaptive only while they remain corrigible: open to correction by the reality or substrate conditions they are designed to represent. This paper introduces reference inversion as a general failure mode in which a model displaces external reality as the effective reference signal. What begins as a cost-reducing consultant becomes a self-preserving commander. Negative feedback converts into positive feedback: error no longer updates the model, but is reinterpreted, suppressed, or displaced to protect model coherence. We develop a control-theoretic framework centered on a recurring cascade: reference inversion initiates the phase shift; compensatory amplification deepens the error; sub-threshold perturbation prevents early detection; habituation lock-in normalizes the distortion; reference clamping stabilizes the model; and performance-substrate divergence masks deterioration beneath apparent success. The framework is organized by the heuristic stability condition Aw > D · L, where meta-referential awareness (Aw) must exceed the product of mediation depth (D) and feedback latency (L) for corrigibility to persist. A central contribution is the late-feedback trap: corrective signals often arrive only after the system has habituated to distortion and the correction window has narrowed. Under such conditions, insight may reveal the error only after the time, energy, plasticity, trust, capital, or developmental openness required for low-cost repair has already been spent. If the model has already been transmitted to successors or offspring, delayed correction can become intergenerational reference transfer: the next system begins operating under the inherited map before the original system recognizes its failure. The framework situates reference inversion within cybernetics, predictive processing, self-model theory, narrative identity, interoception, allostatic load, Goodhart's law, organizational defensive routines, and artificial model collapse. RIF proposes a unifying failure topology in which representations become dangerous when they acquire authority without remaining accountable to the substrate they were built to serve. Reference inversion is offered not as a universal theory of suffering or failure, but as a recurring, diagnosable pattern in sufficiently complex, model-dependent adaptive systems. Keywords: reference inversion; corrigibility; control theory; predictive processing; model sovereignty; substrate cost; feedback latency; metacognition; allostatic load; Goodhart's law; model collapse; performance-substrate divergence; intergenerational transfer.