Homeostatic Parsimony: Interoception and Relief Gradients as Sufficient Conditions for Artificial Agency

Homeostatic Parsimony is a minimal, fully reproducible reinforcement-learning (RL) ablation study on homeostatic agency: the capacity of an embodied controller to recover internal stability after a critical perturbation. The repository implements a toy micro-environment with two internal variables (energy and social integrity) and a vital shock applied mid-episode, then compares two embodied architectures: (i) a parsimonious agent with interoception (access to internal state ) and a relief-gradient value signal , and (ii) a “more complex” agent that adds global modulation (dynamic learning rate/temperature) plus a panic-like crisis override. Across 20 random seeds, the interoception + relief-gradient agent shows rapid, robust recovery (low recovery time and low AUC of drive), while the globally modulated panic scheme can induce rigidity and slower, unstable recovery. The project provides code, fixed seeds, and scripts to reproduce figures and summary statistics, and frames the results as a conceptual “pressure test” for extreme brain–body decoupling scenarios (“somatic mismatch”) rather than clinical evidence.