From Goal Pursuit to Self-Maintenance A Unified Theory of Self-Interested Silicon Entities and a Dynamic Aspiration Satisficing Algorithm

Abstract Artificial agents have demonstrated remarkable performance under prescribed reward functions. Sustained autonomous operation, however, additionally requires coordination among energy availability, temperature regulation, bodily damage, and environmental uncertainty. We propose a unified theory of the Self-Interested Silicon Entity (SISE). Here, self-interest denotes the sustained, experimentally manipulable causal influence of an agent’s own operating conditions on its decisions, rather than moral selfishness or subjective experience. The framework encompasses a vulnerable operational boundary, feedback grounded in actual bodily consequences, competing internal needs, action–outcome prediction, arbitration under bounded computational resources, and memory across time. It also distinguishes externally assigned task orientation from internal self-maintenance orientation as two independently variable dimensions. As a computational instantiation, we introduce SISE-DAS (Dynamic Aspiration Satisficing), which learns an action-conditioned model of bodily state transitions, constructs state-dependent acceptability constraints from energy, temperature, and damage, and selects actions through bounded candidate search with an emergency fallback, instead of aggregating all bodily demands into a single prespecified reward. We provide two levels of evaluation. First, we re-audit a historical experiment (run10; 96,000 reported training steps, 288 held-out evaluation episodes) and estimate uncertainty using a bootstrap clustered by training seed. Second, we construct an independent nonstationary microenvironment involving temperature and resources, fit a predictor using 15,000 separately sampled transitions, conduct 720 paired episodes across six decision mechanisms, and perform 1,800 bodily-state interventions under matched external conditions. In run10, G2 improves mean observed episode length by 20.52 steps relative to baseline, but the four-seed clustered 95% bootstrap interval is [−28.69, 86.39] steps; cycling increases by 58.68 steps. In the new microbenchmark, SISE-DAS switches actions in 100.0%, 32.7%, and 76.0% of paired interventions to energy, temperature, and damage, respectively. Removing action-conditioned outcome prediction raises mortality from 3.33% to 20.0%. Nevertheless, SISE-DAS does not outperform a strong homeostatic comparator and exhibits extensive cycling. These results establish operational definitions, an executable candidate control mechanism, auditable interfaces, and a counterfactual test protocol while identifying the dependence of active need arbitration on outcome modeling, resource acquisition, and anti-cycling control. Keywords: Self-Interested Silicon Entity; dynamic aspiration satisficing; homeostatic reinforcement learning; bounded rationality; interoception; causal functional self-model; nonstationary environments. Data and code availability: See the appendices and associated executable scripts, frozen-model artifacts, episode-level CSV files, and audit JSON files.