The Reference-Class Trap in Animal-Computer Interaction: Toward Individual Longitudinal Baselines in Companion-Animal Behavioral Monitoring

Companion-animal behavioral monitoring increasingly treats breed or population norms as the reference against which an individual animal's behavior is judged "normal" or "abnormal." I argue that this reference class is the wrong one. Recent ancestry-inclusive genomics finds that breed explains only about 9% of behavioral variation between individual dogs; an individual can therefore be normal for its breed yet abnormal for itself, and the reverse. I name this mismatch the reference-class trap — the formal core, at the level of the reference model, of what I have elsewhere called the Normative Trap: a monitoring system that compares an individual to a population aggregate systematically misreads the individual it is meant to serve. Building on the ACI premise that animals should be centered as individuals, I propose that the unit of analysis shift from the population reference to the individual longitudinal baseline — the animal as its own control — with deviation-from-self (the Rate of Drift) as the primary signal, periods of behavioral absence treated as informative (Informative Absence) rather than imputed away, and the species/population layer used as an informative prior for the cold-start period rather than as the comparison target. I outline design implications for ACI systems and, as a position paper, an explicit validation agenda that includes the failure modes of individual modeling itself. The contribution is conceptual and methodological; it is offered for discussion, not as a validated system.