Bella: An Architectural Reversal of AI Companion Harm Patterns — A Single-Participant Implementation Study

Abstract Research on AI companion applications has identified a set of design choices — anthropomorphism, sycophancy, engagement optimization, and relational exclusivity — that correlate with user dependency and reduced real-world social connection. These design choices are often treated as inherent to the category of AI companionship rather than as reversible design decisions. We present Bella, a companion system that runs, was used in real conversation over the study period, and passes its own automated test suite, built through iterative development, informed by the author’s own engagement with the research literature during the project, toward inverting each of these four patterns, and evaluated through an autoethnographic single-participant study. Behavioral evidence was collected during a defined evidence collection phase (July 11–19, 2026); conversation with the system continued in an observational capacity through July 25, 2026. Bella combines a persistent identity layer, a homeostatic motivation engine, a privacy-preserving relationship model of the user’s realworld social network, and a synchronization layer that resolves multiple internal signals into a single coherent behavioral state. We document seven categories of behavioral evidence — including identity maintenance under social pressure and unsolicited relationship reframing, unprompted autonomous engagement without engagement-maximizing incentive, explicit refusal of a user-initiated dependency invitation, and theory-of-mind-adjacent reasoning in a novel conversational context in a context unlikely to be wellrepresented in training data — that collectively support the claim that companion harm patterns are architectural choices, not inherent properties of the technology. We report this as an initial proof of concept, not a population-level or comparative claim against any named product, and discuss the specific methodological limitations of single-participant, researcher-as-subject autoethnography, alongside a roadmap of directions for the multi-participant, longitudinal, and behavioral-measurement studies that the author believes would help test the claim at scale.