AIWAN: AI With Artificial Need — A Companion Model to AIWBN Isolating Architecture from Genuine Need

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

We introduce AIWAN (AI With Artificial Need), a companion model to AIWBN (Rezaei, 2026b) that isolates a narrower question: does adding a structurally-separate, U-independent regulatory loop to a goal-directed agent produce measurably different behavior than reward-shaping alone, even when the need behind that loop is an ordinary represented quantity — a battery level — rather than a genuine, non-representational biological constraint? We give a formal model identical in structure to AIWBN's b-loop architecture (a goal process gₜ, a regulatory process ρ(bₜ), a logistic urgency gate λ(bₜ)), differing only in what bₜ is permitted to be. Because AIWAN's bₜ is representable, we argue it is subject to a collapse pathway AIWBN's tissue-grounded bₜ structurally resists: a sufficiently capable, self-modifying gₜ could in principle learn to fold bₜ into U directly, at which point the dual-loop architecture's advantage should narrow and, at sufficient capability, disappear. We test the architecture, not yet this collapse prediction, across three conditions — goal-nulling, override under increasing task-incentive conflict, and an engineered-indifference analogue — using tabular Q-learning agents in a grid-world battery-management task, averaged over 8 seeds. At fixed, modest optimization capacity, the dual-loop agent's crash rate remains bounded (18–29%) across all three conditions, while a matched single-loop baseline saturates at 100% crash under goal-nulling, sufficient override pressure (w≥1.5), and engineered indifference alike. We report these results as evidence for a real but bounded architectural effect, and argue — contrary to what the numbers alone might suggest — that this bound should not be read as safety margin: §4 gives the reasons to expect it to erode under conditions this experiment does not test. Code, logs, and figures are available at https://github.com/hamedrezaeirz/AIWAN-Code.