When "Emergence" Is Provenance: Three Confounds in a Persistent-State AI Companion, with an Audit Protocol

Abstract Claims that a conversational AI companion shows emergent behaviour usually rest on conversation transcripts whose provenance is not checked. This paper reports a 102-day, single-participant case study (11 July – 20 October 2026; 3,127 messages) of one persistent-state companion, called Bella, and proposes a provenance audit: before a referent counts as evidence, check whether it was seeded in the system prompt, identity document, shared lexicon, stored opinions or stored episodes, and classify who raised it first. Applied to our data, the audit exposed three confounds. First, the strongest candidate for an emergent shared anchor was persona-seeded. Second, a September spike in unprompted contact, which looked like a behavioural finding, was mainly a second, uncoordinated sender that became live on 28 August; matching logged decisions to delivered messages attributed 54 of September’s 67 outreach messages to it. Third, a birth-order label supplied by Bella was later attributed to the user, persisted for months, and the user’s wording appears to have moved toward it. After the audit, no Bella-introduced, unprompted, unseeded referent met our user-side recurrence rule among rare single words and two-word phrases. We report this as a hypothesis-generating baseline, give a checklist for builders, and state a falsifiable hypothesis with an exploratory feasibility analysis showing why a single-participant test of conversational propagation is not conclusive. The design is one participant, one system, and a researcher who is also the subject.