Keeping the patient's words: design of Aperta Health, a clinician-governed AI documentation-support prototype for multilingual and refugee mental health care in Australia

In the 2021 Census, 5.8 million people in Australia (22.8%) used a language other than English at home, and 872,000 said they spoke English not well or not at all. For people from refugee and migrant backgrounds, the first barrier to mental health care is often simply being understood. Services report that many clinicians do not engage interpreters, and when a consultation does cross languages, the patient's own words, idioms and explanations of distress are easily lost between the interpreter, the clinician and the record. Victoria's Royal Commission into its mental health system recommended developing digital technologies to deliver language services. This paper reports the design of Aperta Health, a web-based prototype that uses generative AI to help clinicians and bicultural workers turn multilingual or interpreter-mediated accounts into a structured, safety-conscious clinical record without discarding the original words. Its core design move is a strict division of labour. Deterministic code scores screening instruments and applies coded crisis rules. Generative AI only extracts, translates and drafts within bounded formats, and is told to write "Not stated in narrative" rather than invent findings. Qualified professionals interpret, diagnose, prescribe and escalate. We describe the design principles and architecture, then assess the prototype against Australian regulation. We conclude that several functions, including AI-drafted diagnostic formulations and public self-assessment, would be unlikely to qualify for the Therapeutic Goods Administration's exemption for clinical decision support software, and that the prototype's data flows do not yet meet the privacy expectations set out in national clinical AI guidance. We state these limits plainly. The prototype runs only on synthetic data, has not been clinically evaluated, and is not supplied for clinical use. The contribution is a reusable design pattern for language-inclusive mental health documentation, and an honest map of what stands between such a prototype and safe use. Stratedge Working Paper WP 2026-02.