Language, data cost and sovereignty: a consent-gated advisory for self-builders in Cape Town's informal settlements
Abstract
Most of the housing the world will add in the coming decades will be self-built, without traditionally trained experts, much of it where climate change hits hardest. In Cape Town's informal settlements, much of the guidance to build cooler, drier and more stable dwellings exists, but it rarely reaches the people who build: not in their language, not on their devices, not within their data budget. We describe a multilingual advisory that accepts voice notes and text, designed from those constraints outward. Answers are composed from a knowledge base in which every item carries its source, date, licence and an evidence grade, joined to the dwelling's observed attributes and the settlement's satellite surface-temperature context, and every answer passes a trained community facilitator before it reaches the household. Five rules hold: answers are advisory only and carry citations and evidence grades; official warnings are relayed with their source named, and none are issued, pinned by a test over the relay path's import graph; consent gates every query and its withdrawal propagates through the whole stack; a measured per-language harm rate gates deployment; and no model is trained on household data. Storage, the trace and a community-governed consent ledger sit in a South African cloud region. Speech enters through a commercial recogniser or an open model, behind one interface. The pilot settlement is Afrikaans-speaking; the variety spoken on the Cape Flats, Kaaps, is served by the commercial service only through a call-centre model in beta and appears in no open speech corpus; the scaling settlement speaks isiXhosa, which it does not list but for which a large openly licensed corpus now exists, so the pilot measures which track serves each. We set out the system, four research questions, and a pre-registered evaluation for a pilot with about one hundred households in early 2027, over channels residents already use, facilitator-reviewed first and then direct.