Work continues with no network. Nothing is lost, and nothing skips the quality control every other response passes through.
One screen: the assignments allocated to that person, which are in progress, what is due, and a single action to start the next interview. It is the surface a field team uses on an ordinary working day, and it is built to be usable when the connection is not.
What this shows, and what it does not. A real field surface on synthetic data. It is evidence that assignment and offline collection exist — not a measurement of coverage, of device performance, or of any deployment.
An interview collected in a village and never delivered is worse than one never collected: the respondent gave their time and the record shows nothing. Each behaviour below exists because that is the failure it prevents.
The hardest network failure is not a connection that drops — it is one that carries the interview to the server and loses the confirmation on the way back. The device cannot tell that apart from a failure, so it keeps its copy and tries again. The server recognises an interview it already holds and stores it once.
If a respondent declines to speak, or the room is too loud to hear anything, the interview is saved and the audio is kept. It is marked as inaudible rather than discarded, because a refusal to answer is a finding and the recording is the proof of what happened.
When storage runs low the device says how many more interviews it can hold, and when it is full it refuses to start another one. It never silently fails to save what has already been collected.
A closed app, a flat battery or a household that asks to continue tomorrow does not cost the answers already given. The partial interview is held on the device and picked up from the question it reached.
Recording someone's voice does not depend on which language they are speaking, so a respondent can answer aloud in any of the 57 African languages this platform registers, and that recording is kept as the evidence of what they said. This is the part of "no African voice left behind" that is simply true.
Capturing speech is not understanding it, and the registry keeps the two apart on purpose. Automatic transcription reaches 15 of those languages, on the published list of the speech provider we route to — and not one of the 57 has been accuracy-validated by VoiceInsights against African field audio. Where a language has no transcription, the platform says so rather than producing a confident-looking transcript nobody checked. The recording is still collected, still stored and still reviewable by a person who speaks the language.