Skip to documentation content

Data flow and privacy boundaries

Happy path (default)

This architecture summary groups the required eight-step contribution flow in R5. R5 remains canonical for the recorder’s step-by-step privacy sequence.

  1. User records on device (tunes-ios) — PCM + motion stay local
  2. On-device / local processing extracts derived acoustic features per candidate section
  3. App suggests station-to-station boundaries; user corrects
  4. User trims/deletes; sees exact upload preview
  5. Explicit consent → upload derived package to tunes-web
  6. Pipeline validates, flags, aggregates
  7. Periodic dataset release published under tunes with new version IDs

Trust boundaries

ZoneAllowed dataNotes
Device privateRaw PCM, full traces, draftsUser-controlled deletion
Upload (default)Derived metrics, spectra summaries, section IDs, device model/OS, app+pipeline versions, quality flags, optional subjective answers, consent metadataNo raw waveform
Pipeline internalUpload + server-side reprocessing artefactsAccess-controlled
Public releaseAggregates and/or record-level derived data per licence & anonymisation rulesNo contributor identity; minimise free text
Research-restricted (optional future)May include short excerpts or raw under separate consentNot the default product path

Defaults locked for architecture

  • Prefer derived over raw
  • Nothing leaves device before consent
  • Reprocessing creates a new derived dataset version
  • Objective metrics and subjective responses remain separate fields/layers

Threat themes (for later formal threat model)

  • Incidental speech in raw audio
  • Precise location trajectories
  • Free-text deanonymisation
  • Server logs retaining payloads
  • Malicious fabricated journeys

Recommendation

Ship v1 with derived-only uploads; treat any raw/excerpt path as a separate ADR-gated tier with stronger consent and retention limits.

Confidence: High for default; Medium for research-restricted tier design (needs legal).

Depends on experiment/legal/user-test? Legal + privacy threat model before public beta.

Links: ../research/05-privacy-ethics.md; H14 recorder; [privacy-client-flow.md](../../../../tunes-ios/docs/privacy-client-flow.md); ../decisions/ADR-005-derived-only-default.md.