R1 — Citizen science: trust, validation, and governance

Decision-oriented research for TUNES Wave 1. Practices transferable from environmental monitoring, participatory noise mapping, public mapping, and biodiversity-style observation projects. Do not invent success rates or claim empirical outcomes TUNES has not measured.

Owns: programme posture for contributor trust, validation, moderation, motivation, fraud resistance, and uncertainty communication — decisions that bind all three repos via ADRs in tunes.

Does not own: acoustic metric definitions (R2), device limits (R3), privacy/legal controller roles (R5), open-release mechanics (R10).

---

1. Decision question

What governance and validation posture should TUNES adopt so passenger contributions remain usable for research and quieter-journey comparison without implying engineering-grade accuracy, operator endorsement, or false precision?

Related open questions (assumptions doc): duplicate/fraud thresholds; ethics review before public beta campaigns; how quality tiers appear on the public map.

---

2. What “citizen science” means here

TUNES is not a replacement for Class 1 operator surveys. It is a repeated, real-world passenger sensing programme: ordinary journeys, consumer devices, optional perception reports, versioned open releases.

That framing matches participatory environmental sensing more than regulatory compliance monitoring:

Repo split: tunes defines rules and releases; tunes-ios enforces collection and local review UX; tunes-web enforces honesty on map/pipeline (coverage, tiers, caveats).

---

3. Transferable practices (without fake metrics)

Drawn from established project patterns (e.g. participatory noise mapping such as NoiseCapture / Noise-Planet protocol ideas; OpenStreetMap-style community mapping; biodiversity observation programmes such as eBird / iNaturalist-style verification culture). Cite methods, not invented conversion rates.

3.1 Contributor trust

• Local-first agency: record → review → consent → upload. Trust starts with showing exactly what leaves the device (prefer derived features).
• Transparent method over brand authority: publish methodology, claim limits, and quality-tier definitions in tunes docs; the map must link to them.
• Identity light: prefer pseudonymous contribution IDs over social-login theatre; reputation should improve data handling, not create a public leaderboard that rewards gaming.
• Role clarity: distinguish casual contributors, campaign volunteers, and research partners. Different campaigns may require different consent and ethics paths (see R5).

3.2 Data validation (layered, not one score)

Borrow the multi-check habit of mature mapping/biodiversity projects: automated filters, peer or expert review for edge cases, and disclosed exclusions.

Recommended validation layers for TUNES:

1. Device/session integrity (ios): clipping, interruption, voice-processing mode, Bluetooth route, obstructed mic, pocket/bag flags — emit quality flags, don’t silently “fix.”
2. Journey plausibility (web pipeline + ios correction UX): impossible speeds, missing section assignment, mixed walking/station segments, branch ambiguity — surface for user correction before upload when possible.
3. Acoustic sanity (pipeline): empty/near-floor runs, pathological spectra, duration too short for a claimed station-to-station section, metric fields incomplete (missing duration T with any Leq).
4. Cross-observation consistency (research, later): overlapping sections and times can support relative calibration research (R4) — not automatic trust of absolute levels.
5. Human review queues (ops): campaign-critical or disputed sections; never required for every casual upload if automated layers + tiers are honest.

Confidence must remain multi-dimensional (acoustic level, frequency content, journey assignment, device calibration, metadata, subjective response) — not collapsed into one unexplained score. Objective metrics and subjective perception stay separate (charter / glossary).

3.3 Calibration vs quiet-room checks

Citizen-science programmes often over-claim “calibration.” TUNES working assumption: quiet-room check ≠ calibration. Quiet-room checks may flag obstruction, odd noise floor, or malfunction. Absolute accuracy belongs to R4’s hierarchy (unknown device → model profile → field offset → external mic → professional reference).

Public communication: never say “calibrated iPhone” without stating what was done and which quality tier (A–E working labels) applies.

3.4 Reputation systems — use carefully

Useful patterns:

• Soft reputation for review completion, correction diligence, and low fraud-flag rates.
• Campaign badges for methodology compliance (e.g. phone-in-hand protocol), not for “most recordings.”

Avoid:

• Public competitive rankings that incentivise fabricated journeys.
• Hiding low-tier data behind a reputation wall without labelling why it is excluded.

3.5 Moderation

Moderation covers data integrity and community conduct, not “censoring loud routes.”

• Auto-hold uploads that fail hard integrity checks.
• Soft-downrank or exclude from passenger-facing aggregates when quality flags fire (still keep in research layers with flags).
• Code of conduct for harassment / misuse of forums; scientific disagreement goes through issue trackers and ADRs, not silent deletion of inconvenient measurements.
• Disclose when campaign data or professional reference runs are mixed into public views.

3.6 Open licensing and governance

• Provisional licences live in tunes governance docs; citizen contributors need plain-language consent that matches licence and retention.
• Methodological disputes resolve via ADRs in tunes, not silent edits of peer research docs.
• Long-term maintenance: versioned releases, changelogs, and reprocessing that creates new derived datasets rather than rewriting history (R10).

3.7 Motivation (design for ordinary commutes)

Transferable motivators from successful volunteer sensing:

• Personal utility first: quieter-journey comparison for the contributor’s own routes.
• Visible impact: coverage maps that show where more samples help, without shaming sparse areas into false colour certainty.
• Low friction: minimise mandatory metadata (R8); protocol tips over long forms.
• Respect for time: offline-first collection; review after the journey.
• Community belonging without requiring constant social posting.

Do not promise health outcomes, operator fixes, or “official” status as motivators.

3.8 Handling disagreement

Expect disagreement between: devices; seated vs standing; crowded vs empty cars; perception vs metrics; passenger data vs professional surveys.

Practice:

• Keep metrics layered (R2); don’t average A/C/peak/octave into one blended claim.
• Publish both medians and spread when aggregating sections; show sample size.
• Treat professional surveys (including FOI methodology references) as reference methods, not as a mandate to silence crowd variance.
• Document exclusions the way professional surveys do (conditions and footnotes) — honesty over completeness theatre.

3.9 Fraud and low-quality submissions

Threats (illustrative categories, not measured rates):

• Fabricated timestamps / impossible journeys
• Pocket recordings labelled as held-phone protocol
• Duplicate spam uploads
• Synthetic audio or replayed files (especially if Voice Memos / imports are allowed as lower tiers)
• Strategic gaming if any public ranking exists

Defences:

• Prefer derived-only uploads by default (harder to launder identifiable raw audio; still need integrity checks on features).
• Session provenance: device model, OS, app version, processing pipeline version, quality flags.
• Idempotent upload and duplicate detection (R10 / architecture).
• Tier demotion for imports and incomplete control (Tier D working label for Voice Memos-style paths — finalize in R3/R4 ADRs).
• Do not punish honest edge cases (delays, held trains, late start) — those are journey-alignment problems (R7), not fraud.

Thresholds for auto-reject vs flag-for-review remain an open experiment / ops decision; document them when chosen.

3.10 Communicating uncertainty

Passenger-facing honesty rules (bind tunes-web):

• Show coverage, sample size, recency, and quality-tier mix — not only a colour.
• Label crowd in-service data as survey-grade at best.
• Separate “quieter on average in our samples” from health risk / regulatory exposure claims.
• Prefer intervals / ranges / “insufficient evidence” states over false precision.
• Colour must not be the only quality indicator (accessibility).

Research-facing honesty:

• Export quality flags and confidence dimensions with every record.
• Never silently bake correction factors into published levels without retaining raw-adjusted provenance (methodology notes).

---

4. Recommended posture (decision draft)

Quality tiers (working labels — finalize with R4):

---

5. What TUNES should not copy blindly

• Biodiversity “research-grade” badges that imply taxonomic certainty — our analogue is acoustic/journey confidence dimensions, not a single badge.
• Mapping karma that rewards edit volume — volume without protocol compliance harms acoustics.
• Closed moderation without public method docs — undermines the open-science half of the charter.

---

6. Repo implications

---

Recommendation

Adopt a trust-through-transparency citizen-science posture: local review and derived-only uploads by default; layered validation with multi-dimensional confidence; soft reputation for diligence not volume; quality tiers A–E visible on map and in open data; fraud handled via provenance and plausibility checks without punishing honest journey edge cases. Resolve still-open thresholds and ethics-review triggers through future ADRs after pilot operations evidence — do not invent numeric success rates to justify launch readiness.

Confidence

Medium–High for qualitative posture (aligned with charter and participatory sensing practice). Low for specific fraud thresholds, reputation weights, and ethics-review timing until legal/ops input.

Depends on experiment/legal/user-test?

Yes. User tests for review/consent friction (R8/R9); legal for controller roles and contributor terms (R5, licences); pilot experiments for duplicate/fraud false-positive rates; optional ethics review before public recruitment campaigns.

Links to related docs

• ../governance/project-charter.md
• ../02-research-map.md
• ../01-assumptions-and-open-questions.md
• H10 glossary — shared terms
• ../acoustic-survey-methodology.md — survey-grade honesty; participatory protocol notes
• 02-acoustic-methodology.md
• 03-consumer-device-limits.md
• ../governance/code-of-conduct-outline.md
• Downstream: R4 calibration/uncertainty; R5 privacy/ethics; R10 open data; map honesty ADRs