R9 — Subjective passenger experience (perception survey)

Status: Wave 1 research recommendation
Instrument definition / claim language: tunes
Survey UX timing and presentation: tunes-ios
Aggregation and map layers: tunes-web (must not blend unexplained scores)

Problem

TUNES measures acoustic conditions and passenger experience. Those are related but not the same. A section can be energetically loud yet tolerable, or moderately loud yet sharp, screechy, or conversation-destroying. Parents and sensory-sensitive travellers often care about character and coping, not only a single level.

ADR-010 resolves timing as after the journey. ADR-011 governs discomfort and “risk” wording so claims do not outrun evidence.

Working assumption #5: objective ≠ subjective in the data model; no unexplained blended score (01-assumptions-and-open-questions.md).

What this survey is — and is not

Is	Is not
Optional structured perception report tied to confirmed sections or a whole session	A star-rating popularity contest for lines
Separate layer beside derived acoustic features	A substitute for Leq / spectra / events
Useful for research on annoyance, speech interference, child suitability	A clinical hearing or health diagnosis tool
Versioned instrument with stable item IDs	Free-text venting as the primary measure
Eligible for its own confidence / coverage display	Proof that a route is “safe” or “unsafe”

Passenger reports are first-person experience under stated conditions (device placement, crowding, headphones, etc.), not votes to rank operators. TUNES remains independent and not TfL-endorsed.

Timing: after, not during (default)

Why after is preferred

Recording UX must stay eyes-free and safe on moving trains.
Mid-journey questionnaires increase handling noise and distraction.
Perception of a section often crystallises after the interval ends (contrast effects are real in annoyance research literature; TUNES should not claim a specific bias magnitude without study).
Alignment confirmation (R7) should precede section-linked ratings so users know what they are rating.

Allowed exceptions

Mode	When	Constraint
Session-end survey (default)	After stop + alignment review	Whole-journey items and/or per-section list
Per-section quick mark (experimental)	Optional one-tap “notable discomfort” while riding	Must not require reading scales mid-tunnel; campaign/experiment only
During full Likert battery	Not for v1 casual mode	Rejected as default

Decision lean: ship post-journey optional survey; keep mid-ride interaction to non-survey status (recording health only) unless a named experiment says otherwise.

Instrument design (structured, not popularity)

Design rules

Short core, optional depth — a 30–60 second core; expand for campaigns.
Ordinal scales with labelled anchors — not stars-only, not binary likes.
Section reference clarity — “this journey” vs “section X→Y” must be explicit.
Allow skip / prefer not to say on every item.
No forced ranking of lines against each other in the contributor UI.
Separate items for separate constructs — do not average discomfort + child suitability + avoid-future into one public “score” without a published model (and even then, prefer showing components).
Version the instrument (perception_instrument_id + version); never silently rescale historical answers.

Recommended core constructs (v1 candidate)

Exact wording to be finalised with accessibility advisors and user tests; constructs below are the decision targets:

Construct	Intent	Notes
Overall discomfort	Global affective load	Primary optional item
Perceived loudness	Subjective level	Distinct from measured LAeq
Sharpness / piercing quality	High-frequency / tonal bite	Complements spectra; not a lab sharpness claim unless method cited
Rumbling / low-frequency heaviness	LF character	Optional core or depth
Screech / squeal noticeability	Event character	May be section-linked
Speech interference	Talk / announcement difficulty	High passenger value
Child suitability (felt)	Caregiver planning signal	Clearly labelled as felt suitability, not safety certification
Avoid in future	Behavioural intention	Not a popularity downvote of the operator

Depth / campaign items (optional)

Painfulness / ear-cover urge
Vibration felt through body/seat
Intermittency / startle
Fatigue / after-effect
Difficulty hearing announcements (vs conversation)
Headphones-on vs headphones-off context (link R8)

Explicitly out of core casual survey

Medical symptom checklists
Diagnoses (tinnitus severity scales, etc.) without ethics pathway
Open political complaints as structured data
“Rate TfL” or operator NPS-style items

Objective vs subjective in the model

Contribution
├── objective_acoustic   (derived features, tiers, quality flags)
├── journey_alignment    (sections, durations, alignment confidence)
├── passenger_metadata   (placement, carriage, crowding, …)
└── perception_report?   (instrument version, items, scope: session|section[])

Rules:

Public map may show side-by-side layers (e.g. measured level distribution vs reported discomfort) with independent sample sizes.
Forbidden: one unexplained colour that mixes dB and survey.
Allowed later: published, cited composite indices with formula, inputs, and limitations — as a third derived layer, versioned, never overwriting raw items or acoustic features.
Claim language: prefer “reported discomfort”, “reported speech interference”, “felt child suitability”; avoid “acoustic risk”, “harm”, or “unsafe” unless ADR-011 is replaced by a decision supported by new evidence.

Attachment scope

Scope	Use
Session-level	Default for short surveys; lowest burden
Per confirmed section	Better science; more taps — offer for multi-section trips or campaigns
Event-linked	Optional mark near detected squeal/peak — experimental

If the user rates only session-level, do not fabricate per-section perception by copying the session score into every section.

Popularity vote vs perception measurement

Popularity pattern	Perception pattern (TUNES)
Stars, likes, upvotes	Labelled ordinal constructs
Rank routes to crown a winner	Describe experience under conditions
Incentivise extreme scores for visibility	No gamified “harshest reviewer” rewards
Aggregate as mean stars only	Report distributions, n, instrument version
Ignore context	Join to metadata + acoustics for analysis

Community mechanisms (challenges, coverage) must not reward extreme subjective scores (research-plan.md contributor community caution).

Privacy and ethics

Perception answers can be sensitive (child suitability, pain, avoidance). Treat as personal data in the privacy model (R5).
Default upload still after consent with field-level preview.
Prefer enums over free text; if free text exists, extra retention/moderation rules.
Children as subjects of rating (“suitable for a child”) ≠ children as contributors; under-age contribution needs separate rules.
Ethics review may be required before campaigns that recruit clinical or vulnerable cohorts — flag for partnership / legal track (R11), not blocked for a minimal optional commute survey of adults.

UX ownership (`tunes-ios`)

Flow placement in the existing privacy client sequence:

Record locally
Process / infer sections
User corrects alignment
Optional: perception survey (core chips)
Optional metadata chips (R8) — order may swap with (4); test which abandons less
Upload preview → consent → derived upload (including perception items if selected)

Accessibility: Dynamic Type, VoiceOver labels, no colour-only scale meaning, one-handed taps.

Aggregation honesty (`tunes-web`)

Display when perception layer is on:

Instrument version
n reports (session vs section scope counted honestly)
Distribution, not only mean
Filter by time-of-day / crowding metadata when available
Clear separation from acoustic layers
Caveat: self-selected contributors; not a population census

Do not imply operator endorsement of findings.

Current disposition (ADR-010/011)

Default timing = after journey, post-alignment; not during.
Survey optional for casual contributions; campaigns may require it.
Separate perception object in schema; never blend into acoustic metric records.
Core constructs as in the table above; final copy via user test + advisor review.
No popularity UI; no operator NPS.
Claim language ADR must vet “risk”, “harm”, “safe for children” before public map copy.
Open Q #10 → after.

Open questions

Session-only vs prompting per-section on long trips (usability vs resolution).
Scale length (5 vs 7) — choose via test, don’t invent psychometrics here.
Whether “avoid in future” should be section-scoped only (more actionable).
Need for ethics approval before public beta perception prompts (legal/stakeholder).

Recommendation

Ship an optional, post-journey structured perception survey in tunes-ios, completed after section alignment, with a short core (discomfort, perceived loudness, sharpness, speech interference, felt child suitability, avoid-in-future) and campaign depth items. Keep perception in a separate model object and map layer from objective acoustics; never reduce TUNES to star ratings or an unexplained blended score. Treat wording as experience reports, not health or safety certification, and version the instrument for reproducibility.

Confidence: High for after-vs-during and objective/subjective separation; Medium for exact core item list and scale anchors until user testing and claim-language ADR; Low for any population representativeness claim.

Depends on experiment/legal/user-test? Yes — usability tests for length and timing; accessibility review of wording; ADR-011 review if claim scope changes; legal/ethics check before clinical or child-focused campaigns; not required to draft the schema separation itself.