H18 — Assumptions
This register makes the project’s scientific and product assumptions visible. An assumption is not a result: experiments, user tests, legal review, or ADRs may confirm, narrow, or replace it.
For: researchers, contributors, reviewers, and agents interpreting or changing TUNES.
Assumptions: accepted ADRs are decisions, not entries to reopen casually; unknowns remain explicit rather than receiving invented values.
Foundational assumptions
| Area | Working assumption | Consequence |
|---|---|---|
| Comparison unit | A directed station-to-station section with duration T is useful to passengers and scientifically interpretable. | Every equivalent-level result carries duration; secondary aggregates do not replace sections without an ADR. |
| Evidence grade | In-service crowd data is survey-grade at best. | No Class 1, engineering-grade, regulatory, or health-risk equivalence. |
| Objective and subjective | Acoustic quantities and passenger perception are related but distinct. | Separate schema objects, samples, map layers and claim language. |
| Privacy | Local record → review → consent → derived-only upload is the default. | Raw PCM stays on device; any exception needs a separate high-bar protocol. |
| Coverage | Dense repeated coverage is more informative than shallow network-wide coverage. | Pilot work prioritises repeatability and overlap. |
| Platform | An iPhone-native recorder is the first scientific collector. | Android, browser and wearables require separate validation. |
| Portability | Generic railway concepts can represent London and later networks. | Local topology and labels are versioned profiles. |
Measurement assumptions and confounders
| Topic | Assumption or limitation | How TUNES handles it |
|---|---|---|
| Phone microphones | Response, gain, clipping and processing vary by model, OS, route, case and obstruction. A quiet-room check is not calibration. | Store provenance and calibration state; flag processing and obstruction; run reference experiments before accuracy claims. |
| GPS | Underground GPS is weak or absent and must not be treated as acoustic truth. | Use route prior, topology, motion and user correction; minimise raw location storage. |
| Occupancy | Passenger load changes context and may change the acoustic scene, but casual occupancy reports are imprecise. | Optional coarse crowding metadata; never treat it as measured occupancy or an acoustic substitute. |
| Train stock | Different vehicles and carriage positions may produce different conditions; passengers may not know the stock or unit. | Keep stock and exact unit optional; use versioned network/service data where justified; do not guess. |
| Weather | Weather matters for exposed or professional measurements but may be irrelevant or unavailable underground. | Record only when relevant and supportable; disclose conditions rather than silently correct. |
| Background conversation | Continuous carriage recording may contain incidental speech and speech can alter measured conditions. | Derived-only upload, optional speech-quality flags, no public waveform or transcription by default. |
| Tunnel acoustics | Enclosure, curves, track and vehicle interact; a phone observation cannot isolate a single cause automatically. | Preserve spectra, context and repeated section evidence; source attribution remains provisional. |
| Placement | Sitting, standing, hand, pocket, bag, carriage end and bogie proximity affect observations. | Collect coarse optional metadata; demote or flag obstructed recordings; do not correct to an imagined standard position. |
| Operating state | Acceleration, braking, dwell, doors, holds and disruptions change the interval. | Align and annotate intervals; exclude or flag non-representative holds and walking. |
| Measurement error | Error is multi-dimensional rather than one universal ±dB value. | Persist acoustic level, frequency content, journey assignment, device calibration, user metadata and subjective-response confidence separately; placement and processing remain explicit drivers/flags. |
| Self-selection | Contributors and optional perception reports are not a population census. | Show sample counts, distributions and collection scope; avoid population claims. |
| Time | Conditions may vary by time and published processing may change. | Version observations, methods, network profiles and releases; report scope and recency. |
Assumptions that must not become silent defaults
- Missing metadata does not mean “typical”.
- Missing sections do not mean “quiet”.
- A corrected boundary does not improve microphone calibration.
- A high quality tier does not authorise raw-audio publication.
- A model-adjusted value does not overwrite the source observation.
- A perceived problem does not prove a physical source or operator fault.
- A low observed level does not establish safety or suitability.
- Open operational data does not imply operator endorsement.
Open validation programme
Future work
The project must publish evidence for device repeatability, reference comparison, clipping onset, placement effects, interruption handling, section-alignment error, tier thresholds, aggregate stability, perception-instrument wording, and map comprehension. Exact retention periods, schema evolution policy, GPS policy, and cross-city comparability also remain unresolved or review-gated.
When evidence contradicts an assumption:
- preserve the original observation and method version;
- publish the experiment or source;
- record a decision in an ADR when project behaviour changes;
- issue a new schema, pipeline, or dataset release where interpretation changes;
- update human summaries without rewriting historical releases.
Related Documents
Decisions · Measurement philosophy · Quality tiers · Recorder · Portability · Project charter · Machine assumptions · Consumer device limits · Calibration and uncertainty · Risk register