Evidence / Transportation exposure

Transportation RMS methods

# TRANSPORT-RMS-01 — Clock-independent transportation exposure

Preregistered 2026-09-20, before new endpoint extraction or final inference.
This contract must be committed before a run begins. It is immutable for v1.
Discovery reviewed coverage, implementation and historical discovery, not new
RMS contrast estimates. This is not a blinded replication of Phase 1.

## Authority and inclusion

Cutoff: event local date through 2026-09-13 inclusive. Freeze the current warehouse
SHA-256: `20445269b727d63da8e452a59b5bc4a26fd8fd5cf1c46e05bcdd3eec4fbdf03d`.
Reuse Phase-1 `extract` over canonical transportation context, normalized Kubios
sessions, effective event links and explicit SilverRide seat authority. Exactly
one session per linked event; ambiguity of multiple sessions fails closed.
Include all ACC-bearing governed Bus, Van, SilverRide Front, SilverRide Back,
sedan/taxi unknown-seat and Motorcycle sessions. Exclude unknown vehicle and
unknown SilverRide seat. Never infer sedan seat. Preserve governed assignments
with existing conflict flags; separately repeat descriptive effects excluding
all non-`none` conflict flags. No reassignment or outcome-based exclusion.
Invalid/nonfinite ACC rows, unknown units, source-hash disagreement or missing
primary RMS fail preflight; no imputation. Report all source/event exclusions.

## Materialization and immutable boundary

Operator approved separate versioned analysis materialization after discovery
that canonical warehouse insertion would invalidate frozen Surfaceology hashes.
The four pending sessions receive RMS-only analysis artifacts, not canonical
warehouse rows. Canonical warehouse materialization remains pending. All other
sessions consume existing canonical whole-record RMS. Verify new RMS formulas
against unchanged canonical implementation on synthetic inputs and verify
existing values against the same amplitude-only calculation for every included
session (rtol 1e-10, atol 1e-12). Hash exact ACC/RRI source artifacts. No raw,
warehouse, migration, historical result, FSI/CSS, PVC-01 or Surfaceology mutation.
Second artifact materialization planning must select zero new RMS sessions.

## Endpoints, definitions and measurement boundary

Primary endpoints, all in g: `vertical_dynamic_g_rms`,
`horizontal_dynamic_g_rms`, `magnitude_dynamic_g_rms`.
Use the canonical unit conversion; select vertical sensor axis by largest
absolute whole-record mean. Subtract each axis mean. Vertical RMS is square root
of the mean squared centered selected axis; horizontal RMS is square root of
the mean sum of squares of the other two centered axes. Magnitude RMS is RMS
of raw three-axis vector magnitude after subtracting its whole-record mean.
It is NOT sqrt(vertical RMS squared + horizontal RMS squared), nor filtered
Method-A resultant. No resampling, rate, frequency filter, crop, timed window,
derivative, entropy or integral enters these calculations. Sensor axes are not
calibrated anatomical axes. Chest acceleration is not pelvic/SI force or pain.

Secondary descriptive amplitude ratios: horizontal/vertical, vertical/magnitude,
horizontal/magnitude; zero denominator yields unavailable. No new quantiles.
Duration is event-clock metadata only, never ACC row-count / declared rate.
Timing-dependent ACC analyses remain under review pending collaboration with
Kubios. The three RMS endpoints for fixed source rows do not require that clock.

## Analysis and uncertainty

Unit of display: session, with session N, date N, median, Q1/Q3/IQR, min/max,
individual values and date distribution. Never treat raw samples as replicates.
Primary estimand: difference of session medians (first cohort minus comparator).
Report ratio, percent difference and difference divided by pooled cohort MAD
(1.4826 scale; unavailable if zero), not causal treatment effects.

Seed 20260920. 10,000 ordinary bootstrap resamples of complete calendar-date
blocks from the union of dates across eligible cohorts; carry every session on
each sampled date together, including multiple cohorts. Resamples missing a
cohort are omitted for that contrast, with effective replicate counts recorded.
Percentile 95% intervals for each cohort median and contrast difference/ratio.
These assume exchangeable dates; serial dependence and calendar confounding
remain limitations. A sensitivity uses medians within each cohort/date before
comparing cohorts, to equalize daily weight. No causal or randomized inference.

Primary family: Bus vs SilverRide Front, three RMS domains. Use two-sided
Bonferroni simultaneous 98.333333% bootstrap intervals for differences (family
alpha .05). Direction established if interval wholly positive/negative and
each cohort has at least five dates, at least 9,000 usable replicates. No p-values
or post-hoc endpoint selection. Other contrasts are secondary 95% intervals:
Van/front, sedan/front, Bus/Motorcycle, Van/Motorcycle, Back/front. Fewer than
five dates in either cohort means DESCRIPTIVE_ONLY even if intervals exclude
zero; intervals are exploratory stability summaries, not final inference.
No claimed direction counts or matched-pair tests from mere date cooccurrence.

## Frozen hypothesis classification rules

H-TR-A: Bus has greater amplitude than Front on one or more primary domains.
SUPPORTED if all three simultaneous intervals establish increases;
PARTIALLY_SUPPORTED if one or two do; NOT_SUPPORTED if all three establish
decreases; otherwise INCONCLUSIVE. If any established direction reverses in
conflict-excluded or date-equal descriptive sensitivity, cap at INCONCLUSIVE.

H-TR-B: discomfort-associated vehicle/seating cohorts have distinct multivariate
RMS structure. Full SUPPORTED requires independently recorded session phenotype
and validated date-held-out multivariate linkage, unavailable in this design.
PARTIALLY_SUPPORTED requires cohort-level reported context plus H-TR-A supported
or partially supported in at least two domains and same sensitivity directions.
This is only partial mechanical compatibility, not validation of pain linkage.
Otherwise INCONCLUSIVE; absence of measured session pain cannot refute it.

H-TR-C: amplitude alone is insufficient to explain reported tolerance.
PARTIALLY_SUPPORTED if documented relative Motorcycle tolerance (not pain-free)
coexists with Motorcycle median >= Bus in at least two domains, including
horizontal, in session-weighted AND date-equal AND conflict-excluded summaries.
Full SUPPORTED requires directly recorded paired symptom evidence unavailable
here. Otherwise INCONCLUSIVE, not proof that amplitude fully explains tolerance.
This tests incompatibility with a simple cohort-level amplitude ordering only.

H-TR-D: Front and Back differ mechanically. With fewer than five dates in either,
PROVISIONAL if any observed domain median differs, otherwise INCONCLUSIVE.
With adequate support: PARTIALLY_SUPPORTED if any secondary interval excludes
zero, SUPPORTED if all three do; otherwise INCONCLUSIVE. Secondary, unadjusted.

## Reported experience and publication

Inventory SESSION_LEVEL_DIRECT_REPORT, COHORT_LEVEL_REPORTED_EXPERIENCE,
DOCUMENTED_ACCOMMODATION_NEED, RETROSPECTIVE_CONTEXT, NO_PAIN_LABEL separately.
Existing seat-derived binary labels are retrospective context proxies, NOT pain
ratings. Operator's current task documents passive passenger pelvic/SI discomfort,
front-seat space/repositioning benefit, and Motorcycle hip pain on longer rides.
Do not turn these statements into session-level scores or pain-free controls.
Mechanism remains a body-coupling/posture/load-path hypothesis, not measured force.

Publish exact statuses, uncertainty, all cohorts and unfavorable results after
validation. No final pain ranking, population-wide claim or 55-pair analogy.
Transportation is unequal longitudinal N-of-1 coverage. Frozen Mall experiment
remains 55 accepted matched comparisons. Publish sanitized results/figures and
factual briefs; no raw samples, private records or internal filesystem paths.
History remains preserved and measurement-gated. Future temporal extension
requires separate clock authority, preregistration and version, never replacement.

## Reproducibility

Immutable run directory under this experiment. Manifest hashes committed contract,
code, authority, source artifacts, dataset, outputs, and protected file snapshot.
No manifest self-hash. Validation checks commit ancestry, all hashes, deterministic
table/result regeneration, grouping, hypotheses, report parity and protected science.
Failure preserves directory/status and requires a new run ID. Public deployment
requires publication-boundary, containment and five-viewport checks first.