Trust · Methodology · London beta
FairPath accuracy
How FairWhere measures free travel-time estimates against live TfL — what counts as organic error, what we exclude, and why Google is not the teacher.
Public snapshot
Methodology first · headline figure pending
Public headline MAPE ships once the organic sample is large enough to avoid founder-testing bias. Until then, definitions and exclusions below are the contract.
Ops charts stay on the admin FairPath dashboard. This page will publish a frozen organic MAPE once the sample is large enough that founder testing cannot dominate the average.
Definitions
MAPE (mean absolute percent error)
For each logged leg: |estimated − actual| / actual × 100. We average those percent errors. The logged estimate is pre-residual so residual fitting cannot chase its own shadow.
Organic sample
Rows from real product traffic in estimate_accuracy_log. Excludes seeder runs, soft-arrival taps, and retired Google teacher labels. Primary public series is organic vs live TfL.
Teacher: live TfL
TfL Journey Planner (plus optional soft arrival taps, tracked separately) is the persisted teacher. Google Deep Search may still appear as optional live UX but is not written as a durable training label.
Why absolute minutes still matter
Fairness modes (especially kindest / p80) and the minutes we show on meet-pair pages care about absolute error, not only shortlist order. A uniform ±10% miss is not free for group balance.
Formula
MAPE = mean( |estimated − actual| / actual ) × 100
Estimated seconds are the pre-residual FairPath figure logged at check time. Actual seconds are the live TfL teacher for that leg.
What organic MAPE excludes
Seeder traffic
Warms cache and priors without users. Included in ops dashboards; excluded from organic MAPE so we do not congratulate ourselves for synthetic load.
Soft arrival taps
Lower-fidelity labels (Method J). Tracked as their own series; not fused into the headline vs-TfL number.
Historical Google teacher rows
Google is retired as a persisted teacher (GOOGLE_AS_TEACHER=false). Kept for audit baselines only.
Why this matters for fairness
Meet-pair pages and the Fairness Index only earn trust if the minutes behind a one-minute gap are roughly right. FairPath exists so free rankings approach live quality without burning a Directions call on every person × place combination — then live TfL verifies the shortlist.
Journey overviews are simplified and are not intended for navigation. Always confirm before you book or travel.
Related trust pages
How FairWhere decides
FairPath, FairWhere Score weights, fairness modes, and partner transparency.
London Journey Fairness Index
Research framing for area-pair journey burden — sibling to the meet-pair content factory.
Meet in the middle vs FairWhere
Why crow-flies midpoint is the wrong fairness question for London groups.
Frequently asked questions
- What is FairPath accuracy?
- FairPath accuracy is how close FairWhere’s free travel-time estimates are to live TfL actuals, measured mainly as organic mean absolute percent error (MAPE) on logged legs.
- What does “organic MAPE” exclude?
- Seeder traffic, soft-arrival taps, and retired Google teacher labels. The headline series is organic product traffic versus live TfL.
- Why isn’t Google the accuracy teacher?
- FairWhere’s moat is a self-calibrating estimator that improves without paying Google at scale. Google Deep Search may still verify a shortlist in the product UI, but it is not persisted as a training label.
- Are published times guaranteed?
- No. FairPath times are planning estimates. Live TfL or your usual journey app remains the check before you book or travel. FairWhere journey overviews are not for navigation.
- Where can I see live operational charts?
- Detailed daily/weekly MAPE charts remain on the admin FairPath dashboard. This public page publishes definitions and, when ready, a frozen headline snapshot — not raw ops telemetry.