Paddock Garage

The Problem

Every Number on the Internet About EV Cost Is an Estimate.

Cost-per-mile calculators use national averages for a car nobody owns. Manufacturer range figures are laboratory numbers. Review sites publish figures from vehicles they did not pay for. The result is an entire category of content that reads as authoritative and cannot be checked.

The brief was to build the opposite: a site where every published figure traces back to an invoice or a sensor, where estimates are labelled as estimates, and where the unflattering numbers stay on the page. That constraint is an engineering problem, not an editorial one. It requires the provenance of a value to survive all the way from ingestion into the rendered page.

A number is either measured from a receipt or a sensor, or it is labelled modeled. There is no third category, and the system enforces that rather than trusting the author to remember.

01

Two Sources That Disagree on Purpose.

Vehicle telemetry provides drives, pack-side energy and battery health. The manufacturer Fleet API provides billing-grade charging invoices and live vehicle state. Access tokens refresh themselves; the vehicle is never woken on a schedule, because waking it consumes battery the site itself measures as a cost.

02

The Number Neither Source Publishes.

The invoice reports energy drawn from the charger. The car reports energy that reached the pack. They are not the same, and the gap is real money. Matching them per session across 11 invoices measured 8.82% loss: 422.56 kWh billed, 385.30 kWh stored. Matching is on time proximity at a site, because the two sources stamp different time zones and a date join silently drops sessions across midnight.

03

Aggregate Until It Is Safe to Publish.

Raw drive coordinates resolve a residence and a workplace to within a few metres. They are generalised into city-pair corridors weighted by trip count, and departure timestamps are discarded entirely, because a departure pattern discloses when a property is unoccupied. The site publishes the shape of the driving without publishing the driver.

04

The Build Fails Rather Than Leaks.

A pre-deploy gate encodes the privacy and editorial rules as executable checks: identifier redaction, address and coordinate patterns, earnings figures, dead internal links, page weight ceilings. It is not documentation that asks a human to remember. On its first run against real content it caught a published weekly schedule and a stale asset reference that review had missed.

The reconciliation produced a measurement neither source offers: 8.82% of purchased energy never reaches the battery, which turns a billed rate of 37.8 cents per kWh into an effective 41.5 cents. That figure had previously been withdrawn from the site for lack of evidence. It returned measured.

Alongside it: real charging power curves sampled once a minute, city-level corridor mapping from 61 drives, and a published cost per mile that already includes the loss because it is priced from the invoice rather than the dashboard. Ten pages totalling 211 KB, replacing a single 1,054 KB page whose weight was almost entirely images pasted into the HTML.