TrimIndex

Trim-level scores and buyer briefs for used EVs.

What it is

TrimIndex rates every trim of every battery-electric car from model years 2020 to 2026: 1,134 trims across 25 makes and 82 models. Each trim page shows a score on five dimensions, recalls, complaints, owner quotes, a value rating and a written buyer brief, and readers can re-weight the score in the browser. The inputs are government safety data, EPA range, market pricing and owner sentiment, and pricing and spec facts are stored as claims tied to source documents.

By the numbers

1,134Trim pages, across 25 makes and 82 models

4,537Source-backed fact claims, citing 680 source documents

23 / 24Briefs passed by Claude Sonnet in the writer test, against 8 for the best local model

What it means for your team

This pattern suits any business that publishes numbers: every figure traces to a stored source, AI has a small capped role, and a checker blocks prose whose figures don't match the data. The writing model was chosen by testing candidates on the same cases.

Status

The live site still runs the build from 2026-05-29. A 15-ticket restart was merged into a release branch on 2026-10-03 and has not been deployed. Most live briefs (1,106 of 1,134) come from a template writer, and 654 fail the newer brief audit until they are regenerated. The site has not been submitted to search engines yet, so there is no traffic to report.

AHow it works

Ingest

NHTSA recalls and complaints, EPA range and efficiency data and US sales volumes load into a 19-table database. Pricing pages and owner forum posts come in through Firecrawl.

Provenance

Pricing and spec facts are stored as claims linked to their source documents. A page shows an exact MSRP only when a matching sourced claim exists.

Scoring

Each dimension has a fixed formula, weighted from battery health at 29% down to software at 11%. A local model may adjust a score by at most 8 points, and complaints are counted per 10,000 vehicles on the road.

Buyer briefs

A fact ledger ties every number in a brief to a stored fact before quality, duplicate and policy checks run. In the merged but not yet deployed release, only approved writer versions can publish. On the same 24 vehicles, Claude Sonnet's June sample passed 23 and the local Qwen 3.8 Splash model passed 8, so Sonnet was chosen as the next writer, pending a fresh test against the newer checks.

Release

A static build produces 1,513 pages. Before each deploy, a release gate runs the tests, the type check, a data audit and the copy rules, and database backups are restore-tested.