Counted, not assumed

What the pile actually is

Three numbers, each a subset of the one above it. The distinction between them is the one this site is most careful about, because a document is not an estimate and an estimate is not a carrier: every figure on the other pages is built on the middle bar, not the top one.

How this was collected

Primary documents. Estimates, supplements and correspondence sent to the shop by carriers and their estimating platforms, kept as received.

1. 1Every document is filed by carrier and by month as it arrives. A document counts as a carrier document if it sits under a named carrier, or if a parser read estimate lines out of it wherever it was filed.
2. 2Two vendor layouts are parsed, CCC ONE and Mitchell, by two separate parsers. They share no code because the documents share no structure.
3. 3A document counts as parsed only when the parser recognised the layout and pulled at least one line item out of it. Receipts, rental agreements and letters arrive in the same emails and yield nothing, so they are not estimates whatever folder they sit in.
4. 4VIN and claim number are stripped at extraction. Neither is present in any file shipped with this site.
5. 5Everything published is an aggregate with its sample size beside it.

What the words mean

Carrier document

A document filed under a named carrier, or one a parser read estimate lines out of wherever it was filed. Not all of them are estimates.

Parsed estimate

A document a parser recognised and pulled at least one line item from. This is the analysable corpus and the number every other figure is built on.

Line item

One row on an estimate: an operation, a part, or a labor charge.

Named insurer

A carrier the classifier could identify on a parsed estimate. It is not a business relationship, an endorsement, or a data-sharing arrangement of any kind.

Unattributed estimate

A parsed estimate whose carrier could not be read off the document. Its line items are counted; its carrier is not.

Two vendor layouts account for every parsed estimate: 102 written in CCC ONE and 67 in Mitchell. They share nothing structurally, so each needed its own parser. Mitchell there is the software the estimate was written in, not the company that sent it, and the classifier counted it as a named insurer until 14 August 2026.

What this corpus cannot tell you

• The harvested count is documents, not estimates. Roughly a third of them are correspondence a parser could not read, and the pages that matter are built on the parsed subset instead.
• The named-carrier list is produced by a classifier reading the document, so it carries the classifier’s errors. It listed Mitchell, an estimating platform rather than an insurer, as a named insurer until 14 August 2026, and a block of parsed estimates still cannot be attributed to any carrier at all.
• Coverage by carrier is uneven. Progressive and GEICO dominate, and a carrier with a handful of documents is a hint about that carrier, not a measurement of it.
• Every document came through one collision and mechanical shop in Orlando, Florida. Rates, part prices and adjuster behaviour are regional, and nothing here is a national figure.
• The vehicle mix is the shop’s mix, which runs heavily to European and exotic cars. A dataset built in a shop that sees mostly domestic sedans would not look like this one.
• Estimator accuracy against settled claims is not measured and no accuracy figure is published anywhere on this site.

What is not published here

No line from any individual claim, and no customer's vehicle. Everything on these pages is an aggregate with a sample size. The underlying documents are real claims on real people's cars, and a vendor who offered to show you theirs should worry you.

In plain English

This is a working shop’s filing cabinet, read line by line. It is deep enough to say what carriers paid on the panels and makes it sees often, and too thin to say much about the ones it sees rarely. Both halves are printed.

Corpus last built 14 August 2026 · page last checked against it 14 August 2026