How we source, verify and publish appliance data

What our warehouse holds today

Every share names what it is a share of

Nobody publishes how many units of a given appliance were sold, so the number of machines that could have gone wrong is unknowable. That makes an overall breakdown rate impossible to state honestly, and we do not state one. What can be stated honestly is a share within a corpus we name: a proportion of a specific set of records, over a specific date range, with the size of that set printed beside it. If you see a percentage on this site without those three things next to it, it is a defect — please tell us.

Every share we publish carries all three of:

• the set of records the share is computed over
• the date range those records cover
• the number of records in the set

Why we do not publish fault rates from safety reports

People file a report with the Consumer Product Safety Commission when something frightened them, not when something merely broke. So that corpus is not a picture of what fails most often; it is a picture of what fails dangerously. Publishing a family fault ranking from it would be a real number describing something other than what a reader would take it to mean, which is why we do not do it. We measured the effect rather than assuming it, and the measurements are below.

58% — Reports about fire, smoke, burning or melting

of the 20,581 incident reports in this corpus — the safety skew, quantified

9.3% — Reports we can attach to a specific model family

1,915 of the 20,581 incident reports. The gap is structural, not a parsing problem: we hold certified models, while reports arrive for the whole market

12 of 3,451 — Families with enough reports to say anything

the rest have too few records for any share to be meaningful

What the corpus is genuinely good for is the language people use. Across the narratives, the ice maker is named 1,067 times, the control board 492 and the compressor 279. That tells us which components deserve research and plain-language explanation. It is a reading guide, not a frequency table, and we treat it as one.

Where the data comes from

Each row carries the licence it is used under. That class travels with every record into the page that renders it, so a source cleared only for internal use can never end up cited in public.

• ENERGY STAR certified products

  Held

  Published by U.S. Environmental Protection Agency and Department of Energy · Public domain, open API

  What we take from it:

  Certified per-model specifications: capacity, efficiency, dimensions, configuration, certification dates.

  What it cannot tell you:

  Covers certified models only, and not every category. Gas cooking and microwaves are absent from it entirely.

• Compliance Certification Database

  Held

  Published by U.S. Department of Energy · Public domain, open access

  What we take from it:

  A second certified-specification spine, complementary to the one above.

  What it cannot tell you:

  No electric cooking coverage, so it completes rather than replaces the other spine.

• SaferProducts consumer incident reports

  Held

  Published by U.S. Consumer Product Safety Commission · Public domain, federal, redistributed with the required disclaimer

  What we take from it:

  Narrative signal — what consumers describe going wrong, and which component they name.

  What it cannot tell you:

  A safety-reporting corpus, not a repair-frequency one, and it cannot be attached to most specific models. Detail below.

• Recall notices

  Held

  Published by U.S. Consumer Product Safety Commission · Public domain, federal

  What we take from it:

  Recall notices with the model ranges they apply to.

  What it cannot tell you:

  Our stored copy currently holds 6 notices, because the collector only ever looked at a rolling recent window. Until the historical pull runs, we make no claim about whether a model has been recalled — check the regulator directly.

• Appliance efficiency database (MAEDbS)

  Not yet ingested

  Published by California Energy Commission · Public state database

  What we take from it:

  Intended as the widest public list of models legally sold in California.

  What it cannot tell you:

  Access confirmed; we hold no rows from it yet, so nothing on this site derives from it.

• Federal court records

  Not yet ingested

  Published by Administrative Office of the U.S. Courts, via CourtListener · Public records

  What we take from it:

  Intended for litigation touching specific model lists.

  What it cannot tell you:

  Access confirmed; not ingested, so nothing on this site derives from it yet.

• Manufacturer support documentation

  Not yet ingested

  Published by The manufacturers · Facts only — never republished text

  What we take from it:

  Factual values such as error-code meanings and published specifications.

  What it cannot tell you:

  Facts are not copyrightable but the writing is. We take values and write our own explanation; we never reproduce or reword a manufacturer page.

• Public forums and community threads

  Internal signal only

  Published by Their respective communities · Internal signal only

  What we take from it:

  Helps us decide which topics deserve research effort.

  What it cannot tell you:

  Never quoted, counted or cited on any page. It informs what we look into, nothing more.

• Repair guides published under non-commercial licences

  Not used

  Published by Various · Excluded

  What we take from it:

  None.

  What it cannot tell you:

  A non-commercial licence does not permit use by a commercial site, so these are excluded outright rather than used carefully.

• Third-party articles and blogs

  Not used

  Published by Various · Excluded as a content source

  What we take from it:

  Only to locate where a primary dataset lives, which we then go and read ourselves.

  What it cannot tell you:

  Never a source of facts or wording. Aggregating other people’s prose is not research.

How a record becomes a published fact

1. 1

   Snapshot

   The raw payload is stored first, keyed by source, URL, date and content hash. A fact whose snapshot pointer is missing does not exist as far as our pipeline is concerned.

2. 2

   Extract

   Structured sources are parsed by deterministic code. Language models are used for one job only — mapping free-text narratives onto our controlled vocabulary — and never to invent a value.

3. 3

   Verify

   Every load-bearing number is re-derived from the stored snapshot by code. An assertion that cannot be recomputed from a snapshot is not treated as a fact.

4. 4

   Aggregate

   Variant model numbers are collapsed into families before anything is counted, so a share is never computed against a mixed denominator. Each aggregate records the keyer version it was computed under.

5. 5

   Publish

   Prose may assert only what resolves back to a stored record. A page that cannot clear that check is not published at that level of detail.

What our confidence labels mean

Every fact we store carries a tier, a source and the date of the snapshot it came from. The tier says how directly the fact is evidenced — not how confident we feel about it.

What we do not claim

That open data is ours

Regulator and government datasets are credited to whoever published them, every time. Presenting public data as though it came out of our own workshop would be a false claim about its provenance, and it would also make the number less useful: the version anyone can go and check is worth more than the version only we can see.

A failure rate for any appliance, brand or model

It would need a sales denominator that does not exist publicly. We publish shares within a named corpus instead.

Statistics drawn from our own repair jobs

We are a young company, and the honest description of our service history is that it is too small to carry a statistic. It appears on this site as experience — job notes, photographs, and what our technicians say about specific machines — and never as a percentage. That will change only when the number of jobs behind it can stand up on its own.

That a model has or has not been recalled

Our stored recall corpus is incomplete and we say so above. Recall status should be checked against the regulator, which publishes it directly.

Anything reworded from someone else’s article

Facts are free to use and writing is not. Everything explanatory on this site is written here, from structured data we hold, and the data is credited to its publisher.

Known problems in our own data

• Around 202 groups of model records are stored twice, because two catalogues spell the same model differently and our uniqueness check compares the spellings exactly.

  No published count is inflated — every figure above is counted over distinct models. But where the two copies carry different specifications, which copy is read is not currently decided by anything meaningful.

  Open, tracked, fix is a case-insensitive uniqueness key plus a one-off reconciliation.

• Our stored recall corpus holds 6 notices because the collector was only ever pointed at a rolling recent window.

  We make no recall claims anywhere on the site until the historical pull has run.

  Open, tracked, the collector already supports the full pull.