How we verify · every number, every time
Nothing on this site is written by a model and nothing is ranked overall. A value appears only when the exact sentence it came from was found on the page we fetched, on a date we show, from a source we grade. When sources disagree, we show both. When nothing is published, we say so and when we checked.
- 1Fetch
The official page or PDF, politely, one request a second, honouring robots.txt.
- 2Snapshot
The page is stored with a hash so we can prove what it said on that date.
- 3Extract
A model proposes values, but only from a fixed list of attributes, with the quote it read each from.
- 4Verify quote
The quote must appear word for word in the fetched text. No match, no value.
- 5Normalise
Units to metres, kilograms, metres per second, minutes, USD at that day's rate. Status words to our vocabulary.
- 6Compare
Against the current accepted value. Unchanged values only refresh their date.
- 7Score
Source tier × quote match × corroboration × recency, on a fixed formula.
- 8Review
Anything sensitive, out of range, conflicting or below threshold waits for a person.
- 9Publish
Accepted values go live with their source, date and mark, and the change is logged.
A price, availability or status needs a T1 source or two independent sources before it shows as verified.
Any price change over 25%, any status going backwards, any height or weight change over 10% is reviewed by a person first.
Height 0.2 to 2.5 m, weight 0.5 to 300 kg, speed 0 to 5 m/s, price 50 to 2,000,000 USD. Outside these, never automatic.
Manufacturers can send us a better source. They cannot ask us to change a value without one.
A T1 source, or two independent sources that agree. Shown with the date we read it.
One lower-tier source, usually a reseller or press release. We are looking for a second source.
Nothing published that meets our bar. We say how many sources we checked and when. Absence is recorded, not hidden.
Two acceptable sources give different values. Both are shown until one is resolved. Neither counts in comparisons.
Every value has a report control. Reports go to the same review queue as pipeline conflicts, are answered within seven days, and any change is logged publicly.
Data is CC BY 4.0. Pipeline code is MIT. We use press-kit images only where the kit allows; otherwise we draw our own figure. Names and marks belong to their owners.