Check a source

Every record on this site points to at least one source: the document it was taken from. This tutorial shows where to find a record’s sources, how much weight each kind can carry, and how to tell when many sources are really one.

Where sources appear

• Benchmark. On /benchmark/, the Source column links the first source for each figure.
• Licensing. On /licensing/, the Source column does the same for each regulatory milestone. The Anchor column gives the docket or licence number, so you can look the step up with the regulator.
• Company pages. The Sources section at the foot of each company page lists every source for that organisation, each followed by its independence label in capitals.
• Device pages. Each figure on a device page lists its sources, each followed by its independence label, and its confidence.

Leaderboards do not show sources directly. Follow a row to its device page to see them.

Independence labels

A source’s independence says whether it could have come from the subject itself. It is the most useful single fact about a source. There are five labels.

• Independent: could not have come from us or from the subject. Examples are a peer-reviewed paper, a regulator’s docket, a company registry, a patent, or a named journalist’s original reporting.
• Operator: the subject’s own publication, such as a company’s website or press release. It is good evidence of what the company says. It is not confirmation that what it says is true.
• Aggregator: a directory or database that may be copying someone else’s material.
• Circular: a source shown to reproduce our own material. It is never cited as evidence.
• Unknown: we could not establish which of the above applies. We use this label rather than guess.

A company describing its own machine is always operator, however authoritative it sounds. /coverage/ shows how many sources carry each of the first three labels. Most of our sources are operator sources, and that page says so.

Archived copies

When we cite a web page, we fetch it, store a copy of what we received, and record a SHA-256 fingerprint of it. A fingerprint is a short code computed from the exact bytes. If a single character of the page changes, the code changes. Checks are made against the stored copy, not the live page, so a page that is later edited or taken down cannot silently change what we reported. When a later check finds that a page has gone or changed, the changelog records it as a source that went dead or drifted.

We do not republish the stored copies, because they are other people’s copyrighted pages. If the live page now says something different from what we recorded, the difference is logged, and the changelog shows when a source went dead or drifted.

Circular sourcing

Repeating a claim does not confirm it. One company press release can appear in a dozen trade outlets within a day, each rewording the others. Directories then copy those. The figure ends up in twenty places and still has one origin: the company. Worse, aggregators copy databases like this one, so a search can return our own summary quoted back at us. That looks like independent confirmation and is worth nothing.

To test a claim yourself:

1. Find the earliest version. Sort what you find by date. If every later article repeats the wording of the first, you have one source.
2. Check who published it. A trade story that quotes only the company’s announcement is still the company speaking.
3. Look for the primary record. For a performance figure, that is a peer-reviewed paper or a conference paper with a shot number. For a licensing step, it is the regulator’s notice or docket. For a funding round, it is a filing with a company registry.
4. Search in the original language. For Japanese, Chinese, Korean, German and French organisations, searching the native-language name often reaches a primary source that English coverage never cites.
5. Watch for this site’s own wording. If a source repeats our phrasing, it may have copied us.

A figure that rests only on operator sources can still be right. It has simply not been confirmed by anyone else, and on this site a performance figure in that position is labelled claimed.

Sources

• Methodology, Coverage, and Terms of use on this site