What this dataset holds, and — more usefully — what it does not. The gaps come first, because they are the part you cannot get anywhere else.
A statistics page that shows only totals is marketing. If a number here is unflattering, that is the number most worth publishing: you can only calibrate trust against weaknesses you can actually see. Every gap below states what would close it, because a gap with no route is just a complaint.
Only 9 are independent. A company's own site is evidence of what it says, not of whether it is true. This is the single biggest weakness in the dataset and it is structural: operator sites are what exists publicly, and independent verification means peer-reviewed papers and regulator dockets, one record at a time.
What would close it
Peer-reviewed sources per claim. NIF is the worked example — press said 3.15 MJ, the published paper said 3.1 ± 0.16.
Verified means a human opened the source in that run and confirmed the figure. Everything else is recorded faithfully from its source but has not been independently checked by us.
What would close it
Verification passes. Every one run so far has changed the data.
266 carry field-level verification instead — name, website and summary checked against the archived capture. Classification (sector, approach, fuel, country, founding year) is NOT confirmed by that check, because nothing on a company's own homepage establishes that we classified it correctly.
What would close it
A domain agent reviewing classification against evidence, not more automation.
21 metrics are modelled and empty, including energy confinement time, tritium breeding ratio, capacity factor and LCOE. Triple product has exactly one measurement — W7-X's stellarator record — because the same-plasma condition is hard to satisfy from public sources.
What would close it
Primary-source capture per device. This is the work that would make the benchmark better than Wikipedia rather than equal to it.
Undefined terms are never published in the glossary — a label the site cannot explain is a classification it has not earned. Every term currently attached to a record is defined.
What would close it
Writing them. Domain terms need a domain author, not a template.
Mostly inertial and alternative-concept machines. The IAEA survey gives them a GENERAL INFORMATION table carrying only type, status and ownership — no radius, field or pulse length. Their emptiness is what the source says, not a failure of extraction.
What would close it
Operator publications per device. Absence here is honest, not lazy.
The most important table on this page. 95% of our sources are the operator of the thing they describe.
| Independence | Sources | What it means |
|---|---|---|
| independent | 9 | Peer-reviewed papers, regulators, registries with no stake in the claim |
| operator | 219 | The subject describing itself. Evidence of what it says, not of whether it is true |
| aggregator | 3 | Restating someone else. Repetition is not corroboration |
21 metrics exist in the registry with no measurement behind them. Each is a specific, nameable gap rather than a vague one.
These figures are generated from the dataset at build time, not maintained by hand, so they cannot drift from what the site actually holds. Download it all from open data, or see the changelog for what has changed and why.