Open data
The whole dataset, as files. No key, no login, no rate limit, no scraping required. CC-BY-4.0 — use it, cite it.
The schema and vocabularies ship alongside the data, and that is deliberate. A triple product published without its definition — density × ion temperature × energy confinement time — would reproduce the exact failure this project exists to prevent, one level up. A number without its denominator is not data.
Files
| File | Contents | Size |
|---|---|---|
| meta.json | Licence, schema version, and how to read the status and denominator fields | 2 kB |
| schema.json | The JSON Schema 2020-12 the whole dataset validates against | 48 kB |
| vocabularies.json | Controlled vocabularies — metrics with numerators and denominators, approaches, fuels, roles, regulatory stages | 62 kB |
| organizations.json | Organisations with roles, sectors, location and provenance | 2.3 MB |
| facilities.json | Devices and plants, each with its measurements inline | 718 kB |
| measurements.json | Every measurement, grouped by status | 504 kB |
| glossary.json | Defined vocabulary terms and what uses them | 417 kB |
| changelog.json | Every correction, retraction and verification, derived from the append-only ledger | 69 kB |
| supply.json | Suppliers and evidenced supply categories | 78 kB |
index.json lists all of them with absolute URLs, so a client can discover the set without parsing this page.
Read this before using a number
- status
- Every measurement carries a status. `achieved` was measured on a shot. `claimed` is reported but not independently established. `planned` is a target and is PERMANENTLY INELIGIBLE for any ranking. `superseded` has been beaten. Treating planned as achieved is the most common error made with data like this.
- denominators
- There is no `gain` metric. Fusion gain is meaningless without stating what was counted in — NIF shot N221204 is a target gain of 1.5 and a wall-plug gain of 0.0077, same shot. Gain exists only as gain-target, gain-capsule, gain-plasma, gain-equivalent, gain-engineering and gain-wallplug, each carrying numerator and denominator in the metric registry.
- independence
- Source.independence is independent | operator | aggregator | circular | unknown. A company reporting on its own device is `operator`, never `independent`. Repetition is not corroboration.
- verification
- verifiedAt on a record means a human opened the source in that run. provenance.verifiedFields lists which specific fields were checked against the archived capture — a blanket verified/unverified flag is coarser than the truth.
- missing
- Absence is a finding, not an oversight. Nothing here is imputed to make a record look complete. If a field is missing, our source did not state it.
Spreadsheets
The same data as CSV, for the tool most people actually use. The
status and isDesignParameter columns travel with it — without them
a reader ranks ITER's 3,000 second design pulse above EAST's 1,066 second achieved record and
gets it exactly backwards.
How to cite
Cite the dataset, and state the date you accessed it — figures change when better evidence appears, and every change is in the changelog.
Plain
FusionBenchmark. Fusion and advanced-nuclear device benchmark. 2026. https://fusionbenchmark.com (accessed [date]).
BibTeX
@misc{fusionbenchmark,
title = {FusionBenchmark: fusion and advanced-nuclear device benchmark},
howpublished = {\url{https://fusionbenchmark.com}},
year = {2026},
note = {Dataset. Accessed [date]. CC BY 4.0}
}Licence
The compilation, classification and provenance annotations are ours, released under CC-BY-4.0. Attribute as FusionBenchmark — https://fusionbenchmark.com.
The compilation, classification and provenance annotations are ours and are CC BY 4.0. Quoted source text remains the property of its publisher and is reproduced as short attributed quotation. Archived third-party pages are NOT redistributed.
Archived copies of third-party pages are not redistributed. What is published is the URL and a SHA-256 of what we fetched, so anyone can re-verify a source without us republishing someone else's copyrighted markup.
Stability
Vocabulary term ids are immutable — a term is never renamed or deleted, only marked deprecated with a pointer to its replacement. Record ids are stable. Values change when we find better evidence, and every change is in the changelog with its reason.
Schema version: https://fusionbenchmark.com/schema/v1/fusionbenchmark.schema.json · generated 2026-08-15