On November 5, 2026, Atomic Canyon announced NuclearBench, developed with Idaho National Laboratory. The benchmark evaluates answers to questions drawn from nuclear engineering and regulatory documents. It scores the substance of an answer and its faithfulness to the provided material separately. That distinction is useful whenever an AI tool is asked to answer from a specific document.

Conceptual illustration of a reviewer comparing an answer with a source document; not a benchmark interface.
Conceptual illustration of a reviewer comparing an answer with a source document; not a benchmark interface.

Check the answer and its basis separately

The substance check asks whether the response answers the question and includes the required details. The faithfulness check asks whether the supplied material supports the claims. A response can give the right headline and still attach it to the wrong passage. It can also stick closely to the source while leaving out part of the question.

A fictional example outside the nuclear domain

Suppose a museum’s current visitor guide says the east entrance opens at 10 a.m. on Saturdays and 11 a.m. on Sundays. An older guide says 9 a.m. on Saturdays. Asked about the current weekend hours, an assistant gives the correct Saturday time but cites the older guide and omits Sunday. A reviewer can mark the Saturday fact correct, the cited support wrong, and the answer incomplete. The example illustrates the distinction; it is not a NuclearBench question or scoring result.

Make the review repeatable

  • Write the question and identify which parts the answer must cover.
  • Record the exact document version the assistant received.
  • Mark unsupported claims and missing details separately.
  • Keep notes on disagreements so another reviewer can reproduce the judgment.

Atomic Canyon describes NuclearBench as an evaluation of answers from supplied documents. The launch also distinguishes it from the earlier FermiBench work on finding relevant material. If your own application retrieves documents before answering, record whether it found the right material before judging what it did with that material.

A benchmark result can help decide what to test next. It does not validate a particular deployment or replace expert review for operational decisions.

Check the source

Primary sources

Atomic Canyon’s NuclearBench launch and scoring overview

Atomic Canyon’s November 5 release on the benchmark and retrieval distinction