Moneda Bench, made by Moneda
Moneda CiteBench
Whether the sources a system cites say what it claims. The system writes a short research note on a company, with sources, and every linked claim is checked against the page it links to.
First results
Coming
About this benchmark
- Size
- 100 research notes (about 1,500 claims) on companies in the US, Europe and Latin America
- Public split
- Yes, companies rotate each release
- Grading
- An LLM judge checks each linked claim against the fetched source: factual, hallucinated (the source does not say it) or non-verifiable (dead or dynamic link). Two people blind-audit 15% of verdicts.
- Licence
- Items CC BY 4.0, harness MIT
- Submission
- Open harness at release
- Products allowed
- Yes
- What it does not cover
- Whether the note's conclusions are sound, and claims that carry no link at all beyond counting them.
How it will run
Each system gets the same plain retail prompt per company. Claims are split into numeric and descriptive. The judge is checked against people on a held-out set before the rubric is tuned, and we report agreement next to the scores. Below 0.7 agreement the board is labelled experimental and not ranked.
Status: first results coming.
Results
First results coming.
Papers and links
Paper, data and harness will be linked here at release.
Last checked 1 Nov 2026. Something out of date or wrong? Send a correction.