Approving Feynman, on one page.
Feynman critiques research a person brings to it: a hypothesis, a manuscript, a proof or a plan. It finds where the reasoning could fail, shows the published evidence it relied on, and names the test that would settle it. The researcher stays responsible for the work.
How it fits a university AI policy
- The researcher stays accountable
- Feynman critiques; it doesn’t write papers, certify proofs, replace peer review or make clinical decisions. Every conclusion is the researcher’s.
- Use is easy to disclose
- Every study gives a run ID, the model it ran on, and a paragraph for the methods section, with BibTeX for the review and for every source it retrieved.
- Unpublished work stays private
- Studies are private by default. Feynman doesn’t use a researcher’s work to train a general or cross-member model, and nothing is shared between members unless the researcher chooses to publish it.
- Its output can be checked
- Each citation is checked against the records the search actually returned, and a citation to anything it didn’t find is removed. Sources link to their DOI or record.
- The researcher keeps control
- Claims are shown for correction before anything runs. Export or delete everything at any time; exports stay with the researcher after cancelling.
Where a researcher’s data goes
Text and files a researcher submits, with relevant private history and retrieved evidence, are sent to Anthropic’s API to reconstruct and critique the work. Model processing may take place outside the UK. The provider’s handling is governed by its API agreement and retention settings, and no zero-retention promise is made. Searches go to public scholarly indexes: Crossref, arXiv and Europe PMC. The privacy notice has the detail.
What isn’t in place yet
- UK or EU data residency. Not offered. Processing may take place outside the UK.
- A data processing agreement. Not available yet. Institutional terms need review by counsel before they are offered.
- Single sign-on. Not offered.
- Security certification. None is claimed.
- Independent evaluation. No one outside the company has yet assessed the quality of its reports. Until they have, Feynman makes no claim to be better than one strong model asked the same question.
A paragraph for a methods section
Each study fills this in with its own run. The researcher edits it to say what they did with the review.
We used Feynman (feynman.works), an AI research-review tool running on [model], to critique [the study] on [date] (run [run ID]). It reconstructed [n] claims, retrieved [n] published records, and raised [n] objections through reasoning frameworks drawn from the published work of [lenses]. Each citation in its output was checked against the records it retrieved, and we read every source we relied on. Its objections informed our analysis; the conclusions, and any errors, are our own.
Membership is by application, and checkout remains closed until launch validation is complete. Membership and licences · Terms · Privacy