An independent project to make AI-assisted analysis auditable.
Reprosia grew out of open-source R packages for regulated statistics. It exists because the next question after "is the result right?" is "can you show how you know?".
How the work is approached.
- Open source first
- The evidence tools are public on CRAN and GitHub, so anyone can read the code that produces the record.
- Claims match evidence
- Software is described as designed around a regulation, never as compliant with it. Features that are not built are labelled as plans.
- Statisticians decide
- The tools record and verify. Judgement about methods and models stays with the people accountable for them.
- Humans review AI output
- An AI draft is not evidence until a named person has reviewed it and the decision is on record.
- Plain about status
- What is available, what is being built and what is only an idea are separated on every page.
Founder.
Ndoh Penn is a biostatistician and R developer based in Belgium, with experience in pharmaceutical clinical development and assay statistics.
He is the author of regulog, lineager, reproducr and bayprior, and writes about them at reprostats.org.
Work you can check.
Everything here is public, so you can read it yourself.
- Open-source packages on CRAN
- regulog, lineager, reproducr and bayprior, with documentation and source.
- A worked example
- A real run of regulog and reproducr on public data, including a tamper test and its limits: read the example.
- Writing
- What the FDA and EMA AI principles ask of your R workflow, and two design notes on evaluating AI coding agents and AI-assisted independent QC.
- Free templates
- An evidence kit: workflow intake sheet, AI-use register, context-of-use page and reviewer checklist.
Where things stand.
- Open source
- The packages are on CRAN. They are usable on their own.
- Audited AI assist
- In development. No release date is promised.
- Customers and endorsements
- None are claimed. Early conversations with a small number of teams are open.