Reprolith

Point it at a modeling paper. Get back proof of whether the model reproduces its own results.

About half of published biomedical models can't be reproduced from their own paper. Reprolith rebuilds the model, re-runs it, and checks every result against the paper's own figures and tables. You get a certificate: for each result, reproduced, partially reproduced, or not, and why.

Checking a biology claim usually takes a lab. Checking a paper's figure doesn't: it's a computational result, and re-running the model either produces it or it doesn't. That test is free, repeatable, and exact to a stated tolerance, and Reprolith does nothing else.

• A verdict per result, not per paper. A curve has to match on average and at its worst point, so a doubled peak can't average its way into a pass.
• Honesty built in. If a result only reproduced because Reprolith had to assume a missing value, the certificate says so. It never takes credit for its own guesses.
• A list of what was missing. When a paper falls short, you get the exact parameter, unit, or condition it left out, which is the thing an author can actually fix.
• The same answer every time. Pinned engines, deterministic runs, and a certificate anyone can regenerate.

A certificate says one thing only: the model regenerates its own published results. That is not the same as biologically correct, and it is not clinical advice. Every certificate says so.

Work from a clone; the published certificates are repository data.

pip install -e ".[dev]"              # add ,engine to run models (COPASI + libSBML)
reprolith catalog                    # the papers in the catalog
reprolith self-validation            # how often Reprolith's verdicts match known answers
reprolith certificate <digest>       # one certificate, human-readable
reprolith gaps <digest>              # what the paper left out
reprolith archive-check <file.omex>  # for authors: what a reproducer will find in your archive

reprolith --help lists every command. AI agents get the same answers over MCP: run reprolith-mcp and see docs/mcp-server.md.

Pre-alpha, and honest about it. Six model classes are built and self-validated in the open: pharmacokinetic/pharmacodynamic, constraint-based, kinetic, logical, stochastic, and spatial. Every certificate, blind track record, and known limit is published in this repository.

For the full account, including every number, every command, and what isn't built yet, read docs/in-depth.md. The spec is in openspec/, and CONTRIBUTING.md covers building and testing.

Code: MIT (LICENSE). The data Reprolith produces (certificates, catalog, reports) is CC BY 4.0 (LICENSE-DATASET). Third-party model files under datasets/ keep their own licenses, and some are non-commercial: see datasets/THIRD-PARTY-NOTICES.md.