Private Data, Public Findings - Mareforma

View on GitHub

examples/04_private_data_public_findings/

Lab A discovers a finding and publishes its provenance trace. Lab B reads the trace, replicates on its own private dataset, and publishes back. The graph answers three questions automatically:

Independent paths? Different data sources, or both from LLM prior?
Genuinely reproducible? Stable across datasets, or a data artifact?
Provenance distance? How far is each conclusion from raw data?

Also demonstrates a spurious convergence on the stored support_level ladder (a REPLICATED label, deprecated for v0.4.0): two INFERRED claims repeating the same LLM prior, and how to detect it from classification and the derived independence axis rather than trusting the stored label.

Run

pip install langchain-core
cd examples/04_private_data_public_findings
python 04_private_data_public_findings.py