Before a clinical model goes live, prove it won't fail quietly.
Pasteur is our open-source clinical AI stress-testing toolkit. It subjects tabular and clinical predictive models to Blackout (missing features), Jitter (Gaussian noise), and Flipper (counterfactual recourses) locally on Parquet and ONNX.
A model with a great AUROC can still be dangerous.
Validation today is a static 80/20 split on data that looks like training. Deployment isn't static — and the failure shows up in production, on a real patient.
Distribution shift
Performance on a held-out split is not performance on tomorrow's patients — the static metric never sees the cohort move.
Silent overconfidence
Models rarely say “I don't know.” They guess — confidently — on impossible inputs, and nothing in the AUROC flags it.
Rare cohorts vanish
Synthetic data generators drop rare phenotypes, so the model fails exactly where the stakes are highest.
We map where a model is safe to use — and where it isn't.
Pasteur simulates thousands of realistic clinical perturbations and traces how predictions deform under each, producing a map of the model's safe operating envelope. Coupled with TRAILED (topological representation learning) and Phil (multiverse imputation), it keeps synthetic and perturbed trajectories clinically faithful — including the rare cohorts standard generators erase.
Blackout masks combinations of clinical features to simulate missing labs; Jitter perturbs continuous vitals to test sensor noise resilience; Flipper finds minimum counterfactual changes that flip a prediction.
Built toward the bar regulators and clinicians actually set: you should be able to judge a model's failure boundaries locally on your own cohort before deployment.
Run Pasteur locally on your models today.
Install the Python bindings (`uv pip install pypasteur`) or the native Rust CLI (`cargo install pasteur-cli`) and explore the full API and interactive guides on docs.krv.ai.