Pasteur¶

Clinical AI stress testing

Pasteur

Find where clinical models become brittle before deployment.

Test model behavior under realistic data failures. Pasteur simulates missing measurements, measurement noise, and transitions between patient cohorts, then quantifies how predictions respond. All computation stays on your machine.

Get started → View on GitHub

Blackout

Remove a feature and its measurement companions to test missingness.

Simulations

Jitter

Add controlled measurement noise and measure prediction stability.

Simulations

Flipper

Interpolate between differently labeled patients and locate decision flips.

Simulations

100% local	Runs on your machine’s CPU. No server, daemon, or admin rights.
No network, no telemetry	No network calls at runtime. Data, models, and results never leave the machine.
Standard tooling	`pip install pypasteur` or `cargo install pasteur-cli`. BSD-3-Clause open source.

Note

Reviewing Pasteur for a hospital or health system? Start with Security and data handling, or download the one-page overview (PDF).

What Pasteur produces¶

One simulate run creates a reproducible bundle containing the clean cohort and each stress-test variant. evaluate scores one ONNX model; compare places several models on the same rows and can write a parquet of per-row predictions.

Command	Input	Result
`simulate`	Local parquet data and optional cohort labels	Clean, blackout, jitter, and flipper parquets
`evaluate`	Simulation bundle, labels, and one ONNX model	Baseline and stability metrics (Reading the results)
`compare`	Simulation bundle and multiple ONNX models	Model comparison and optional row-level predictions
`card`	Simulation bundle and provenance	A Hugging Face-compatible dataset card

Quick look¶

pasteur-cli simulate \
  --input patients.parquet \
  --id-col patient_id \
  --feature glucose \
  --output ./output

This writes clean/, blackout/, and jitter/ beneath ./output. Add --labels groups.parquet to generate flipper pairs. For a complete run from CSV to a model comparison, see Walkthrough: choosing between models.

Designed for evidence, not a pass/fail badge¶

Pasteur does not claim that a model is safe. It produces concrete evidence about model behavior under declared perturbations: what changed, where it changed, and how strongly. Those results belong alongside intended-use documentation, local validation, clinical review, and deployment monitoring. Pasteur is a research and evaluation tool, not a medical device.