Simulations

Pasteur applies controlled perturbations to a clean tabular cohort. Every run keeps source-row identity and records deterministic row order so clean and shifted variants can be compared.

Blackout

Blackout sets the selected feature to null for a seeded sample of rows. Companion columns can be nulled at the same time:

pasteur-cli simulate \
  --input patients.parquet \
  --feature glucose \
  --blackout-companion glucose_measured \
  --blackout-rate 0.2

Use companions for observation flags or other fields that would become internally inconsistent if the value disappeared alone.

Jitter

Jitter adds seeded measurement noise to the selected numeric feature and can generate several variants:

pasteur-cli simulate \
  --input patients.parquet \
  --feature glucose \
  --jitter-scale 0.5 \
  --jitter-iters 5

The resulting jitter_0.parquet through jitter_4.parquet retain the same source-row IDs as the clean cohort.

Flipper

Flipper samples patients from opposite labels and interpolates between their feature vectors. The resulting path can expose where a model crosses its decision threshold.

pasteur-cli simulate \
  --input patients.parquet \
  --labels groups.parquet \
  --positive-group-id 1 \
  --flipper-pairs 100 \
  --flipper-steps 20

Flipper rows include pair_id, step, t, source-row references, and endpoint labels. Their source_row_id is synthetic because each row lies on an interpolated path rather than representing one source patient.

Reproducibility

All simulators use --random-state (default 42). Record the command, source-data revision, and model contract alongside results; a simulation metric without those inputs is not independently reproducible.

Interpretation

• Blackout tests behavior under declared missingness, not every real-world missing-data mechanism.

• Jitter tests a chosen noise scale, not instrument calibration in general.

• Flipper examines interpolated paths between observed endpoints; clinical plausibility still requires domain review.

These simulations are evidence about a model under explicit conditions. They are not a substitute for local validation or clinical safety review.