Ask what bounded change reaches another class.

Select the destination class and constrain the search so the result answers a precise question.

Run the search

vraven counterfactual model.pt \
  --model-format state_dict \
  --factory my_project.models:Classifier \
  --input input.npy \
  --sample-index 0 \
  --target-class 2 \
  --steps 200 \
  --max-linf 0.05 \
  --clamp-min 0 \
  --clamp-max 1 \
  --output counterfactual.json

Read the result

Check whether the target class was reached, the final perturbation size, the score change and the largest changed input coordinates. A failed bounded search means no candidate was found by this procedure—not that none exists.

Use the correct validity claim

For the simpler bundle workflow:

vraven analyse ./my-model \
  --task counterfactual \
  --target-class 2 \
  --preset thorough \
  --output counterfactual.json