Guide 03 · Explanation

Explain one model decision.

Start with the beginner command. Choose a deeper mode only when the question requires it.

Recommended: use a bundle

vraven analyse ./my-model \
  --task explain \
  --preset standard \
  --target 1 \
  --format html \
  --output explanation.html

Presets control depth: quick favours speed, standard is the default balance, and thorough collects more evidence.

Or load a state dictionary directly

vraven explain model.pt \
  --model-format state_dict \
  --factory my_project.models:Classifier \
  --input input.npy \
  --target 1 \
  --mode gradient \
  --report-format html \
  --output explanation.html

Choose the evidence mode

Mode	Use it when
fast	You need reduced activation evidence with low overhead.
gradient	You need target-sensitive gradients and input attribution.
causal	You need limited controlled intervention evidence. It is slower.

Interpret carefullyActivation or gradient magnitude is observational evidence. VRAVEN labels causal claims only when they are supported by an intervention.