← Observatory home
Reverse · 2 of 12

Adversarial Mechanism Audit

Which mechanism changes under bounded adversarial pressure?

image_pattern_cnn2d · Open result image ↗
Adversarial Mechanism Audit result for the image_pattern_cnn2d fixture
Real output from the small deterministic CNN-2D documentation fixture.

What it shows

A perturbation-by-perturbation view of prediction and internal mechanism stability.

Claim boundary

This is evidence for this input and model state, not a universal statement about every input.

Run this with the installed CLI

Point VRAVEN at a validated model bundle. This produces one focused HTML report and works after a normal PyPI installation.

vraven analyse ./my-model \
  --task explain \
  --preset standard \
  --format html \
  --output vraven-report.html

Start with the model-folder guide if your bundle is not prepared yet. The larger documentation fixture is a contributor workflow described in Develop from source.

Generate this view

AdversarialAuditor(model).audit_sample(input_tensor, target=target).write_html(output_path)
InputModel bundle, one reference input, selected target
Display modepublication_light
Evidence levelInterventional or perturbational
Machine outputreverse/adversarial_audit.html
Related APIvraven.AdversarialAuditor
Fixtureimage_pattern_cnn2d

How to read the result

  1. Start with the title and target metadata in the generated report.
  2. Read the strongest structures first, then inspect weaker or opposing evidence.
  3. Check the stated evidence level and status before treating a relationship as causal or formal.
  4. Retain the trace, model version, preprocessing, and VRAVEN version with any published claim.
Reproducibility noteThe screenshot is public documentation evidence. The source run remains local-first; raw provenance files are not published because they can contain machine paths.