Public result gallery

One model, many evidence views.

These four previews are a small selection from the wider VRAVEN output system. Every preview below was generated from torchvision MobileNetV2 and an openly licensed Labrador image. Select a view to open its code, result and interpretation guide; use the walkthrough for the complete case study.

VRAVEN MobileNetV2 input attribution
Input attributionSigned input-times-gradient over deterministically sampled image positions.Open code and result →
 VRAVEN MobileNetV2 decision flow
Decision flowSupporting and opposing local signals across model depth.Open code and result →
 VRAVEN MobileNetV2 activation energy
Activation energyObserved signal magnitude—not causal importance—across captured layers.Open code and result →
 VRAVEN MobileNetV2 causal interventions
Layer interventionsTarget-score change under three temporary zero-ablations.Open code and result →

What a run can contain

Human-facing

  • Interactive dark HTML
  • Publication-light figures
  • Combined explanation or certificate reports
  • PDF, SVG, PNG, JPEG and WebP exports

Machine-facing

  • JSON traces and report payloads
  • Mechanism IR
  • Capability and provenance manifests
  • Executable Python or C decision programme
Case-study scopeThe result applies to one image, one model version and the declared interventions. Observational visuals are not converted into causal claims.
Read the walkthroughOpen the full report