Scientific discipline

Know what the evidence permits you to say.

A compelling visual is not automatically a causal result. VRAVEN keeps observational, interventional and formal evidence distinct.

LEVEL 1ObservedWhat occurred inside this run?
LEVEL 2AssociatedWhat varies with the decision?
LEVEL 3IntervenedWhat changes the outcome?
LEVEL 4CertifiedWhat is formally guaranteed?

Observational evidence

Activations, gradients, attention, attribution and representation structure describe a recorded computation. They can locate evidence and generate hypotheses; alone, they do not prove causal necessity.

Interventional evidence

Ablation, patching, perturbation, counterfactual search and adversarial testing modify a declared part of the system or input. Claims are limited to the tested intervention, bounds and sample coverage.

Formal evidence

A certificate is valid only under its stated assumptions and verified region. When prerequisites are missing, VRAVEN should report that status rather than silently upgrading an empirical result.

Read every result with four questions

  • Which model, input and target produced it?
  • Was the result observed, intervened or formally verified?
  • What population, region or perturbation budget is covered?
  • Which computation remains unresolved?

Publication rule

Exported figures should preserve labels, provenance, comparable scales and uncertainty. Decorative graphics must never be presented as measured data.