One figure. One scientific question.
Choose a visual by the question it can answer. Select any entry to open its runnable code, real result, interpretation guide and evidence boundary. Badges identify the evidence basis; they are not interchangeable.
20 core visuals
Where is computational activity concentrated across the network?
How does evidence travel from the input toward the selected decision?
Which pathways carry the largest share of decision evidence?
How does the representation spectrum change by depth?
Which units change the output when intervened upon?
Which units are most associated with the current prediction?
What routing pattern is expressed by attention?
How does this sample move through representation space?
What is the compact signature of the model's reasoning path?
Where do supporting and opposing signals collide?
How does class evidence accumulate or reverse through depth?
Which tested mechanism changes the selected output?
Do observational and interventional measures agree?
Which input regions are associated with this target?
Which features matter at each position or step?
Which internal concepts and paths form the explanation graph?
Which internal elements form functional communities?
How is relevance redistributed through the network?
Does the explanation remain stable under bounded perturbations?
What changed between the original and counterfactual decision?
7 signature views
What are the model's structural and computational landmarks?
How do discovered concepts evolve through the model?
Which mechanisms survive direct causal testing?
Can an executable programme reproduce the model's decisions?
Why this decision rather than the strongest alternative?
What bounded change is sufficient to alter the decision?
How much computation remains unexplained by the recovered mechanism?
Population and evolution views
Which samples or units form coherent activation communities?
How are samples organised by their internal reasoning?
Which samples occupy unusual internal regions?
Which examples are internally similar despite surface differences?
How does reasoning change across checkpoints?
Where do two models agree or diverge internally?
18 mechanism audits
Which internal circuit carries the decision?
Is the model relying on a simpler unintended cue?
How early does the eventual prediction become recoverable?
Which mechanisms are specific to a class?
How far has the internal representation moved?
Which internal stages distinguish errors from correct decisions?
What is the shape of the local stable region?
Where do explanation methods contradict one another?
Which model capacity is inactive or irrelevant for this task?
How does an attack redirect the mechanism?
How much computation is spent for the evidence obtained?
Are multiple internal circuits sufficient for the same decision?
Do components matter more together than separately?
Which activation mediates the effect of an input change?
Does the internal graph reorganise under change?
Which attributed elements are stable across repeated explanations?
Does a discovered mechanism transfer to another model or domain?
When is the model confident outside its supported representation region?