vraven.darkside · function
discover_decision_circuitry
discover_decision_circuitry(model: 'torch.nn.Module', inputs: 'Any', *, target: 'int | None', layer_names: 'Sequence[str] | None' = None, group_size: 'int' = 8, relative_score_tolerance: 'float' = 0.05, absolute_score_tolerance: 'float | None' = None, baseline: 'str' = 'zero', maximum_groups: 'int' = 256, config: 'VRavenConfig | None' = None) -> 'AnalysisResult'Find an approximate epsilon-sufficient sparse internal core.
The candidate universe is formed by contiguous unit groups from the selected
layer outputs. The routine first measures each group's individual ablation
effect, then greedily removes low-effect groups while the target prediction
and declared score tolerance are preserved. The result is order-dependent
and is not claimed to be the globally minimal circuit.
Usage
from vraven.darkside import discover_decision_circuitry
result = discover_decision_circuitry(... )Source of truthThis page is generated from the checked-out public symbol, signature, and docstring. Pin the package version for reproducible work.