CIRCUS: Circuit Consensus under Uncertainty via Stability Ensembles

Circuit analysis in mechanistic interpretability has a hidden fragility: change the pruning threshold and you get a different circuit, with no principled way to know which one to trust. CIRCUS fixes this by running the same attribution graph through multiple threshold configurations and keeping only the edges that survive all of them. This consensus circuit is 28x smaller than the union of all views, yet retains 86% of the full graph's causal influence — validated by activation patching against the actual model. This advantage holds across 94% prompts on Gemma and 95% prompts on Llama, and replicates on 50 diverse reasoning tasks spanning arithmetic, commonsense, and science questions. We show that the fraction of influence retained is a certificate: the smaller it is, the less faithful the circuit, provably. As a direct production application, this signal flags low-confidence model outputs for human escalation with precision 0.78 and recall 0.90, at only 16% added compute cost and no extra model runs.