Required or Redundant? Tracing Evidence Collaboration with Factorized Causal Activation Patching

Large language models often receive multiple pieces of evidence, yet existing behavioral and mechanistic analyses typically measure each piece's marginal influence and therefore cannot determine whether evidence is integrated independently, redundantly, or collaboratively. We introduce Factorized Causal Activation Patching, a framework that measures whether the causal contribution of one evidence statement depends on the state of another. We construct a controlled 400-claim fact-verification dataset spanning eight domains, with matched conditions in which two evidence statements are either jointly required or individually sufficient. A 2x2 factorial intervention independently manipulates both statements, enabling a dependency contrast between their non-additive interactions. In Qwen3.5 and Gemma-4, semantic counterfactuals yield strong behavioral dependency contrasts of 3.81 (95% CI: 3.60-4.02) and 4.21 (95% CI: 3.87-4.53) gold-margin units, respectively. Word-order permutations produce smaller but positive contrasts of 0.90 and 0.31, demonstrating that the effect is not limited to opposing semantic content while confirming that semantic compatibility is its dominant driver. Bidirectional disruption and restoration interventions establish causal directionality, and the layer-zero factorized estimator exactly recovers the behavioral interaction. Position-wise interventions reveal that the dependency signal is causally recoverable at evidence spans in early layers and at the decision position in late layers, with no strong signal at intermediate readout positions, while impossible reverse autoregressive pathways remain exactly null. These results distinguish genuine partner-conditioned evidence collaboration from marginal evidence sensitivity and provide a unified behavioral-to-mechanistic account of multi-evidence decision formation.