What Survived the Handoff? Causal Measurement and Value-Aware Targeting in Multi-Agent LLM Communication

When one language-model agent hands work to another, endpoint correctness cannot reveal whether information was transmitted across the handoff or reconstructed by the receiver from its priors and context. This distinction poses an acute reliability problem for novel, proprietary, or unfamiliar facts, where reconstruction can fail silently. We model the relay as a non-ignorable assignment mechanism over epistemic atoms, separating causal availability value ${\Delta }^{\text{av}}$ from budget-neutral displacement value ${\Delta }^{\text{bu}}$. Across financial filings, causal decomposition shows that receiver reconstruction accounts for a $0.676$ omission baseline against $0.842$ endpoint fidelity, while the preregistered hypothesis linking reconstructability to prior access remains an unsupported negative result. In fixed-budget allocation (Experiment C), predicting ${\Delta }^{\text{bu}}$ from source-only pre-treatment features via CausalRelay achieves $0.858$ endpoint accuracy ($+0.067$ over random, within $0.014$ of oracle), significantly outperforming natural-policy imitation ($0.817$), ${\Delta }^{\text{av}}$ targeting ($0.824$), and direct textual compressors (LLMLingua-2 at $0.829$, DAC at $0.836$). Causal measurement thus prevents apparent endpoint success from masking handoff unreliability and enables value-aware, leak-free message targeting.