As LLMs become reusable infrastructure for evidence access, NLP evaluation must move beyond verifying whether generated outputs are faithful to their sources. We study source-framing propagation: whether the directional effect of institutionally authored evidence is preserved or altered through LLM-mediated representations. We introduce Evidence-to-Action Propagation (E2A), an evaluation protocol that traces how fixed evidence affects reasoning, belief, and action across role-conditioned LLM receivers. In a controlled primary evaluation over 94 post-knowledge-cutoff earnings events and 24,816 LLM-agent decision traces, supplemented by targeted mechanism and scope diagnostics, raw disclosures produce a strong positive-return skew. Source-faithful canonical extraction produces no detectable change in this skew (80.1% vs. 80.0%, p=0.889); no-style synthesis shifts action away from buy, while its belief and quality effects remain unresolved. Shared summaries synchronize evidence paths and rationales. A separate diagnostic shows that randomizing evidence order reduces the skew by 4.30 percentage points, although its quality effect remains unresolved. These findings reveal a faithfulness paradox: reducing hallucination risk through source-faithful representations may still preserve a source’s downstream directional influence. More broadly, the paper argues that NLP evaluation should examine not only whether models reproduce sources accurately, but how language interfaces shape the path from evidence to reasoning and action.