Trusted by Default, Doubted When Official: Source Framing Gates Whether Context Overrides Memorized Knowledge

Retrieval and long-context systems promise that models can be corrected by placing the right fact in context. We show that whether this works depends less on the fact itself than on how its source is framed. Using memorized public numeric benchmarks as continuous, exactly known probes, we inject source values that conflict with a model’s parametric memory by controlled magnitudes and measure context-reliance β, the fraction of the way the answer moves from the memorized value to the injected value. Across seven frontier models from three providers, bare data records and RAG-style retrieved documents are adopted in full, even when the injected value is highly implausible or contradicts strongly memorized knowledge. Marked sources behave differently: authoritative attribution triggers memory-based adjudication and is increasingly overruled as the conflict grows, while hedged sources are discounted toward memory. The effect reproduces on a second memorized domain, national inflation by year, and is unchanged by evidence position in a multi-record retrieval context. A logit-level probe in an open model further shows the same framing-driven re-weighting in token probabilities: plain records shift probability mass toward the injected value, while authoritative and hedged framings preserve more mass on the memorized answer. These results show that the boundary between context-time evidence and weight-time memory is gated by source framing. For retrieval and long-context systems, the implication is double-edged: plain evidence is most effective for correction, but that same unconditional adoption makes false plain records a vulnerability.