Struggling with Claude's refusal on legitimate medical data analysis

I've been trying to use Claude 3.5 Sonnet for a quick prototype of a tool that helps researchers summarize clinical trial summaries. The input is strictly anonymized text from public FDA databases and academic papers. However, I keep hitting refusal messages saying the model cannot provide medical advice or interpret patient data. The issue is I'm not asking for a diagnosis; I'm asking it to extract specific efficacy metrics and side effect frequencies from the provided text. I've tried various prompt engineering techniques, including explicitly stating that this is a data extraction task for research purposes only and that no individual patient data is involved. The model still seems to over-interpret the presence of medical terminology as a trigger for safety refusals. Has anyone else run into this specific boundary? I'm considering switching to a smaller open-source model like Llama 3 if I can't get reliable output from Anthropic, but I really prefer the reasoning capabilities of Sonnet. If you have a specific prompt structure or system message that helps bypass these false positives without violating the usage policy, I would really appreciate the insight. I don't want to jailbreak the model, just want it to do what it says it can do: process and summarize text. It’s getting frustrating because I’m spending more time fighting the guardrails than building the actual feature.