Follow a single UCC risk request all the way through, naming the track at each step. Assemble the context — the architected system prompt (T2), retrieved and reranked filings (T3), shaped tool results and a fused multi-source picture (T3), the analyst's memory and profile (T4). Position and compress it — the disqualifying lien at an edge, the bulk compacted (T5). Govern what goes in (injection defense, T6) and what comes out (a schema-enforced, PII-safe, audited decision, T6). Run the whole thing inside an agent loop that escalates the borderline cases to a human (T7). And observe, evaluate, and version every bit of it (T8).

That's the point of the capstone: a production answer is the product of the entire pipeline, not any single technique. The disqualifying lien only changes the decision if it was retrieved (T3) and positioned where the model reads it (T5) and not compressed away (T5) and not overridden by an injection (T6) — and you can only know it worked if you measured it (T8).

The organizing structure underneath is M02's lifecycle: assemble → position → govern → observe → iterate. The capstone is that loop, running in production, over the UCC domain.

Context EngineeringMastering the LLM Context Window

Context Engineering
Mastering the LLM Context Window