v1 ships a sequential multi-agent pipeline (Classifier → Extractor (chunked) → Mapper (per-obligation, sequential) → Report). v6 plan: refactor pipeline.py to use LangGraph as a typed StateGraph with explicit conditional routing. Parallel Mapper execution (4-8 concurrent) targeting 5-8× wall-clock improvement on full-document runs. Critic agent reviews low-confidence mappings (< 0.80) and either confirms, suggests alternatives, or marks for human review. Conditional routing on Classifier output: in_scope=False with low confidence routes to a Clarifier agent that fetches table of contents and re-classifies. Total scope: 10-13 hours. Tracked for v6 — the writeup against v1 ships first, then orchestration becomes the v6 work and a follow-up post.
v1 ships a sequential multi-agent pipeline (Classifier → Extractor (chunked) → Mapper (per-obligation, sequential) → Report). v6 plan: refactor pipeline.py to use LangGraph as a typed StateGraph with explicit conditional routing. Parallel Mapper execution (4-8 concurrent) targeting 5-8× wall-clock improvement on full-document runs. Critic agent reviews low-confidence mappings (< 0.80) and either confirms, suggests alternatives, or marks for human review. Conditional routing on Classifier output: in_scope=False with low confidence routes to a Clarifier agent that fetches table of contents and re-classifies. Total scope: 10-13 hours. Tracked for v6 — the writeup against v1 ships first, then orchestration becomes the v6 work and a follow-up post.