Benchmark run complete: 595/600 evaluations, grounded cuts misstatements 56% vs naive

Phase 3 complete: two-layer judge with deterministic citation checker, LLM entailment judge, Cohen's kappa — smoke test passed

LangSmith tracing wired into all five configs and eval harness

Phase 2 complete: all five configs + eval harness working end to end

Config 5: grounded/citation-forced RAG pipeline complete

Config 4: KG-augmented RAG with cross-reference graph expansion (KG working)

Configs 1-3: naive, hybrid, rerank RAG pipelines + citation format fix

Config 1 (naive) and Config 2 (hybrid) RAG pipelines working

Phase 1 complete: retrieval layer (dense+BM25+RRF) + naive RAG config end-to-end

Retrieval layer: dense search + BM25 + RRF combiner (smoke test verified)

Embed all corpus chunks with text-embedding-3-large (11613 chunks, /bin/zsh.23)

Clause-aware chunker: paragraph ID preservation, header fix, SCHEDULE tagging, APRA artifact filter

Clause-aware chunker: paragraph ID preservation, header fix, SCHEDULE tagging, APRA artifact filter

Add Neon connection + pgvector enablement (src/db.py)