Two-Tier 'Librarian + Writer' Architecture Removes 6,845 Cross-Section Contradictions From LLM Research Reports to Zero
This system separates a deterministic 'librarian' that ingests timestamped sources into a trust-tiered ontology (evidence cards, an authoritative metric ledger, a claim graph) from a portable multi-agent 'writer' that composes reports at any cutoff T, reading only evidence with as_of <= T. Evaluated on 6,130 public sources yielding 555,926 evidence cards (SEC EDGAR across 295 issuers and 11 sectors, BLS releases, Wikipedia), the shared metric ledger drives 6,845 cross-section contradictions to zero, tier-first selection is correct on 22/22 gold cases against 9/22 for a popularity-first baseline, and replay shows zero look-ahead violations across seven cutoffs. Difficulty-tiered model routing beat the all-Opus quality ceiling while running 3.7x faster than serial.
↳ Follow the thread