We turned SEC 10-K risk-factor sections, Federal Register rule notices, FinCEN guidance, and DOJ/OCC enforcement orders into a living, interlinked knowledge base of 98 cross-linked pages using a four-stage process, then hardened it with an integrity layer that fact-checks itself. Here's how it works at a high level.
01 - Extract
We read the document archive and identified the knowledge structures inside it: entities (FinCEN, the OCC, the Federal Reserve, supervised banks), regulations (the Bank Secrecy Act, the Corporate Transparency Act, the CDD Rule, OFAC sanctions), filings (10-K BSA/AML risk factors, SAR and CTR forms), and enforcement actions (consent orders and DOJ resolutions). Each source was read once, deeply, and decomposed into atomic, structured fragments - not summarized, but pulled apart by meaning.
02 - Structure
We designed a template for each page type - regulation, entity, filing, enforcement, source, synthesis - so every fragment follows a consistent schema. A regulation page always carries its citation, effective date, and penalties; an enforcement page always splits the penalty by agency. That consistency is what makes the system compoundable: new information slots into a known structure instead of floating as free text, and exact figures and citations are never rounded away.
03 - Connect
Every fragment was cross-linked into a knowledge graph - 1,097 internal links, all resolving. The record $3.09B TD Bank AML resolution links to the Bank Secrecy Act it violated, the CDD Rule behind it, and the four agencies that brought it. Where the law itself changed, the base preserves both states: the 2024 Beneficial Ownership Information reporting rule and the March 2025 interim rule that exempted domestic companies are both kept, each with its effective window, and the conflict is flagged - not silently overwritten.
04 - Compound
With the base built, every new source and every new question strengthens the whole. A query like "which enforcement actions cite transaction-monitoring failures" draws across the TD Bank, Danske Bank, USAA, and Capital One pages to produce an answer no single search could assemble - and that answer is filed back into the base, making the next query richer. The knowledge compounds instead of scattering.
The integrity layer
In regulated finance, a knowledge base is only useful if a compliance officer can trust every number. Every claim carries its primary source - a Federal Register citation, a CFR section, an SEC accession, a DOJ release - and nothing is stated as fact without provenance. When a figure can't be verified, it is flagged, not invented.
Before the base was signed off, an adversarial fact-checking pass attacked its uncertain claims against primary sources. It corrected a mis-stated OFAC civil-penalty figure to the real $377,700 cap from the current Federal Register inflation table, and it caught a real-estate reporting rule that a federal court had vacated nationwide - refusing to present superseded law as current. A regression suite of golden queries runs on every update, so if a load-bearing fact silently changes, a red check catches it. The base is auditable and self-checking - it can't quietly rot.
The result: 98 interlinked pages - 16 regulations, 17 entities, 12 filings, 12 enforcement actions, 23 source citations, and 7 cross-cutting syntheses - every wikilink resolving (1,097/1,097), 9/9 regression checks green, and all 18 uncertainty flags fact-checked against primary sources. Compiled and verified by a fleet of AI agents in days, then hardened - browsable in Obsidian or any markdown reader, and fully owned by the team.