About Compliance for Claude

Want to get involved? Partner with us, contribute to the corpus, or just reach out — see the ways to engage & get in touch →

Our why

The more compliance professionals use AI, the better compliance will be with AI.

Also, once it’s set up, the more operators use Claude, the better they are at compliance.

There’s a flywheel layered into this. When the P&P your team is refreshing happens to be the firm’s AI Governance Policy — its AI Inventory, its AI System Risk Assessment, its third-party AI vendor sweep — you’re using AI to build the guardrails that let the rest of the firm safely adopt more AI. Artifacts stay confidential inside your compliance program. Rails enable use, use produces artifacts, artifacts govern use, governance enables broader adoption. The compliance team isn’t catching up to AI — the compliance team is the surface area through which the firm can adopt it safely.

What makes this corpus different

• Verbatim quotes only. Every paragraph quote in every register has been verified as a literal substring of an immutable PDF or HTML snapshot fetched from the regulator. We do not paraphrase regulatory text. Paraphrase fails the gate.
• Audit-defensible by construction. Each obligation links back to its source snapshot, fetched date, and content hash. Examiners can verify the cite chain end-to-end.
• Auditor-minimum evidence pathways. Where we have built them out, each obligation has explicit OR-pathway evidence requirements — what an examiner will literally ask to see, broken down by pathway (and selection logic between pathways).
• Mechanism, not assertion. The anti-hallucination defense is in the pipeline gate, not in a "trust us" assertion. Same content goes into the gate, same content comes out.

Copilot, Claude, or both?

The library is AI-agnostic — the files work in Microsoft Copilot, Claude, or ChatGPT — so start where your shop already is. For most people that’s Copilot: it’s already inside your approved Microsoft tenant, so there’s no new vendor to security-review, and the data stays in your tenant.

The tools genuinely differ in one place: size. A quick check or a single P&P section works anywhere. But a full servicing P&P runs 100–300 pages, and Microsoft caps Copilot’s rewrite at about 3,000 words — so it can’t redline a P&P that big. That’s where you move the heavy job to Claude, which holds the whole P&P plus the regulations in context (200K tokens, more on Enterprise) and redlines in one pass. See the context-window comparison →

What Compliance for Claude is not

• Legal advice. Compliance officers make compliance decisions.
• A replacement for a competent attorney for novel or high-stakes situations.
• A live regulatory monitor (yet). Snapshots are point-in-time; we record when each was fetched. Compliance officers should verify against the latest regulator publication for time-critical questions.

Honest disclosures

• Examiner-cite verification. The library contains hundreds of "examiner topic pointers" (CFPB SEM-style section references). Most of these have not yet been mechanically cross-referenced to real exam-manual section IDs. They are flagged in the rendered output with an (operator-verify) prefix until verified.
• Licensing. The library content is licensed for free use by compliance officers, lenders, auditors, and researchers. License choice (MIT vs Apache 2.0 with patent grant) pending. Until selected: the corpus is offered as-is for read-only public use.