For regulated industries

The only AI platform regulated industries can actually deploy.

Deterministic compliance. Zero PII to the LLM. Audit-ready from day one. ARC generates high-stakes customer documents the way your MRM, legal, and compliance teams would — if they could scale.

SR 11-7 Documented
Model risk governance
Zero PII to LLM
Architectural firewall
Hash-chained audit
Tamper-evident by default
WCAG 2.1 AA
Plain-language gated
The gap

Every GPT pilot in regulated finance fails the same four questions.

ARC is built around the answers.

Problem
Hallucination risk

LLMs invent citations, misstate terms, and fabricate regulatory language under pressure.

ARC
Deterministic Logic Engine

A json-logic rules engine selects every clause. The LLM only phrases pre-approved content — it never decides what goes in.

Problem
PII exposure

Sending customer names, balances, or SSNs to any third-party AI provider violates data-residency and privacy mandates.

ARC
Architectural PII firewall

The model receives logic flags and jurisdiction context only. Customer data is injected locally after generation via {{PLACEHOLDER}} tokens.

Problem
Audit gaps

Regulators need to verify who approved what, when — and prove nothing was altered between draft and delivery.

ARC
Immutable hash chain

Every audit entry is SHA-256 chained to the previous. R2 WORM archives per batch. End-to-end verifiable.

Problem
Regulatory drift

ECOA, FCRA, and provincial rules change. Templates referencing old citations ship silently for months.

ARC
Regulatory change management

Effective-dated clauses with scheduled activation. Impact analysis shows every template affected before the change goes live.

How it works

Four load-bearing capabilities your compliance team will want to interview.

The PII Firewall

Your customer data never reaches the AI model.

ARC splits every document into two streams. The AI sees the logic plan — regulation IDs, jurisdiction, product flags. The local injector sees the PII — name, balance, account number. They meet only at render time, inside your own runtime.

  • Registered placeholder tokens for every PII field
  • Hallucination guard rejects any fabricated token
  • Relational identity (household, referrer) stays local too
Logic stream
jurisdiction: "ON"
regulation: "ECOA"
flag: VIP_REFERRAL
LLM
GPT-4o
firewall
PII stream
name: "J. Smith"
balance: $4,250.00
email: j@...
Local injector
CF Worker
Streams meet only at render time.
Neuro-Symbolic Pipeline

Deterministic where it counts. Fluent where it matters.

Seven stages: Trigger → Logic Engine → Content Plan → AI → Auditor → PII Injection → Delivery. The symbolic layer owns compliance. The neural layer owns voice. The seams are auditable.

  • Model-agnostic — bring your preferred AI provider
  • Secondary adversarial auditor with veto power
  • Golden Template cache — identical regulatory profiles share one generation call
Step 1
Trigger
Step 2
Logic
Step 3
Plan
Step 4
LLM
Step 5
Auditor
Step 6
PII Inject
Step 7
Deliver
Step 6 is the first time customer data touches the pipeline.
Immutable Audit Trail

Prove what went out, and who signed off.

Every audit entry is SHA-256 chained to the previous entry. A chain-walk endpoint verifies the full history in one call. Per-batch snapshots are archived in WORM mode. Two-person approval is enforced in code, not policy.

  • Business + Compliance dual sign-off with separation of duties
  • Statutory deadline engine with AT_RISK / BREACHED alerts
  • Role-based access with full approver identity per action
Entry 1a3f1...7c2prev: generated
Entry 29d8b...42eprev: a3f1...7c2audited
Entry 3b7c4...051prev: 9d8b...42eapproved
Entry 4e2a9...8f3prev: b7c4...051delivered
Each hash = SHA-256(prev_hash ∥ payload)
Fair Lending & Accessibility

Ship evidence, not assertions.

Built-in disparity index across intent × jurisdiction × demographic proxy. Golden Template equality proofs show identical regulatory content reached every proxy group. Every document is gated on Flesch-Kincaid readability and WCAG 2.1 AA.

  • Fair Lending dashboard with heatmap + equality proofs
  • Plain-language retry loop — fails closed if simplification fails
  • Multilingual output with per-jurisdiction language requirements
US-CA
US-NY
CA-ON
Adverse Action
Mortgage Renewal
Collections
Statement
EqualLow disparityFlagged
Coverage

Pre-built for the document types that actually keep you up at night.

Retail Banking
  • Account opening
  • Overdraft notices
  • Fraud alerts
  • Rate changes
  • Monthly statements
Credit & Lending
  • Adverse action (ECOA)
  • Mortgage renewals
  • Collections (FDCPA)
  • Credit dispute (FCRA)
Insurance
  • Benefit determination (ERISA)
  • Breach notifications (HIPAA)
  • Policy disclosures
  • Claim status
Jurisdictions
  • US Federal (ECOA, FCRA, FDCPA, ERISA, HIPAA)
  • Canadian (OLA, PIPEDA)
  • EU data residency
FAQ

Questions every risk officer asks in the first meeting.

How is this different from a GPT wrapper or RAG pipeline?

A wrapper asks the LLM to pick clauses. ARC uses a deterministic Logic Engine — json-logic rules — that selects every clause before the LLM runs. The LLM phrases; it never decides compliance-critical content. Retrieval-augmented generation still lets the model hallucinate citations; ARC's hallucination guard rejects any unregistered token before delivery.

Is our customer data ever sent to the AI provider?

No. The model receives the logic plan only — regulation IDs, jurisdiction, product flags, relational flags. PII (names, balances, emails, household members) is injected locally after generation, inside your own runtime. The two streams meet at render time, not in the model's context window. ARC is model-agnostic; you choose the AI provider that fits your risk posture.

What about data residency?

ARC supports per-region deployment on either Cloudflare or AWS. A data-residency binding routes AI calls and storage to the required jurisdiction. EU tenants can pin to EU-resident AI endpoints and EU-resident storage. No customer data crosses regional boundaries.

How do you handle model risk (SR 11-7)?

ARC ships with SR 11-7 documentation out of the box: model inventory, conceptual soundness, monitoring plan. A live diagnostics endpoint exposes audit pass rate, readability distribution, retry rate, and citation warning rate — exactly the KPIs your MRM team will ask for.

SOC 2? Penetration testing?

We deploy into your Cloudflare or AWS account, not ours — so your existing SOC 2 and penetration testing scope extends naturally. For managed deployments, we pair with your security team to scope attestation. Reach out for the current posture.

How are regulatory changes handled?

Clauses carry effective_from and effective_until dates. A daily cron activates scheduled changes and supersedes older ones. Before a change goes live, impact analysis shows every template and every intent affected — forcing re-approval on anything downstream.

Book a 30-minute architecture review.

Bring your MRM, compliance, or platform lead. We'll walk the pipeline end-to-end against your highest-risk document type — no slides, no sales deck.