The Short Answer
SR 11-7 compliance for AI isn't a vendor checkbox — it's an architecture requirement. ibl.ai is built around the architectural pattern SR 11-7 actually demands: self-hosted runtime inside the bank's environment, model-agnostic (so the MRM team can swap models without vendor coordination), open-source platform code (so the bank can audit the orchestration layer), and full audit logs inside the bank's SIEM.
What SR 11-7 Actually Requires of Bank AI
OCC SR 11-7 (and the joint Federal Reserve / FDIC interagency model-risk-management guidance) frames model risk through three lenses:
1. Model development soundness — the bank validates the model's conceptual approach, data inputs, and theoretical fit for the workload. A managed AI vendor controlling the model selection forces the bank to accept the vendor's choices; the bank's validation is constrained to inputs and outputs.
2. Implementation soundness — the bank verifies the model is implemented correctly in the production system. With a managed vendor, the bank can't inspect the runtime; implementation soundness collapses to "we trust the vendor's SOC 2 attestation."
3. Use soundness + ongoing monitoring — the bank monitors the model continuously, captures version metadata, and re-validates after material change. A vendor that updates the model on the vendor's release cycle (not the bank's) creates monitoring gaps that audit will flag.
Self-hosted on ibl.ai aligns with all three — because the runtime, the model artifacts, and the audit chain all live inside the bank's MRM scope.
How ibl.ai's Architecture Satisfies SR 11-7
Self-hosted runtime inside the bank's environment. OpenClaw or NemoClaw executes in the bank's VPC, on-prem data center, or air-gapped enclave. Implementation soundness becomes a normal IT validation, not a vendor-trust exercise.
Model-agnostic + version-pinned. Run Claude (any tier), GPT-5, Gemini, Llama 4 (self-hosted), DeepSeek-R1 (self-hosted). The bank's MRM team pins specific model versions, validates the chosen version against the workload, and signs off on the validation pack. Model swap is a new validation event — not a vendor surprise.
Open-source platform code. OpenClaw is MIT-licensed; the platform license is perpetual. The bank can inspect the orchestration logic, document it in the validation pack, and reproduce it indefinitely. Development soundness becomes auditable.
Audit logs in the bank's SIEM. Every AI call logs to the bank's existing SIEM with model version, prompt template, input hash, output, decision flag, and disposition. Use soundness + ongoing monitoring runs through the same observability the bank already uses for everything else.
For the broader segment context: AI Cost Math for Financial Services: Per-Seat vs Usage-Based in 2026 + Air-Gapped AI for Banks.
What SR 11-7 Validation Looks Like on ibl.ai
A typical validation pack for an AML triage workload:
- Conceptual soundness — workload definition (alert narrative drafting), input sources (transaction monitoring system, customer profile, sanctions list, KYC data), output specification (drafted narrative + disposition recommendation + cited reasoning), benchmark against existing manual process.
- Implementation soundness — runtime version, model version + provider (e.g., Claude Sonnet 4.6 via Anthropic API or self-hosted Llama 4), prompt template version, tool integrations + auth flow, deployment environment (which VPC, which subnet, which IAM role).
- Use soundness — sample-based output validation, comparison against human-analyst gold-set, accuracy + precision + recall on labeled examples, performance under edge cases (multi-account scenarios, jurisdictional variations).
- Ongoing monitoring — model drift detection (output distribution change over time), provider availability + fallback behavior, audit-log completeness, re-validation triggers (model version change, prompt template change, material workflow change).
Every component lives inside the bank's existing MRM tooling. No vendor coordination required for the validation cycle.
Workloads SR 11-7 Touches
Any AI workload that affects bank decisions falls under SR 11-7's scope. In practice:
- AML alert triage — narrative drafting + disposition recommendation
- KYC document review — beneficial-ownership reconciliation, PEP screening
- Credit underwriting assistance — borrower-context summarization, decision-rationale drafting
- Advisor copilot for private wealth — investment-recommendation justification, portfolio-rebalancing rationales
- Trading-desk research synthesis — market-news triage, position-justification drafting
- Compliance + audit response — policy Q&A, audit-finding response generation
For per-workload cost math: What AI AML Alert Triage Actually Costs in 2026.
Why Managed AI Vendors Struggle With SR 11-7 at Scale
The managed-vendor pitch typically includes: "we're SOC 2 + we run on a BAA-equivalent + we follow industry best practices." That satisfies third-party-risk reviews. It doesn't satisfy SR 11-7's specific requirement that the bank validate the model and govern it on the bank's cycle.
Three specific failure modes:
- The vendor changes the model. Bank's validation pack was for v1.2; vendor pushes v1.3. The bank either pauses the workload pending re-validation OR accepts uncontrolled change. Neither is acceptable.
- The vendor's audit log lives in the vendor's cloud. Examiner asks for the reasoning behind a flagged alert; the bank requests from the vendor; the vendor supplies what the vendor has retained. Chain of custody.
- Multiple model providers create multiple validation cycles. A bank running Sonnet for AML + Opus for appeals + Haiku for triage has three vendors with three change cycles. Self-hosted with the bank controlling model versions consolidates the validation cycle.
Run the Numbers
- AI Cost Math for Financial Services: Per-Seat vs Usage-Based in 2026 — segment cost math
- What AI AML Alert Triage Actually Costs in 2026 — per-alert math + SR 11-7-aware deployment
- Air-Gapped AI for Banks: Why FINRA + SR 11-7 Make It the Default — broader deployment-shape argument
- Financial Services AI Reference Architecture on ibl.ai — full SR 11-7 / FINRA / SOX / PCI architecture
- Financial Services Blueprint: Air-Gapped AI in 90 Days — staged deployment recipe
- Self-Hosted AI vs ChatGPT Enterprise for Financial Services — deployment comparison
Why Family-Owned and New York Matters Here
For a bank's AI vendor relationship — touching MRM scope, examiner audits, and regulator subpoenas — the structure of the vendor matters. ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned, long-term partner with a perpetual platform license and no investor exit pressure. The runtime is open source. The transaction data stays inside the bank's VPC. The validation pack stays valid because the bank, not the vendor, controls the model version.
SR 11-7 compliant AI isn't an attestation. It's an architecture.