In-House AI Control Plane Requirements for Examiner-Ready Evidence at a Bank

What changed in supervisory guidance, and why it matters for evidence design

Many internal requirements documents still map AI evidence to SR 11-7. That mapping is now partly outdated. SR 26-2 updates the model risk management guidance first issued as SR 11-7, clarifying principles and emphasizing a risk-based approach tailored to a bank's model risk profile and the size and complexity of its operations. Its attachment states that its principles apply to traditional statistical and quantitative models and non-generative, non-agentic AI models. For tools, processes, or systems it does not cover, it says the bank's own risk management and governance practices should guide the choice of governance and controls.

The practical result is that evidence for generative and agentic systems cannot be mapped one-to-one to SR 26-2 clauses. The bank needs to record its own rationale for how it governs those systems and show that the evidence design follows from it. Confirm with the model risk function how the stated scope applies to each use case.

Separately, the OCC issued Bulletin 2026-54 on November 4, 2026, a supervisory framework that helps examiners evaluate bank use of what it calls super intelligence (SI) agents. It sets no new expectations, and banks are not expected to use it to assess their own SI use. It is still useful as examiner context. It separates Control, meaning controls that may prevent unauthorized, inappropriate, or unsafe actions, from Assurance, meaning evidence that those controls continue to operate effectively. It also tells examiners to tailor reviews to materiality rather than treat SI as a separate risk category, and it refers to human review and risk-based retention. No regulator prescribes record fields, retention periods, or tamper-evidence methods. Those are design decisions the bank must define and defend.

Retention, residency, tamper-evidence, and completeness

Evidence stores that hold sensitive content may warrant access controls and risk-based retention periods, so retention should be configurable by use case or risk tier rather than fixed globally. Materiality drives this: control intensity can rise with the consequences of an agent's failure or misuse, and retention can follow the same logic. Set the periods with legal and compliance, and make access to the store a separately governed permission.

Data residency and deployment location are bank-defined requirements. Running the enforcement and evidence components in the bank's own cloud keeps prompts, outputs, and evidence inside boundaries the bank already controls and audits.

Tamper-evidence is a testable property, not a label. Common approaches include append-only or write-once storage and hash chaining across records, with periodic verification by a party independent of the operators. The requirement is that a reviewer can detect altered or missing records.

Completeness is the harder problem. Evidence is only as good as the coverage of the enforcement point, so the bank needs an inventory of models, agents, tools, and MCP servers, a way to detect traffic that bypasses enforcement, and a periodic reconciliation of the inventory against recorded activity. MCP traffic is a coverage point because agents retrieve context and reach tools through it. An unmonitored tool is an evidence gap regardless of how good the records for monitored tools are.

How these requirements map to shipped Trussed capabilities

Trussed is runtime control and evidence infrastructure for regulated enterprises. It enforces policy in the execution path across models, agents, tools, MCP servers, AI sites, and coding assistants (allow, redact, block, rate limit, escalate for human approval), at the level of individual interactions and tool calls. Audit-ready evidence is generated by enforcement and mapped to HIPAA, NAIC, NYDFS, EU AI Act, and NIST. Deployment is self-managed in the customer's cloud with configurable retention and data residency. It does not provide model lineage, dataset provenance, or end-to-end tracing of agentic workflows.

Trussed has run continuously for 18+ months in a regulated enterprise, with 99.95% uptime on live production traffic, 100+ models under one policy plane, and 30+ GB of audit evidence generated per day. Customers span four regulated verticals: insurance, pharma, healthcare, and education. Insurance, a regulated financial sector, is where Trussed is deepest. The company holds SOC 2 Type II, HIPAA, GDPR, FERPA, ISO 27001, and NIST AI RMF credentials. Use the checklist above to test these capabilities in the same way as an in-house build.