Financial services

SR 11-7 was written for models. Your agents are the operators.

Model risk management assumed a model produces an output a person acts on. An agent produces the action. Validation, governance and ongoing monitoring still apply — but the record has to come from something that is not the agent, and it has to survive an examiner asking what happened on a Tuesday in March.

/001 · THE PROBLEMEXAMINER-READY OR NOT

Three questions an examiner asks that a dashboard cannot answer.

Each one is a record question, not a monitoring question. Detection tells you something looked wrong. An examination asks you to prove what actually happened, to someone who does not take your word for it.

“Which agent took this action, under whose authority?”

Requester identity propagates into every event. No ambient authority, no service account standing in for a person. The delegation chain travels with the record, including causal depth and parent event.

“Show me the policy that was in force at the time.”

Policies are version-controlled Rego with an audit-logged approval workflow. The policy version hash is on the event, so the rule that actually evaluated is recoverable, not reconstructed.

“How do I know this log was not edited?”

Every event is Ed25519-signed and SHA-256 hash-chained. Verification runs offline against a trust anchor frozen at enrollment. Your examiner does not need a Behavry account, or our cooperation.

“What was the unapproved tool nobody told us about?”

Surface discovery cross-references IdP apps, SaaS admin APIs and endpoint signatures through a four-state model — Licensed, Enabled, Active, Governed — so the gap between what is running and what is authorized is a number, not a guess.

/002 · CONTROL MAPPING13 CONTROLS · 6 REGIMES

Mapped to the rules you are actually examined against.

Each control below is backed by shipped enforcement, not a roadmap entry. Coverage is stated honestly: full where the control is satisfied by shipped capability, partial where Behavry supplies evidence but the obligation extends beyond what any platform can discharge for you.

RegimeSectionControlCoverage
OCC SR 11-7§IV.AModel ValidationFull
OCC SR 11-7§IV.BModel GovernanceFull
OCC SR 11-7§IV.COngoing Model Risk MonitoringFull
NYDFS Part 500§500.7Access PrivilegesFull
NYDFS Part 500§500.6Audit TrailFull
SEC Rule 206(4)-7§275.206(4)-7(a)Written Compliance PoliciesFull
SEC Rule 206(4)-7§275.206(4)-7(b)Annual Review + MonitoringFull
SEC Rule 206(4)-7AI governanceEmployee Acceptable Use & Unauthorized Tool BlockingFull
SEC Rule 204-2§275.204-2(a)Record CreationFull
SEC Rule 204-2§275.204-2(e)Record Retention (5-year / 2-year accessible)Full
SEC Reg S-P§248.30NPI Safeguards — Policies & ProceduresPartial
SEC Fiduciary DutyDuty of CareData Management & Model Bias MonitoringPartial
SEC Fiduciary DutyDuty of LoyaltyForm ADV & Client Disclosure SupportPartial

What “mapped” means. Our shipped controls satisfy the named clauses today. It does not mean an independent auditor has certified them. Current status on the Trust Center. Full framework-to-control mapping with PDF export in the dashboard.

/003 · SR 11-7 IN PRACTICEVALIDATE · GOVERN · MONITOR

The three-pillar structure, applied to something that acts.

§IV.A · Validation

Behavioral baselining with Welford EWMA. Six-dimension anomaly scoring. Drift detection by Bray-Curtis divergence. Each agent carries a baseline tool manifest, and calls outside it are denied rather than noted.

§IV.B · Governance

Version-controlled Rego policy-as-code, a change-request workflow with audit-logged approve and reject, and a tamper-evident hash-chained record of what the policy did.

§IV.C · Ongoing monitoring

Real-time behavioral monitor over an event bus, six-dimension risk scoring, per-agent rate limits coupled to risk tier, and a fleet-wide kill switch when a model or an agent has to stop now.

§500.7 · Least privilege

Session-scoped tokens with risk-adaptive TTL, requester identity propagation instead of ambient authority, a restricted enforcement tier, and a per-agent kill switch.

/004 · NPIREG S-P · GLBA

Nonpublic personal information, stopped before the call executes.

Reg S-P and the GLBA Safeguards Rule do not care whether you detected the disclosure afterwards. Behavry evaluates the tool call before it runs, so NPI in an argument is a denial rather than an incident report.

Inline denial

33 DLP patterns including SSN with structure checks, account and routing numbers, and card numbers with Luhn validation. Critical-severity matches auto-block before policy evaluation.

Fragment reassembly

Cross-session correlation catches NPI split across sequential requests — the exfiltration path that per-request scanning is structurally blind to.

Pipeline, not just a scanner

Classify, redact with pseudonymization, BYOK envelope encryption, then retention purge with a decryption audit trail.

Retention that matches the rule

Configurable retention policies with TimescaleDB compression, plus SIEM export as an off-platform durable copy for the 5-year and 2-year-accessible obligation under 204-2(e).

/005 · NEXTBOOK A BRIEFING

Bring the agent your examiner would ask about.

We will walk your agent surfaces, show where the record gets produced, and hand you a signed Decision Trace you can verify yourself with no Behavry service in the loop.

BOOK A BRIEFING →

Related: compliance · decision trace · trust center · integrations