BankCo · In-House LegalSR 11-7 governed
Model oversight · SR 11-7

Can the institution trust the engine?

The screen the reviewing lawyer never opens — and the one a Chief Risk Officer signs. Every number is derived from the same cases the queue routes, so governance and operations never disagree.

Auto-routed
60%
3 of 5 decisions, unattended
Sent to a lawyer
40%
2 held below the gate
Calibration margin
+2%
mean headroom above the gate, auto-routed
QA sample
1
33% of auto-routed, pulled for review
Exceptions this period
reviewer overrides of the engine
Calibration · confidence you can defend

When the engine says 95%, observed accuracy must hold at 95%. Nothing auto-routes to the left of the gate — those cases are escalated to a human. That invariant is the SR 11-7 commitment.

gate 95%80%100%
auto-routed (≥ gate) escalated to a lawyer (< gate)
Ongoing monitoring · human-review sample

Validation is not one-time. A share of auto-routed cases is pulled back for a lawyer to confirm — continuous evidence the gate is still calibrated in production.

How precedent compounds

Every lawyer override becomes an institutional rule. As the bank's own precedent accrues, the engine defers to a human less often — a moat a generic model can't copy, because it's thisbank's judgment.

3
rules learned
2
cases now governed by precedent
60%80%
auto-route rate, with 1 escalation now precedent-backed
  • JV NDAs with regional banks → Senior Corporate Counsel (not Junior pool, even at low standstill exposure)
    learned 6 weeks ago · now governs Westfield Bank
  • Vendor MSAs touching customer PII → Privacy Counsel ahead of standard vendor review
    learned 2 weeks ago · now governs DataSpring Analytics
Audit posture

Every decision carries its extracted signals, policy basis, calibration, and the reviewer's disposition. The model-risk pack bundles this into the evidence a CRO hands a regulator.