Control framework

Governing AI from registry to runtime: an integrated risk, legal and cyber control framework

One AI product registry, a risk assessment that gates production, runtime guardrails and evidence from every decision, with the three lines built in.

← All whitepapers

By Richard White · · Owner: Richard White · Next review: by 2 December 2026, and when a cited rule changes status

Free, no form.

In brief

AI governance fails when risk, legal and security run separate processes: three reviews of three descriptions of the same system, on three calendars, while the product that ships is one none of them saw. This paper sets out one integrated control framework for banks and insurers. One AI product registry records every model, agent, prompt, tool, data product and vendor in use; if it is not in the registry, it does not run in production. One risk assessment carries the risk, legal and cyber questions together, sets a tier and gates every release, with independent second-line approval for the higher tiers. Guardrails, identity and monitoring enforce at runtime what the registry approved, and every decision leaves evidence. The three lines of defense run through every control point: the first line registers and operates, the second line sets standards and approves independently, internal audit tests the gate and the evidence, and agents never approve.

Who it’s for
Chief risk and compliance officers, general counsel and regulatory lawyers, CISOs, chief audit executives, and the CIOs, CDAOs and product owners who build AI at banks and insurers.
You’ll be able to
Stand up one AI product registry and tiering questionnaire, run a production gate with evidence by tier and named independent approvers, assign the three lines at every control point, map US, EU, UK, Canadian and Singapore obligations and the OWASP and MITRE ATLAS threats to controls, and report to the board from the same record.
Short on time?
Read the executive summary and sections 03 and 06, then the 90-day plan in section 17 and the appendix for your role.

Abstract

AI governance fails when risk, legal and security run separate processes: three reviews of three descriptions of the same system, on three calendars. This flagship paper sets out one integrated control framework for banks and insurers, written from the seats of the chief risk and compliance officer, the technology and regulatory lawyer and the CISO.

One AI product registry records every model, agent, prompt, tool, data product and vendor in use. One risk assessment sets a tier and gates every production release, with independent second-line approval for the higher tiers and no approval by agents. Guardrails, identity and monitoring enforce at runtime what the registry approved, and every decision leaves evidence. The three lines of defense run through every control point, with a RACI for intake, tiering, the release gate, runtime, incidents, change and retirement.

It maps US federal and state, EU, UK, Canadian and Singapore obligations to registry fields and controls, maps threats to the OWASP 2026 lists and MITRE ATLAS, and closes with how RJL’s products carry each control, a 90-day plan and six appendices: a registry template, a tiering questionnaire, a control matrix, a red-team plan, an incident playbook and a sourced register. It is not legal advice.

Decision it supports

How risk, legal and security approve and control AI as one process, and who in each line decides what.

What you will produce

A registry record and tier for one product, with its evidence by tier, named approvers and runtime checks.

See it worked through in a guide

Claims package 1.1.0 · 26 October 2026. Worked example: team A, group disability, region 2. Read-only preparation of the assigned case; examiners retain every case decision. Hold candidate 1.0.0 after an unrelated employee’s claim is returned. In the example decision, accept 1.1.0 for a monitored team A rollout only after prior-claims history is disabled and the remaining scope passes a fresh evaluation. All figures, evidence and sign-offs in that guide are invented; the paper provides the underlying method.

In RJL’s reference implementation

What RJL runs of this paper today.

  • Live No record, no production: every audit record carries the release ID, and a changed composition without a passed evaluation serves no agent.Section 04 · Release ID and policy hash on every audit record; a changed composition without a passed evaluation serves no agent See it working (invited sign-in)
  • Live The registry is the AI inventory, with a discovery check.Section 04 · AI inventory with discovery check and sealed export See it working (invited sign-in)
  • Live Independence rules are checked by the system, not left to memory.Section 03 · Decision-rights register with independence checks applied by every workflow See it working (invited sign-in)

Honest gap. Several second-line controls are built but run in warn mode rather than enforcing, and RJL’s own fallback model route was evaluated on 20 November 2026 and did not meet the threshold; fixes are being made and the route will be re-tested before it is relied on.

Status as of 20 November 2026, from the product records; statuses never exceed a recorded live check. See every row for this paper or the whole reference implementation.

In the paper

Three passages, as published

The rule

If it is not in the registry, it does not run in production.

Independence

Agents and automation can prepare evidence and block a release. They never approve one.

Fallback

A fallback that was never tested is a second incident.

Keep reading

Read next

Stay informed

Subscribe to whitepapers and updates

Get an email with the link when a new whitepaper or blog post is published, or a paper is revised. Subscribing is optional; every whitepaper remains free to read.

Next step

Discuss this paper against your own program.

Set up time with Richard to discuss your top gaps. The first conversation is free.