The Doctrine Is a Standard. Agencies Fund Programs.

Read OMB M-25-21 as a document type rather than as policy and something stands out. It is not a framework. It is a program plan. A Chief AI Officer within 60 days. A governance board chaired at the Deputy Secretary level within 90. A public AI strategy at 180. Updated IT, data, cyber and privacy policies at 270. Minimum risk-management practices for high-impact AI documented at 365, with the instruction that a use case which cannot meet them must be safely discontinued. An inventory every year. Owners, clocks, consequences.

Now read the Governed Autonomy Doctrine. Five Laws, five Architectural Planes, six Framework Pillars, a three-level Maturity Model, a proposed conformance layer. No owners. No clocks. No budget cycle. No sentence that tells a Deputy Secretary what to decide before the end of the quarter.

That is not a defect in the doctrine. It is the doctrine doing its job. A standard has to be stable, neutral and citable, and the moment it starts naming dates and roles it stops being one. The Framework page is explicit that it sequences what to build; the RFC is explicit that the doctrine names no mechanism because mechanism decays fastest. Those are correct choices for a published standard.

They are also the reason a federal CIO cannot act on it. Agencies do not adopt standards. They fund programs, and a program has a shape the doctrine does not supply.

What a federal program looks like

Every successful federal control regime of the last twenty years arrived in the same four layers, whether or not anyone designed it that way.

An executive layer. One document a Deputy Secretary can read in twenty minutes that says what the problem is, what the agency is going to do about it, who owns which piece, and what the first decision is. M-25-21’s main memo is three pages. The Appendix is where the requirements live. That split is deliberate.

A phased program with gates. Not a maturity model. A calendar. Phase one produces a baseline and an inventory; phase two proves the control set on a small number of bounded use cases; phase three turns the proven pattern into a service other components onboard to. Each phase has an exit condition somebody signs. The GAO accountability framework is organized the same way: governance, data, performance, monitoring, each with practices an auditor can check, because GAO writes for the people who will be audited.

A procurement method. M-25-22 exists because M-25-21 on its own would have produced requirements no contracting officer could put in a solicitation. Capabilities have to be stated in a form that can be evaluated, and vendor claims have to be graded by the evidence behind them, not by the confidence of the slide. FAR Part 39 has required that discipline for IT since before anyone said “agent.”

A worked example. One concrete system, threaded through every section, so the reader never has to imagine what the abstraction means for their own environment. The CISA and NSA agentic guidance is stronger for the scenarios it walks through than for the control list it ends with.

The doctrine supplies none of the four. It supplies the thing the four layers are built on top of, which is the harder thing to write and the wrong thing to hand a CIO on its own.

Where the gap actually bites

Three places, in our experience.

The first is the decision-rights question. The doctrine says governance must span systems (Law 3) and that policy lives in an engine outside the agent (Plane 3). It does not say who owns the policy when the CAIO, the CISO and a component’s mission owner disagree, or who may stop a running agent versus who may restore it. An agency cannot deploy against an architecture that has no answer to “who signs.” M-25-21 answers it at the top (CAIO, governance board, risk-accepting official with a signature) and leaves the operational layer to the agency. Somebody has to fill that layer in, per agency, before the first agent runs.

The second is the fiscal cycle. A control program that cannot be expressed as FY27 and FY28 line items does not exist to an appropriations staff. Token spend, inference placement, the cost of a governed transaction versus an ungoverned one: none of this is doctrine, all of it decides whether the doctrine gets funded. The doctrine’s own budget-adjacent construct, cumulative limits on what an agent may do, is the same mechanism a program office needs for cost. One meter, two audiences.

The third is assessment. The doctrine’s conformance layer is proposed and not in force, and it says so. An agency cannot wait for it. What an agency can do today is what NIST SP 800-53A has always done: separate the requirement from the assessment method, state preconditions and expected outcomes, sample, and keep the evidence. The doctrine’s Laws are the requirements. The assessment procedures have to be written by whoever is running the program, and they have to be written before the vendor demonstrations, not after.

What we take from this

The temptation, when a standard does not reach the decision-maker, is to add to the standard. More laws, more planes, an implementation guide bolted to the side. That would be the wrong fix. It would trade the doctrine’s neutrality, which is the property that lets an agency cite it in a solicitation without endorsing a vendor, for a program layer that has to be different for every agency anyway.

The right fix is to keep the doctrine as the citable standard and build the program layer separately, per agency, in the agency’s own vocabulary and on the agency’s own calendar. Executive explainer, phased roadmap with signed gates, procurement method with graded evidence, one worked example. Cite the Laws on every control the program claims to implement, so that the program can be checked against something that will not move under it.

That is the work. The doctrine tells an architect what must be true. A program tells a Deputy Secretary what to fund, and by when, and how they will know it worked. The two are not in competition. One is unreadable to the other’s audience, and a practice that serves federal customers has to write both.