MissionHarness.ai Analysis
First-party analysis of the mandates, memos, and acquisition rules that govern federal AI. Where Insights tracks what others report, Briefings is what we conclude — every claim sourced to the primary document.
VA designated its ambient documentation tool high-impact and applied the safeguards that follow. It did not designate the general-purpose chat tools staff were using to draft clinical notes, on the reasoning that those tools are analogous to search engines. VA's own Inspector General called the analogy flawed. The gap is not a VA problem. It is what happens when the unit of governance is a product rather than a use.
Every VA IT contract that touches sensitive personal information carries a liquidated damages clause: $37.50 per affected individual, uncapped, in place of actual damages, and the AI clause that now travels with it requires the substance to flow down to every subcontractor at any tier. On an enterprise AI platform that ingests health records, that clause is the largest financial exposure a reseller prime signs, and it is the one an OEM's commercial paper will not accept as written.
Between January 2025 and September 2026 the federal government issued six AI executive orders, five OMB AI memoranda, an AI Action Plan, the OneGov AI agreements and the first proposed contract clause written specifically for LLMs. The six years before that produced three orders and one memo. VA is the agency where the stack lands hardest: a public AI inventory that grew nine-fold in two years, most of it high-impact, and three solicitations in one summer that describe a single strategy.
Most of the coverage of the Logging Reference Architecture is about deadlines. Agency Logging Plans are due in November, maturity gates follow. Section 10 is the part worth reading. Where agencies point AI at their security telemetry, CISA now requires named human review points, documented fallback procedures, and a chain of custody that separates event records from whatever the model said about them.
GSA's proposed data-safeguarding rule for large language models (91 FR 36559) flows down the entire LLM supply chain through four role-specific sub-clauses. The comment window closed August 3, 2026. What remains is compliance posture: a training ban, 72-hour incident reporting, and automated controls keeping human eyes off government data.
Two OMB memoranda issued April 3, 2025 replaced the previous administration's federal AI policy in a single stroke. M-25-21 governs how agencies use AI. M-25-22 governs how they buy it. Together they are the baseline every federal AI program, and every vendor selling into one, is now measured against.
A word-rate comparison of eight federal AI mandates, five agentic-governance publications from three standards organizations, and the full text of the Governed Autonomy Doctrine. On almost every term that matters for governing an agent, the standards bodies sit between the mandates and the doctrine. It is a gradient, not a gap, and knowing which end of it a document sits on tells a program office how to read it.
OpenAI declared GPT-6 Astra the first model to cross its Critical cybersecurity threshold, restricted what the released version will do, and published evaluations from five external organizations. Two days earlier it committed a billion dollars to putting that class of model inside water utilities and the grid. A private release-gating apparatus now exists and works. No federal acquisition instrument reads its output.
OMB grades AI by the use case: a high-impact determination, an impact assessment, a waiver, a discontinuation order, each attached to one application. The Governed Autonomy Maturity Model grades the organization and lets the level license what it may deploy. Federal practice has the better instrument, and the doctrine's own Laws are the reason to adopt it.
A control model for autonomous agents does not reach a federal CIO as a set of laws and planes. It reaches them as a program: owners, dates, gates, a budget line, and a way to buy against it. OMB already wrote federal AI governance that way. The Governed Autonomy Doctrine deliberately did not. This is what has to sit on top of the doctrine before an agency can act on it.
Fail closed is the right default for an agent that cannot reach its governing system. It is the wrong end state for a benefits claim, a care referral, or a pharmacy order that was halfway done when the governor went dark. OMB asks agencies to assess the harm of not using AI. NIST has distinguished fail-safe from fail-closed for years. The doctrine's Law 5 asserts a right to intervene and says nothing about the obligation not to strand. At VA, that silence is the whole question.
NIST AI 600-1, the Generative AI Profile to the AI Risk Management Framework, is the federal government's most detailed catalog of generative AI risk: 12 risk categories, 212 suggested actions. It is also a document whose authorizing executive order has been revoked, whose home agency has been ordered to rewrite three of its risk categories, whose 212 actions carry no minimum tier, and whose threat model predates the agentic AI federal health programs are now deploying. Federal health components citing it today are citing guidance in visible transition, not a settled standard.
On August 27, 2026 a federal judge permanently enjoined the Department of War's supply chain risk designation of Anthropic, calling the measures illegal and baseless. Contractors sitting on purge deadlines should not read that as an all clear. The same judge issued a preliminary injunction in March, and purge memos kept arriving for months afterward. The gap between what a designation is legally worth and what your contracting officer told you to do is the thing to understand.
OMB's December 2025 memo implements Executive Order 14319's two Unbiased AI Principles (Truth-seeking and Ideological Neutrality) as procurement requirements for every LLM the government buys. Vendors selling LLMs into federal contracts now face contractual compliance terms and documentation expectations tied to how their models behave.
Briefings are original MissionHarness.ai analysis. Every factual claim links to a primary source — official memoranda, the Federal Register, and government publications. Doctrine terms link to the independent standard at governedautonomy.org.