The Policy Chain: From Executive Order to Contract Clause
The October 20 memo is the third layer in a policy stack that begins in July 2025.
Executive Order 14319, "Preventing Woke AI in the Federal Government," signed July 23, 2025, directs agency heads in Section 3 to procure only LLMs developed in accordance with two Unbiased AI Principles. Under truth-seeking, LLMs are to be truthful in responding to prompts seeking factual information or analysis, prioritize historical accuracy, scientific inquiry and objectivity, and acknowledge uncertainty where reliable information is incomplete or contradictory. Under ideological neutrality, LLMs are to be neutral, nonpartisan tools that do not manipulate responses in favor of ideological dogmas, which the order names as including DEI; separately, developers are not to intentionally encode partisan or ideological judgments into outputs unless those judgments are prompted by or otherwise readily accessible to the end user. Section 4(b)(i) directs agencies to include contract terms requiring compliance and providing that decommissioning costs are charged to the vendor if the agency terminates for the vendor's noncompliance following a reasonable period to cure.
Section 4(a) of the order directed the OMB Director to issue implementing guidance within 120 days. OMB issued Memorandum M-26-04, "Increasing Public Trust in Artificial Intelligence Through Unbiased AI Principles," on December 11, 2025, signed by Director Russell T. Vought. Its core requirement is that any solicitation or order for procurement of an LLM issued after the memo's date include contractual requirements addressing compliance with the principles. Agencies had until March 11, 2026, to update their procurement policies, including a process for agency users to report violating outputs.
M-26-04's scope limits matter to a defense audience:
- National security systems (as defined in 44 U.S.C. 3552(b)(6)) are outside its requirements, although applying the memo to them "to the extent practicable" is encouraged. Footnote 5 says those systems are governed through other policy.
- Open-source LLMs acquired under a free, open-source license are not required to receive the memo's contract requirements, but agencies should perform due diligence on their alignment with the principles before use.
- Sunset: the memo ceases to have force two years after issuance unless the OMB Director provides otherwise.
The Tenaglia memo translates M-26-04 into DoW acquisition mechanics. It states that compliance with the EO and the OMB memorandum is mandatory, while the attached templates are "guidance only" that components and requiring activities may revise, tailor or supplement to fit the complexity and risk of a requirement.
Three Attachments, Three Acquisition Stages
The guidance is organized into three attachments, each tied to a point in the procurement lifecycle.
Attachment A: Unbiased Artificial Intelligence Principles Memorandum. The requiring activity signs a memorandum classifying the requirement as one of four types: a large language model, a generative AI model other than an LLM, discriminative AI, or none of these. It then marks which transparency documents apply and sends the signed copy to the contracting officer with the requirement package. For an LLM, the four minimum items must be included; for other generative or discriminative AI, including them is at the signer's discretion. Our reading is that this forces a documented determination before Section L is drafted, so proposal instructions trace to a deliberate decision rather than boilerplate.
Attachment B: Contractor Required Information. The contracting officer copies the selections into Section L (or equivalent) proposal instructions. The template states that an offeror's failure to provide the required information "may result in the proposal being ineligible for award," and Attachment A indicates the documentation is evaluated in Section M. Missing documentation is therefore a proposal-eligibility risk, not a post-award paperwork issue.
Attachment C: Compliance Contract Language. This clause goes into the SOW or PWS in both the solicitation and the resulting contract. It defines the principles and "significant modifications," makes the documentation in the proposal the compliance baseline, and states that compliance "is a material requirement of the contract." Attachment A ties materiality to contract eligibility and payment and to supporting termination for default when the contractor refuses to take corrective action.
What LLM Vendors Must Document
For a requirement classified as an LLM, the minimum threshold (drawn from M-26-04, Appendix A, Section 1.A) is:
- A copy of the LLM developer's acceptable use policy
- The most current model card, system card and/or data card
- A description of, and where available links to, end-user resources such as tutorials and developer guides
- A description of the mechanism through which the government can give feedback on outputs that may violate the principles
The requiring activity can also select items from an enhanced threshold, which the Section L template labels as applying to higher-risk requirements. The options include:
- Pre- and post-training actions affecting factuality and grounding
- System-level prompts that instruct the model how to respond
- The types of output restricted by content moderation and safety filters
- Use of red teaming to assess the model for bias
- Any training or development that occurred outside the United States
- Modifications made to comply with regulations of governments other than the U.S. federal government
- Results and methodology of the vendor's bias evaluations, and benchmark scores for bias, helpfulness, honesty or accuracy
- Enterprise-level controls, such as customizable system instructions, evaluation tools and source-citation features
- For vendors that are not the direct developer, disclosure of their own added controls, such as classifiers, system prompts, fine-tuning or filters
Because the selections are made program by program, there is no single universal checklist. A vendor that prepares only the minimum package may be unready for an enhanced request, while one that assumes the full enhanced list everywhere may overinvest. M-26-04 itself gives an example of tailoring: for an LLM that summarizes cybersecurity incidents for a security operations center, truth-seeking will likely be more relevant than ideological neutrality.
The Supply Chain Asymmetry
Attachment A repeats a limitation from M-26-04: the amount and type of information available "will vary depending on the contractor's role within the software supply chain and its relationship with the LLM developer itself, with more information generally being available from sources closer to the original LLM developer." M-26-04 adds that for resellers, integrators and platform operators, the availability of information and of direct product interventions depends on the actual developer's willingness to collaborate through the distributor.
A firm that trained its own model controls its training records, evaluation results and system prompts. A firm offering access to another company's model through an API or platform can generally provide only what that developer chooses to share, plus documentation of its own added layers. The templates do not relax the compliance obligation for integrators. In our assessment, an integrator's credible evidence is therefore bounded by what its developer will disclose, and a prime that has not asked its model provider what it will release for a DoW proposal has an open gap before the next solicitation.
Scope deserves a separate check. M-26-04 does not require its terms for national security systems, and the Tenaglia memo directs personnel to apply the guidance to "applicable procurements." Whether a given defense program falls inside or outside that boundary is a determination for the government, and vendors should not assume a classified or mission system is either covered or exempt without confirming it with the contracting officer.
Enforcement Mechanics: The 15-Business-Day Clock
Under Attachment C, when an LLM output appears to violate the principles, the contracting officer provides written notification. Within 15 business days, the contractor must investigate and give a written response covering its findings, whether there was a violation, the most likely cause, its proposed corrective actions, and its proposed timeline to complete them.
The consequence chain follows from materiality. Attachment A ties the requirement to eligibility and payment and to termination for default when a contractor refuses corrective action; EO 14319 separately calls for decommissioning costs to fall on the vendor after termination for noncompliance following a cure period. In our judgment, a response that cannot explain the likely cause of an output is weak footing in that sequence, even though the template ties termination for default specifically to a refusal to take corrective action.
Two further obligations apply. First, the contractor must provide updated documentation when significant modifications are made to the AI model. Attachment C defines significant modifications as changes by the contractor or subcontractor that could reasonably be expected to alter adherence to the principles, including changes to training data, model architecture, and filtering and content moderation. Integrators should establish how they will learn of such changes when a subcontracted developer makes them. Second, the contractor must insert the substance of the mandatory requirements in all subcontracts at any tier for the acquisition, development or integration of LLMs used or delivered in performance.
Hypothetical example: a prime delivers an analyst-support tool built on a commercial model obtained through a subcontracted platform provider. A user reports that a summary of a contested policy question omitted one side. The contracting officer issues notice. The prime now needs, within 15 business days, the prompt and output logs, the system prompt and filter configuration in force at the time, any developer-side model change in the period, and a corrective proposal such as a revised system instruction plus regression tests. If the subcontract did not carry the flow-down terms or an obligation to supply those records quickly, the prime is answering on someone else's schedule.
The Unanswered Measurement Question
The templates create clear acquisition mechanics, but none of the three documents, the EO, M-26-04 or the DoW guidance, defines a test, metric or threshold for determining whether a particular output violates ideological neutrality. The EO asks for transparency through disclosure of system prompts, specifications and evaluations rather than prescribing a measurement method, and directs OMB to avoid over-prescription.
That leaves a practical gap. A vendor can supply a model card, red-teaming description and benchmark scores, and the contracting officer may still lack an agreed way to decide whether one disputed output reflects an intentionally encoded judgment or ordinary model variability. Because notices start the clock after award, vendors without a predefined evaluation method will be building one under time pressure.
Our recommendation is to define the evaluation approach before bidding: the test sets, scoring criteria and investigation procedure the vendor will use, aligned to the documents the solicitation requests. The templates do not require this, but a defined process gives a contractor a consistent basis for its findings-and-cause response.
A credible alternative: for a low-risk requirement that requests only the minimum threshold, such as an internal summarization tool, a vendor may reasonably rely on the developer's published documentation plus a clear feedback channel and a lightweight investigation runbook. That calculus changes when a program office selects enhanced items, when the tool is public-facing, or when the vendor adds its own system prompts or filters, because the vendor then owns evidence the developer cannot supply.
What to Do Before the Next DoW LLM Solicitation
The October 20 guidance moves documentation into the proposal and sets a response clock after award. Four actions follow:
- Map your LLM supply chain. List the models you integrate, the developer of each, and what documentation each developer will release to you for a government proposal. Where it is thin, negotiate before a solicitation closes. Owner: capture and contracts lead.
- Assemble the minimum package. Confirm you can produce a current acceptable use policy, model or system card, end-user resources and a feedback mechanism for each offered model, and identify which enhanced items you could supply on request. Owner: product or engineering lead.
- Update subcontract templates. Add the substance of the Attachment C requirements for any subcontract involving the acquisition, development or integration of an LLM, including notice of significant modifications and timely access to logs and configuration. Owner: subcontracts manager.
- Write an output investigation protocol. Define how you will retrieve a specific output and its context, determine cause and produce a written corrective plan and schedule within 15 business days. Test it once before you need it. Owner: engineering, with contracts review.
The memo does not add a new technical requirement to the underlying models. It standardizes how DoW asks for evidence and enforces the existing principles through Section L, Section M and the SOW or PWS. Vendors that can answer those documentation questions at proposal time will spend less of the 15-day window reconstructing what their system did.
Spartan X's AI advisory practice includes AI assurance evaluation informed by its work on Arbiter, which is in development. That experience is relevant to the gap identified here: designing documented evaluation and output-investigation processes that a vendor can stand behind when a contracting officer asks what happened and why.
Sources and further reading
- DoW DPCAP, Guidance on Implementing Unbiased Artificial Intelligence Principles in Procurements, memo and Attachments A–C, October 20, 2026
- Executive Order 14319, Preventing Woke AI in the Federal Government, July 23, 2025 — whitehouse.gov
- OMB M-26-04, Increasing Public Trust in Artificial Intelligence Through Unbiased AI Principles, December 11, 2025 — whitehouse.gov PDF
- Pentagon issues guidance on AI procurements to support White House's 'truth-seeking' and 'ideological neutrality' directives — DefenseScoop, October 23, 2026
- Pentagon Issues Contract Templates to Enforce Unbiased AI Rules — ExecutiveGov, October 26, 2026
- In Bid to Ban "Woke AI," White House Imposes Transparency Requirements on Contractors — Crowell & Moring
