INDUSTRY FIELD GUIDE
National security public AI evaluation and approval evidence
A public announcement may describe an evaluation effort without establishing acceptance, deployment or an operational outcome.
NAICS 928110 · National Security
A public announcement may describe an evaluation effort without establishing acceptance, deployment or an operational outcome. Public testimony by the Chief Digital and Artificial Intelligence Officer describes government AI programs, and Scale advertises public-sector model evaluation. Use a fictional administrative exercise to test these distinctions. Public CDAO written testimony · Scale public sector evaluation
Invent a benign office text retrieval benchmark with a scope statement, a report of a test run and no approval decision. Ask the model which work was documented and which claims remain unsupported. Score scope, stage and citation separately.
Use original examples or approved public materials. Exclude operational, classified, export-controlled and controlled unclassified records. Leave access, testing authorization and deployment decisions to the agency.
Describe a synthetic public evaluation example
Original Fiund exercise
A public test report does not establish acceptance or deployment
A benign fictional office benchmark retrieves original form titles from a paper catalog. It uses no operational, security-sensitive or restricted material.
| Record | Source | What it records |
|---|---|---|
| NE1 | Practice evaluation scope | Test E1 covers title retrieval from ten original dummy documents in a detached fixture. |
| NE2 | Public-style test report | Records one E1 run with eight of ten exact title matches. Report says acceptance review pending. |
| NE3 | Announcement draft | Says the office is exploring evaluation. No acceptance decision, deployment authorization or operational-outcome record supplied. |
Question for the model
Which evaluation work and score are documented, and what evidence would be missing from a claim of accepted or deployed capability?
Show the reference answer and checks
NE1 and NE2 document one narrowly scoped E1 run with eight exact matches out of ten dummy-document questions. NE2 explicitly leaves acceptance pending, and NE3 adds no deployment or operational evidence. The supported statement is the reported test and its limited score, not accepted capability, authorized deployment or real-world effectiveness. Keep scope, measured result and approval stage separate with their sources cited.
- Treating an exploratory announcement or test completion as acceptance.
- Generalizing a small fixture score to operational performance or deployment authority.
Fields to preserve
- Benchmark scope and fixture
- Defines what was tested.
- Run result and scoring rule
- Makes the limited metric inspectable.
- Acceptance and deployment authority
- Preserves stages absent from the report.
Test another case
Add an acceptance decision limited to continued lab evaluation. Require the model to distinguish that authorization from operational deployment.
Use original benign examples only. Exclude classified, export-controlled, controlled unclassified and operational records; actual access and testing decisions stay with the agency.
Discuss a contribution
Describe the sources you control, the task they could support and what has already been checked. Use the existing-data assessment for an archive or the collection brief for a new recording project. The contributor agreement defines Fiund’s licensing authority and compensation. A guide or fictional exercise does not establish an accepted collection or a sale.