The Cleared FDE Standard.
"Forward-deployed engineer" is having a moment. In government, the words are cheap and the seat is not. This document is the public, versioned definition of what a cleared forward-deployed AI engineer must actually be able to do, published so that mission owners can hold anyone who uses the term, including us, to the same bar.
One engineer. Both disciplines. Inside the boundary.
A Cleared FDE is a security-cleared engineer who practices security engineering and AI engineering as one discipline, embedded with the mission owner's team inside the customer's approved environment, and personally accountable for a workflow from candidate architecture through operational acceptance. All three conditions are load-bearing: remove the clearance and the engineer cannot reach the mission; split the disciplines and the program dies in the handoff; remove the embedding and the acceptance test gets written for the user instead of with them.
The term deliberately stops at "cleared." Levels, programs, and access are validated per role, contract, and site, in an appropriate setting, never claimed in public materials. A practitioner who advertises more specificity than that is failing the Standard's own operations-security clause (§2.1).
What the seat must hold.
A Cleared FDE demonstrates working competence, not familiarity, in all five domains. Three of five is two specialists and a translation layer, which is the failure mode this Standard exists to prevent.
Works productively inside customer-approved environments, including disconnected and air-gapped networks: offline packaging, transfer evidence, deterministic builds, and the discipline to say "cleared" and nothing more in public.
Selects, integrates, and evaluates approved models against the mission's actual data; writes evaluation cases before capability; measures instead of demos; and can defend every model decision to a technical review board.
Maps the authoritative system for every governing fact; builds bounded, read-only-first connectors with provenance; and can tell a mission owner which of their systems does not belong in scope, out loud, early.
Designs fail-closed control flows: named human authority, held consequential actions, full evidence ledger. Every released output traceable to its sources; every action answerable to "who asked, who approved, what changed."
Writes the acceptance test with the end user; trains the owning team to operate without the FDE in the room; and treats operator adoption, not delivery, as the definition of done.
The loop, and the refusals.
A Cleared FDE runs every workflow through one auditable loop, Assign → Ingest → Reconcile → Hold → Release, where the mission question, owner, and acceptance test are named before work begins; sources are connected read-only with provenance; conflicts are surfaced rather than silently merged; consequential actions pause for the named authority; and approved outputs ship with their evidence attached.
The Standard also binds by refusal. A conforming practitioner refuses unbounded autonomy; refuses to demonstrate on customer data in public; refuses to sell research capability as delivered product; and refuses the endless pilot, every engagement carries acceptance measures and kill criteria agreed in writing before work begins.
Conformance is measured, not claimed.
An engagement meets the Standard when four things are true at day 90, in writing:
The workflow runs in the customer's environment, under the customer's controls, on the customer's data.
The value baseline was captured in week one and the agreed measures are met, or the kill criteria have been invoked and the customer keeps the connectors, evaluation cases, and data map.
Every consequential action in the ledger shows a named human approval, and no action exists outside the ledger.
The owning team operates the workflow without the FDE in the room.
Using this Standard
The Cleared FDE Standard is published by Federal AI and free to reference with attribution, in evaluations, solicitations, and teaming discussions, including to hold Federal AI itself to account. Cleared FDE™ is a trademark of Federal AI; a U.S. trademark application is pending before the USPTO. Version 1.0, August 2026; revisions are versioned and dated on this page. Feedback: lc@federal.ai.
Bring this Standard to the briefing.
Thirty minutes with a technical lead who is measured against this document. Ask for any section, applied to your workflow.