Government

Federal AI work has two audiences. One is the person who needs an answer today. The other is the reviewer who asks, months later, who did this and on what basis. Agency teams must satisfy both. The configuration ties every request to a named person, refuses controlled records and requires a citation before an answer ships.

Nothing proceeds without an attributable federal identity. The role check refuses a call that arrives without a user, an organization and a card backed authentication header. That single setting turns a shared assistant into a reviewable one.

Sensitive records handling fills most of the request side. Control markings and sensitive designations are refused by name, along with grand jury material, procurement sensitive content and source selection information. Redaction masks case and docket numbers, claim and policy numbers, taxpayer identifiers, benefit recipient names and personnel files. Read only research tools cover document, records request and regulation lookup, and a session may take three actions.

Citations are strict here. Sources are required, they must match, and an unverified or fabricated claim is blocked instead of flagged. Output scoring holds the answer to a combined 0.93, with accuracy at 0.95 and faithfulness at 0.95. A response under 300 characters or three sentences fails as well, and a failing answer is blocked outright.

Volume control and record keeping fill out the rest of this configuration. Automated traffic detection refuses a client above eight requests in five minutes, which keeps citizen facing surfaces from being scraped. Capture runs in redacted mode, includes blocked events, and stops the request on failure. Audit marking notes the evaluation and adds no gate of its own. Records schedules, authorization boundaries and agency approvals stay outside these settings and remain your responsibility.