govFMP: prompt quality assurance and governance, inside your boundary
Model Single-tenant, self-hostedData path to vendor NoneIntegration One base-URL change
Staff already use AI. govFMP puts a measurable, logged control between them and
whatever model your agency has approved, so every prompt is scored,
cleared, and recorded before its output is trusted. It runs entirely
within your network. The vendor never sees a prompt.
How it sits in your network
Staff tools
ChatGPT, Copilot, agency chatbot, IDE
→same API
govFMP gate
score · clear · log
→pass / review
Your model
Azure OpenAI Gov, Bedrock, or local
▲ block (enforce mode) → a clear refusal, logged with its reason, and it never reaches the model
Agency network
Everything above lives here. govFMP holds the only network route and the only model credential,
so a prompt cannot reach a model except through the gate.
🔒 Enforced by your controls: egress rule makes the raw model endpoint unreachable except from govFMP · model key held only inside govFMP
No outbound connection leaves the boundary except to your approved model. The vendor has no telemetry, no console, no access.
What it does per prompt
Quality score, 0–100. Seven weighted criteria (goal, audience, format, constraints, context, tone, examples). Below the agency's floor, a rewrite is offered.
Security clearance: PASS / REVIEW / BLOCK. A separate check for protected data going into the wrong model (PII, CJI, PHI, CUI), for prompts trying to get around the model's controls, and for requests after genuinely harmful instructions. It judges intent and the person's role against the agency's own policy, and it does not screen on the subject.
Immutable audit record. Who, what, score, clearance, reason, rubric version, model, timestamp. Append-only and hash-chained, written to your sink. This log is what an agency shows when asked how AI use is controlled.
Deployment and control
▸ Runs where you say
OCI container in your VPC / GovCloud subscription or on-prem. Air-gap capable, offline license, no phone-home.
▸ Your identity, your logs
SSO/SAML/OIDC, PIV/CAC where required. Scores and audit records write to your datastore. The vendor has no read path.
▸ Your rules, versioned
Mode, thresholds, prohibited data categories, role exceptions, and rubric "packs" are declarative config under your change control. The vendor cannot alter them remotely.
▸ Monitored in your stack
/metrics for Prometheus, decisions streamed to your SIEM (Splunk, Sentinel), plus a governance dashboard with an over-block watch.
Compliance posture
Not a SaaS data processor. Single-tenant and in-boundary with no vendor data path, which removes the specific authorization gate that blocks hosted services. It rides your model's existing authorization (Azure OpenAI Gov, Bedrock GovCloud) rather than needing its own.
Auditable both ways. Every block is logged with a reason and is appealable, so the control can be reviewed for over-blocking as well as under-blocking.
Section 508. Accessibility conformance is in scope for any deployed interface.
Adopting it
Pilot. Point one team's tools at govFMP in observe mode. It scores and logs, blocks nothing, and produces the audit trail from day one.
Enforce. Once the log shows the policy is right, flip to enforce for the routes that need it.
Extend. Add rubric packs per use case (public comms, benefits, records, law enforcement) as your program matures.