AI FactSheet · GovAI Coalition template
Development status. govFMP is in active development and has no production deployments. This FactSheet is filed so that agencies evaluating it can do so on accurate information rather than on a sales description.
Fields marked none yet are ones where the honest answer is that the work has not been done. They are marked rather than omitted, because a reviewer who discovers an unmarked gap is right to distrust everything else on the page. Each will be filled in as it becomes true, and this document is the version of record.
GDM Digital Media Group. Carolina Shores, North Carolina, United States.
govFMP.
Brief summary of the AI system.
govFMP is a self-hosted control that sits between agency staff and an approved language model. It evaluates each staff-authored prompt before the model call, scores it for quality, screens it against the agency's acceptable-use and data-sensitivity rules, and writes an immutable record of the decision to a log the agency owns.
It is deployed as a single-tenant container inside the agency's environment and calls a model endpoint the agency configures. No prompt, output, or log is transmitted to the vendor, and the vendor operates no service in the data path.
What function does the AI system perform, and for what purpose? If the system performs multiple functions, list each discretely. For configurable features, describe configuration options and default settings.
Four discrete functions:
1. Prompt quality scoring. Scores a prompt against a rubric of seven quality axes and returns a score with a per-axis breakdown. Purpose: consistency of AI-assisted work across staff of varying skill. Configurable: rubric packs are agency-authored and versioned; the default pack is the published PQA-1 standard.
2. Acceptable-use and security screening. Evaluates the prompt against the agency's rules, including a data-sensitivity dimension mapping data classes to permitted models and roles. Returns a clearance with a stated reason. Configurable: the agency defines data classes, role mappings, and rules. There is no vendor-defined default policy, because acceptable use is jurisdiction-specific.
3. Rewrite suggestion. Returns an improved version of a prompt that scored poorly. Configurable: can be disabled. Default: enabled, advisory only, and never applied automatically.
4. Audit logging. Writes a record of every evaluation to an agency-owned sink. Configurable: retention period, and whether the full prompt or only its hash is stored. Default: hash-only, which is the more privacy-protective setting.
Operating mode spans all four. Observe scores and logs without blocking. Enforce additionally refuses prompts the rules reject. The default is observe, deliberately, so an agency sees real traffic before any rule becomes restrictive.
What domain is the AI system intended to be applied in?
Governance of general-purpose generative AI use by employees of public-sector organizations: state agencies, cities, counties, K-12 school districts, and higher education.
It governs staff use of AI. It is not intended for, and must not be used as, a component of any automated decision affecting a member of the public. It makes no determination about any individual, and it should not be placed in a benefits, enforcement, employment, or eligibility decision path.
How was the AI system trained? What data was used? How often is data added to the training set? Was all training data legally obtained and fully licensed?
govFMP does not train, fine-tune, or host a model, and it has no training set. This is the most important answer on this sheet and it changes several answers below.
Scoring and screening are performed by rule evaluation combined with calls to a language model the agency selects and configures. The model's training data is therefore the responsibility and the choice of whoever supplies that model, and it should be assessed through that provider's own documentation.
Agency prompts are never used as training data by the vendor. The vendor has no data path and is technically incapable of collecting them. Whether prompts may be used for training by the configured model provider depends on the agency's contract with that provider, which the agency controls. Agencies commonly point govFMP at an endpoint with a zero-retention agreement, and that is the recommended configuration.
The rubric content shipped as a default is the PQA-1 standard, authored by GDM. It is original work, published openly, and free to adopt.
What data was used to test system performance? Under what conditions has the system been tested?
Limited Scoring has been exercised against prompt sets constructed in-house, covering the categories the rubric addresses and a set of deliberately non-compliant prompts written to test the screening gate.
Two limits should be stated plainly. The test material is synthetic and vendor-authored, so it reflects what the authors anticipated rather than what public employees actually type. And there has been no testing against real agency traffic, because there are no deployments yet. First deployments are expected to run in observe mode partly for this reason: it produces the real-world test data that does not currently exist, under the agency's control.
General description of the model(s) used.
Two components:
Deterministic rule evaluation for structural checks and for the data-sensitivity and role gate. Not machine learning. Given the same prompt and the same configuration it returns the same result, which is what makes the gate auditable and appealable.
A large language model, supplied by the agency, for the judgment-based portions of quality scoring and for rewrite suggestions. govFMP is model-agnostic and is normally pointed at whatever model the agency has already approved. The vendor supplies no model and does not require any specific provider.
How often are models updated? Will the user have a choice about moving to an updated model? What documentation is available for new versions?
The agency controls model updates entirely, because the agency owns the endpoint. The vendor cannot change which model an agency uses, and cannot push a model change.
For the software itself: releases are versioned, distributed as container images, and installed by the agency on its own schedule. Nothing auto-updates. An agency may stay on a version indefinitely, subject to the support terms of its agreement. Release notes accompany every version, and any change to default scoring or screening behaviour is called out explicitly rather than left in a changelog, because a silent change to a governance control invalidates the audit record's comparability over time.
Rubric packs are versioned separately and every audit record stores the pack version in force at the time, so a historical decision can always be interpreted against the rules that actually applied to it.
What are the inputs and outputs? What interfaces and integrations are supported?
Input: a staff-authored text prompt, plus the authenticated user identity supplied by the agency's identity provider and, where the agency configures it, a data sensitivity classification.
Output: a quality score with per-axis breakdown; a clearance decision with a stated reason; optionally a suggested rewrite; and an audit record. In enforce mode a rejected prompt is refused and the refusal reason is returned to the user.
Interfaces: an MCP server so the control is available inside AI assistants staff already use; a reverse-proxy mode that sits in front of the model endpoint and requires one base-URL change; and a direct API. Integration with the agency's identity provider is required for user attribution in the audit log, which is what makes the log evidentiary.
What are the performance metrics? What is your current level of performance? How can the user monitor performance in the deployment environment?
none yet No field performance figures exist, and none are quoted here. There are no production deployments, so any accuracy, uptime, or satisfaction number on this sheet would be invented. Agencies should treat a pre-deployment vendor quoting such figures as a warning sign.
The metrics that will be reported, once there is real traffic to report on: screening precision and recall against agency-reviewed samples, over-blocking rate in enforce mode, evaluation latency added to a model call, and availability.
Monitoring does not depend on the vendor. Because the audit log is written to the agency's own sink, the agency can compute all of the above from its own data, without asking us and without trusting a number we supply. Over-blocking in particular is measurable directly: every refusal is logged with its reason and is appealable, so the rate and the pattern are visible to the agency at all times.
What biases does the tool exhibit and how does it handle that bias?
Three distinct sources, worth separating because they have different owners.
Inherited model bias. The judgment portions of scoring call the agency's chosen language model, so that model's biases are present in govFMP's output. govFMP does not remove them and does not claim to. An agency's model selection is therefore a bias decision, not only a procurement one.
Rubric bias. A rubric encodes a judgment about what a good prompt looks like, and that judgment was made by its authors. The default pack was written by a small company in the United States in English. Prompts written in another language, in a regional register, or by someone whose professional idiom differs from the authors' may score lower for reasons that have nothing to do with quality. This is why rubric packs are agency-authored and replaceable rather than fixed, and it is a real limitation rather than a hypothetical one.
Screening bias. A sensitivity gate can systematically burden roles that legitimately handle sensitive data, such as caseworkers, more than roles that do not. Observe mode exists partly to surface this before enforcement makes it consequential.
Handling. Every decision is logged with its reason, which makes disparate effects measurable rather than anecdotal, and rules and rubrics are agency-editable so a measured problem can be corrected without waiting for a vendor release. No claim is made that the system is unbiased.
How does the system handle outliers? Do overwritten decisions feed back into the system to help calibrate it?
Prompts that fall outside what a rubric anticipates receive a low-confidence score, which is surfaced rather than hidden, and an unclassifiable prompt is not blocked in observe mode.
Overridden decisions do not automatically retrain anything, and this is deliberate. Automatic calibration from overrides would let a governance control be gradually loosened by the people it governs, with no record of the policy having changed. Overrides are logged as override events, and they are reported to whoever administers the rules, who may then change a rule as an explicit, versioned decision. Calibration happens through a human changing a rule on the record, not through the system quietly moving.
What conditions does the model perform best under? Minimum requirements for quantity of records?
Best case is text prompts in English for drafting, summarising, and analysis, where the agency has defined its data classes and has a rubric pack matched to the work.
No minimum record count. Each prompt is evaluated independently, with no training threshold and no cold-start period, so the audit log is complete from the first day rather than after a learning window.
What conditions does the model perform poorly under? Limitations? What errors can it make and what makes them more likely?
Screening is not a guarantee. It reduces the rate at which sensitive material reaches an inappropriate model. It will not catch everything, and it should not be presented to a governing body as though it does. Sensitive information described obliquely, embedded in a long document, or expressed in domain shorthand is materially harder to detect than a recognisable identifier.
Both error directions are real. False negatives let a prompt through that policy would have refused. False positives block legitimate work, which is the more corrosive failure, because staff who are blocked without cause route around the control and then nothing is logged at all.
Inherited errors. Rewrite suggestions come from the agency's language model and can be wrong or fabricated in the ordinary way. They are advisory and are never applied automatically.
It governs one path. govFMP covers the model endpoints it is placed in front of. A staff member using a personal AI account on a personal device is outside its reach, and that gap is a device and network management problem rather than something this product solves. Any vendor implying otherwise is overstating.
Errors are more likely with very short prompts, mixed-language text, heavy jargon, and unusually long inputs.
How does the system explain its predictions? Are outcomes understandable by subject matter experts, users, impacted individuals, and others?
Every score is returned as a per-axis breakdown rather than a single opaque number, so a user can see which axis cost them and why. Every clearance decision returns the specific rule that produced it, in the agency's own wording, and rule-based decisions are deterministic and reproducible.
A staff member sees the reason at the moment of the decision. An administrator can reconstruct any past decision from the audit record, including the rubric version in force. Because govFMP makes no determination about any member of the public, there is no impacted-individual explanation surface, and it should not be deployed anywhere that would create one.
Considerations relevant to local, state, industry, or other jurisdictional regulations.
Public records. Prompts and audit records held by an agency are agency records and may be subject to that state's public records law. Retention is configurable specifically so an agency can align it with its own retention schedule, and the hash-only default exists so an agency can hold a complete decision record without holding prompt text it does not want to retain.
K-12. Student data handling falls under FERPA. Districts typically classify student records as a data class barred from external models, which the sensitivity gate is designed to express.
Regulated data. Because the deployment is self-hosted and calls an agency endpoint, govFMP introduces no new external data path for CJI, PHI, or CUI. It does not by itself confer authorization for any of them; that depends on the agency's environment and its model provider.
Data residency and sovereignty follow the agency's own infrastructure, since nothing leaves it.
List data protection frameworks you comply with and certifications held.
Designed against: the NIST AI Risk Management Framework. The product's core deliverable, an attributable per-decision record, is intended as evidence for the human-oversight and monitoring functions that framework describes.
none yet Certifications held: none. GDM holds no SOC 2, no FedRAMP authorization, no HIPAA attestation, and no CJIS certification, and no third-party audit of the company has been performed. Stating this plainly is more useful to a review team than a page of framework names, and a vendor of this size claiming otherwise should be checked.
The architectural point that matters more for a review: the vendor holds no agency data at all, so the vendor is not a processor in the deployment and there is no vendor environment for an agency's data to be assessed in. What must be assessed is the agency's own environment and its chosen model provider.
How is the tool monitored to identify problems in usage? Can outputs be overwritten by a human, and do overwritten outputs help calibrate the system?
Monitoring is performed by the agency against its own audit log, which contains every decision with its reason. False positives, false negatives, and quality complaints are all traceable to specific logged decisions.
Every output can be overridden by a human. Scores are advisory. Rewrites are suggestions. In enforce mode a refusal is appealable to a designated administrator, and the whole system can be returned to observe mode at any time by the agency without vendor involvement, which matters: a control an agency cannot switch off is a control it will eventually remove entirely.
Overrides do not automatically recalibrate the system, for the reason given under Robustness. They are logged and reported so a human can change a rule deliberately.
Rubric packs and screening rules are agency-authored, versioned, and editable without a vendor release, so a measured problem can be corrected by the agency directly. Every decision is logged with its reason and its rubric version, which is what allows bias to be measured by role, department, or time period rather than argued about. Observe mode lets an agency quantify effects before any rule becomes consequential.
There is no parameter that toggles model bias, because the model belongs to the agency. The available lever is model selection, which the agency holds.
none yet No. Neither the vendor nor an independent party has conducted a bias, accuracy, or disparate impact study, so there is no study for an agency to review. There are no deployments, and therefore no real-world data that such a study could be conducted against.
What is offered instead, stated as an intention rather than an accomplishment: the audit log is designed so that an early agency, or a third party it appoints, can perform exactly this analysis on its own data without vendor cooperation and without taking the vendor's word for the result. An agency willing to run that analysis on a pilot would be doing work we would want published, including if it is unflattering.
How can the agency and its partners flag issues related to bias, discrimination, or poor performance?
Within the agency, every refusal is appealable to a designated administrator and every decision is retrospectively reviewable from the log, so past system actions can be audited without vendor involvement.
To the vendor, through the contact form at gmediagroup.us/contact. Reports of bias, discriminatory effect, or unsafe behaviour are treated as defects rather than as feature requests. There is no requirement to send us data in order to report a problem, and no contractual restriction on an agency publishing what it finds.
How has the HCI aspect been made accessible, such as to people with disabilities? Has it been assessed against usability standards, and what was the result?
none yet No formal accessibility assessment has been conducted, so there is no WCAG conformance level or VPAT to report and none is claimed. An agency with a Section 508 or WCAG 2.1 AA procurement requirement should treat this as an open item.
Two structural points are true today. Most staff interaction happens inside the agency's existing AI assistant through the MCP interface or transparently through the proxy, so the primary surface is software the agency has already assessed. What govFMP returns is text, which inherits the accessibility of whatever surface displays it.
A conformance assessment and a VPAT are committed deliverables before any general availability release, and this entry will be updated with the result rather than removed.
Relevant information, links, or resources regarding your organization's responsible AI strategy. URL to any broad AI policy or strategy.
Published and freely usable, including by organizations that never become customers:
Prompt QA Governance Standard (PQA-1) — the full scoring standard the product applies, including its authorship and provenance: gmediagroup.us/standard.html
Technical summary — deployment model, data flow, integration: gmediagroup.us/govfmp-brief.html
Security and data handling — where data goes across all our products, including what we cannot see and what the product cannot do: gmediagroup.us/security
Product overview — gmediagroup.us/govfmp
The position underneath all of it: a governance product that overstates its own coverage makes an agency less safe, because it converts an unmanaged risk into a documented false assurance. This FactSheet is written on that basis, which is why the gaps are marked instead of skipped.