Deterministic software handles
- Intake validation and completeness checks
- Routing and approval rules
- Permissions, inherited from source systems
- Every calculation in the ROI model
- Audit logging and retention
Home / Approach
Approach
Built so IT can say yes.
Deploy into the agency's own stack. Measure before and after. Keep the model to one bounded step, and a named employee on the signature. The architecture is the governance framework.
§ 01 Order of operations
Measure first. Scale only what worked.
The work runs in a fixed order, and each phase can stop with something the agency owns. If the procedure a pilot would follow isn't written down, writing it becomes part of the pilot.
- 3 weeks
Assess
Inventory the workflows, time them with the staff who do them, and score each one. Leave with one recommended pilot and its ROI case.
- 30–60 days
Pilot
One workflow in your tenant, with an evaluation set from real cases and a before-and-after metric defined at kickoff.
- 3–6 months
Deploy
SSO, permissions, integrations, audit logging, training and documentation, only after the pilot is measured.
- Ongoing
Operate
Evaluations on every change, monitoring, and change control, with the employee still on the signature.
§ 02 Deterministic by design
The model is a component behind gates, not the system.
Agencies have sat through a dozen AI demos and still have no clear first step. That's rational: adjectives can't be verified. In a Munivera system, ordinary testable software does everything except one bounded task.
Intake
Rules check the request is complete before work starts.
Retrieve
Existing permissions decide what may be read.
- The only AI step
Model
Extract, draft or summarize. One task, bounded.
Validate
Code checks the output against the source.
Approve
A named employee signs. The model never does.
The model handles
- Extracting fields from documents
- Drafting text for employee review
- Summarizing and citing source material
Always behind a validation check, and always ahead of a named signature.
§ 03 Architecture
Requirements we won't waive.
Hosting
Your cloud tenant: Azure and Microsoft 365, AWS or Google Cloud. Munivera doesn't host agency data.
Identity
Entra ID or your existing SSO. Role-based permissions inherited from source systems.
Data minimization
The model receives only the context the task needs.
Auditability
Every prompt, retrieval and output is logged and retrievable for records requests.
Employee in the loop
A named employee approves any consequential action: an award, a payment, a permit decision, a letter sent.
Retention
Follows the agency records schedule. No shadow copies.
Accessibility
WCAG 2.1 AA for any staff- or public-facing interface.
Evaluation
An accuracy test set built during the pilot, rerun on every model or prompt change.
Portability
The agency owns prompts, evaluation sets, configuration and documentation at exit.
§ 04 Assessment method
The same measurement, every time.
For each candidate workflow:
Annual volume × minutes per task × fully burdened labor rate × automation share
The result is framed as recoverable administrative capacity across the workflows we inventory, then narrowed to one recommended pilot. Each workflow is also scored on data readiness, integration complexity, records and privacy sensitivity, and political risk.
§ 05 Employee in the loop
The employee still owns the decision.
Awards, payments, permits and correspondence never leave without a named person. That isn't a slogan. It's how a warranted contracting officer designs an agent: the accountability has a name on it, and software doesn't get to dilute it.