Application workflow
Purpose-built narrative, fact-review, instruction, policy-source, and warrant-preparation flows keep the user inside a defined task.
How LeoPen works
The model is one part
of the system.
Task-specific workflows, fact-preservation checks, official and agency sources, application rules, and review states determine how the AI service is used inside LeoPen.
Provider, hosting, retention, routing, and data-flow decisions are documented for the exact deployment under review.
System-level differentiation
What surrounds the generation step matters
Better documentation depends on how facts, sources, prompts, review states, and people are connected—not on claiming that one model always knows the answer.
Fact-preservation checks
The supplied fact set is compared with the working draft so omissions and unsupported language can be surfaced for review.
Official and agency sources
Supported Florida jury instructions and agency-provided policy material remain connected to their retained source documents.
Review-first guardrails
Product language, status states, and provenance keep uncertainty and human responsibility visible.
Request path
Generation stays inside a defined source-and-review loop
The model is one service inside the application. The workflow around it determines what context it receives, which source material is available, what checks run afterward, and who is responsible for the result.
- 01
Define the task
The user enters a narrative, fact-review, instruction, or policy workflow rather than an unrestricted general-purpose chat.
- 02
Preserve user facts
Officer-provided notes and statements remain the working fact set used to organize and review the draft.
- 03
Retrieve approved sources
When the workflow calls for it, LeoPen resolves supported official instructions or configured agency policy material.
- 04
Generate inside the workflow
The configured AI service produces task-specific language while application rules keep the output in working-draft status.
- 05
Run application checks
Fact comparison, reviewed instruction matchers, source links, and explicit manual-review states add context around the generation.
- 06
Return control to people
The officer and agency inspect the facts, wording, sources, prompts, and decisions before the work moves forward.
Architecture questions
Pin down the exact AI service used in the pilot
“AI model” is not enough detail for security, procurement, or operational review. The final architecture should answer these questions in writing.
Provider and version
Which model and provider serve each workflow, and how are model or routing changes approved and communicated?
Data exposure
Exactly which fields, documents, prompts, and metadata are sent to the provider, from which region, and for what purpose?
Retention and training
What does each provider retain, log, or use for training or service improvement under the proposed service tier?
Failure and fallback
What happens when a provider is unavailable, a request is interrupted, a source is missing, or output falls outside acceptance rules?
Evaluation
Which representative, incomplete, contradictory, and adversarial scenarios are tested before and during the pilot?
Change control
How are new models, prompts, source packs, matcher logic, and guardrails evaluated before they affect agency users?
Pilot risk management
Govern, map, measure, and manage the complete workflow
NIST's AI Risk Management Framework is a useful structure for evaluating LeoPen as a system across design, deployment, use, testing, and change—not as a one-time model demo.
Read the NIST AI RMFGovern
Assign accountable owners, approved uses, prohibited uses, review duties, escalation paths, and change authority.
Map
Document users, scenarios, affected people, data, sources, providers, dependencies, limitations, and potential harms.
Measure
Test fact preservation, source grounding, incomplete inputs, unsupported requests, reviewer behavior, reliability, and security.
Manage
Set acceptance thresholds, human controls, monitoring, incident response, remediation, retirement, and recurring review.
Human and legal boundaries remain unchanged
LeoPen supports drafting and review. It does not determine what charge applies, whether probable cause exists, whether evidence is admissible, whether an agency is compliant, or whether a person is guilty.
Evaluate the complete system proposed for your pilot.
Ask about workflow boundaries, model providers, source handling, data flow, deployment, evidence, and operational responsibility together.