# StormTrust AI

Product design

This page describes the product model and requirements. Examples illustrate the model, not a guaranteed API or completed feature. Check the [application contracts](/stormtrust/apps/) and [dataset inventory](/stormtrust/product/platform-and-data/datasets-and-source-contracts/) for implemented scope. Use [operations guides](/stormtrust/engineering/operations/) for runnable procedures.

AI helps people interpret, search, summarize, prioritize, and operate StormTrust. It does not replace the canonical evidence graph.

## High-value uses

### Storm summaries

AI can explain what happened and why StormTrust assigns a confidence level.

### Property explanations

AI can explain why a property receives a particular priority.

### Action plan assistance

AI can suggest deployment zones within explicit capacity and policy constraints.

### Report assistance

AI can turn structured evidence into understandable narrative without changing the evidence.

### Natural-language search

AI can translate a request into a structured search. For example, a user could request commercial roofs above a size threshold that experienced estimated hail within a time range.

### Operational assistance

AI can identify assignments that need attention or opportunities approaching a policy deadline.

## Grounding

AI answers use StormTrust tools rather than model memory for tenant or storm state. Representative tools include:

```txt
searchStorms
getStormEvent
getEvidenceTimeline
searchProperties
getImpactAssessment
getActionPlan
searchOpportunities
getCampaignPerformance
```

## Output provenance

Generated content records the model, available model version, prompt or template version, tool calls, source object IDs, and generation time.

## Guardrails

AI cannot independently relabel modeled evidence as observed, certify evidence, alter provenance, fabricate weather data, confirm roof damage, bypass authorization, reveal protected contact data, bypass consent or suppression, activate paid advertising, or perform destructive changes outside an approved workflow.

## Evaluation

Evaluation measures factual grounding, unsupported claims, citation accuracy, policy compliance, recommendation quality, refusal correctness, latency, and cost. Historical storm fixtures support repeatable evaluation.