Five phases from hackathon MVP to a 12-month pilot evaluation. Each phase ships shippable artifacts and is judged by an explicit success criterion tied back to the Figure 3 timeliness milestones.
Scope. One vertical (VBD), one happy path (mail-in tick), three MCPs, one agent pipeline.
| Ship | Owner |
|---|---|
schema/deep/application.sql — new node types: observation, symptom, exposure_factor, consent_profile, wearable_metric |
repo |
mcp/great-az-tick-check-mcp/ — submission tracking + species feedback (mock if Walker lab API isn’t ready) |
repo |
mcp/knowledge-graph-mcp/ — read-only kg query over DuckLake (regions_at_point, pathogen_by_vector, outbreak_check) |
repo |
app/ — Next.js / React Native shell with the tick mail-in flow end-to-end |
repo |
| Agents: Intake, Geo-Enrichment, Validation, Triage (VBD only), Enrichment (vectorsurv + greatAZtickcheck + kg), Notification (in-app + email) | repo |
Live integration with vectorsurv-mcp for nearby pool context |
already shipped |
Success criterion. A user can mail in a tick using the app in under 3 minutes from open-to-confirmation, with a mailing label and a 14-day symptom watchlist generated. The observation lands in the kg with edges to pathogen, county, resource, and focus nodes. Detect milestone fires automatically.
Scope. Add the Heat vertical and three more MCPs.
| Ship | |
|---|---|
mcp/nws-heatrisk-mcp/ — daily HeatRisk + active alerts |
new |
mcp/mag-hrn-mcp/ — Heat Relief Network locations + (where available) supply / occupancy |
new |
mcp/adhs-mcp/ — read-only ADHS arbovirus + heat-mortality summaries |
new |
| Triage Agent — Heat branch with the vulnerability-score model | extend |
| App — three new flows: CHW heat check-in, anonymous heat self-report, “where can I cool off?” | extend |
| Notification Agent — SMS fallback + Spanish localization | extend |
| Cluster Detection Agent (vertical-scoped, hourly during heat season) | new |
Success criterion. A CHW with the field app can take an unsheltered check-in and dispatch transport to a cooling center in under 90 seconds. A clustering of 5 heat-exhaustion check-ins in one ZCTA / 2 h triggers a county heat-emergency alert to MCDPH.
Scope. Engage tribal partners, add offline + low-bandwidth modes, expand the linguistic surface, and harden data-sovereignty guardrails.
| Ship | |
|---|---|
| MOU + DUA framework with ITCA-TEC | external |
Optional navajo-ec-mcp proxy, gated by tribal DUA |
new (gated) |
| Field app: full offline capture, sync-on-reconnect | extend |
| SMS-only flow for users with no smartphone (Twilio / agency short code) | new |
| Multi-lingual UI: Spanish, Diné Bizaad, Tohono O’odham (with native-speaker review) | extend |
| Validation Agent: row-level tribal-data suppression at write time | extend |
whispers-mcp and inaturalist-mcp for wildlife mortality + citizen-science context |
new |
211-az-mcp for utility-assistance + transport referrals |
new |
Success criterion. A pilot tribal community can use the app end-to-end in the local language without uploading any data to shared infrastructure unless explicitly opted in. Tribal-data suppression audit passes a tabletop review.
Scope. Real users, real reports, evaluation against Figure 3 timeliness milestones.
| Ship | |
|---|---|
| Agency dashboard for MCDPH, ADHS, AZGFD, Coconino HHS | new |
agent_run audit table + Figure-3-milestone joins |
new |
| Cluster-detection thresholds calibrated to historical AZ data (1993 Four Corners, 2021 Maricopa WNV, 2023/2024 heat seasons, 2025 Coconino plague) | calibration |
| Public-facing aggregated dashboard (“AZ One Health Today”) | new |
inaturalist-mcp + AZGFD wildlife-mortality cross-references on every wildlife observation |
extend |
Success criterion. During the heat season and the WNV season, the median Detect → Notify interval for reports flowing through the app is at least 30% shorter than the 2024 baseline for the same counties. CHW + 211 satisfaction surveys ≥ 4 / 5.
Scope. Statewide expansion, federated learning with tribal partners, formal evaluation.
| Ship | |
|---|---|
| 15-county deployment | rollout |
| Federated cluster-detection that lets tribal partners contribute model gradients without releasing line data | new |
| Wearable integration: HealthKit + Health Connect skin-temp, HRV, sweat-rate (for Heat) | new |
| Annual evaluation report scored against Figure 3 milestones, with sector-by-sector intervals | external |
| Open-source release notes + governance handoff | external |
Success criterion. Independent evaluation against the Figure 3 milestone framework shows measurable timeliness improvement for both verticals, with no significant disparities by county, tribal status, or demographic group.
These run across every phase.
consent_profile enforcement; immediate-revoke pathway for
any user.schema/deep/datasets_apis.sql
catalog).| Risk | Mitigation |
|---|---|
| Upstream API drift breaks an MCP integration | Versioned OpenAPI snapshots in each mcp/<name>/openapi/; nightly diff alerts; env-overridable paths so a fix is deploy-without-rebuild |
| VectorSurv credentials expire / rotate | MCP server runs token refresh; key rotation playbook documented in each .env.example |
| Tribal partners object to a feature | Governance board has hard veto; tribal-data MOU is opt-in per tribe |
| LLM hallucinates a triage class | Triage Agent’s output is gated by a deterministic rule layer that maps {pathogen, vulnerability, current-MCP-readings} to a small enumerated class — LLM can only choose from the enumeration |
| Cost / token blow-up under heat-season load | Per-agent budgets in the orchestrator; Intake uses Haiku exclusively; aggressive caching of MCP responses |