Four worked scenarios, one per persona × vertical. Each shows the
Minimum-Dataset fields populated, the MCPs called, the agents
invoked, and the Figure 3 timeliness milestones triggered.
Scenario A — Hiker mails in a tick
Persona: Recreational hiker, age 38, has a tick attached after
a weekend hike near Patagonia (Santa Cruz County).
Channel: Mobile app.
Steps
User opens the app. “Submit a tick” flow.
Camera capture of the tick on a coin for scale. App-side
client-side ML estimates engorgement; user enters date attached,
approximate hours attached, location on body.
Intake Agent populates the Minimum Dataset:
General: age, sex, ZIP, GPS coordinates, date of report.
Exposure: tick bite, attached for ~6 h, leg.
Auxiliary: photo of tick.
Geo-Enrichment Agent resolves GPS → county.santa_cruz,
not on a reservation; pulls
vectorsurv-mcp.agency_region_intersect to find the responsible
county vector-control agency.
Validation Agent photo-quality check passes; species
tentatively identified as Rhipicephalus sanguineus (brown dog
tick) but with low confidence — flagged for Walker lab.
Triage Agent (VBD branch):
Matches vector.rhipicephalus_sanguineus → transmittedBy →
pathogen.rickettsia_rickettsii (RMSF) from
schema/deep/pathogens.sql.
No symptoms reported; triage class = mail-tick-to-walker-lab
self-monitor-for-14-days.
Enrichment Agent:
great-az-tick-check-mcp.create_submission(user, tick_meta)
returns a mailing label PDF.
vectorsurv-mcp.get_pools(arthropod="tick", county=Santa Cruz,
last 90 days) — nothing reported nearby, so no live signal.
Looks up schema/deep/outbreaks.sql for any active RMSF
event — historical tribal-lands cluster noted but not active
near Patagonia.
Persistence: observation node lands with edges to
pathogen.rickettsia_rickettsii, county.santa_cruz,
resource.ua_extension_tickcheck, and focus.rmsf.
Notification Agent: in-app card with the mailing label, an
ICD-10 reference (A77.0), a 14-day symptom watchlist, and a
one-tap “if symptoms appear” booking link.
Milestones triggered
Detect — at step 2 (community reported a tick exposure).
Verify (potential, deferred) — when Walker lab returns species
pathogen results, the Knowledge Update Agent ingests the
result and edges the observation to a lab_result node.
Scenario B — Symptomatic patient calls 211
Persona: Construction worker, age 52, Phoenix, fevers + severe
muscle aches for 3 days, no obvious bite memory.
Channel: Voice via 211 Arizona (transcribed by a downstream
voice-MCP).
Steps
211 operator opens the worker’s intake on the Sentinel
web console; voice transcript streams in.
Intake Agent extracts:
General: age, sex (M), Phoenix ZIP, occupation = construction.
Human: fever, severe muscle aches, mild headache, 3-day onset,
has not sought care.
Exposure: outdoor occupation, no remembered bite, no recent
travel.
Geo-Enrichment Agent:county.maricopa.
Validation Agent: no duplicate report; flagged as “symptomatic,
no exposure history — open differential.”
Triage Agent (VBD branch):
Symptoms match WNV, dengue (low likelihood given no travel),
and — importantly — leptospirosis (occupational exposure).
Calls vectorsurv-mcp.get_pools(target_acronym="WNV",
county=Maricopa, last 30 days) — WNV pool positivity at
8.2 per 1000 in nearby ZCTAs, well above threshold.
Calls knowledge-graph-mcp.outbreak_check("WNV", Maricopa) —
edges to a fresh outbreak.maricopa_wnv_2026 if one exists.
Triage class = urgent-care + clinician-alert-WNV.
Notification Agent:
Worker: “Given your symptoms and current WNV activity in
Maricopa, please go to an urgent care or ED today. Here is
the nearest one open now.” (One-tap call + directions.)
211 operator: same in real time.
Agency channel: an aggregated pin on the MCDPH analyst
dashboard, no PII.
Milestones triggered
Detect — symptom onset 3 days ago (computed from date-of-
illness field).
Notify — the agency dashboard pin is the local Notify event;
Maricopa Vector Control is paged if cluster threshold reached.
The interval Detect → Notify lands in the agent_run audit
table; this is the headline metric.
Scenario C — Unsheltered resident heat check-in
Persona: Unsheltered resident, no phone, downtown Phoenix on a
115 °F Magenta-HeatRisk day. CHW is doing block-by-block outreach
with a tablet.
Channel: CHW-mediated, on the CHW’s tablet.
Steps
CHW opens the Sentinel field app; “Heat check-in” flow.
CHW enters: approximate age, sex, currently unsheltered,
sweating profusely but mentation intact, no chronic medications
reported, last drank water ~3 h ago, current GPS.
Intake Agent suppresses Email, Phone, Household member ID
per the anonymous-outreach consent profile.
Geo-Enrichment Agent: ZCTA 85003.
Validation Agent: accepts; flags as
“high-risk-population observation.”
Triage Agent (Heat branch):
Computes vulnerability score: unsheltered (+3),
outdoor-exposure (+2), no AC (+2), Magenta HeatRisk (+3),
symptomatic (heat exhaustion: heavy sweat + headache, +2).
Total = 12 (max ~15).
Triage class = go-to-cooling-center + dispatch-CHW-transport.
Notification Agent:
CHW: card with the nearest center + a one-tap “request
transport” button.
Center operator: a heads-up message via mag-hrn-mcp
(peer-to-peer between operators).
Aggregated: observation lands in the Cluster Detection
Agent’s stream — five unsheltered heat-exhaustion check-ins
in the same ZCTA in 2 h triggers a county heat-emergency
alert.
Milestones triggered
Predict — the Magenta HeatRisk forecast 24 h prior was the
predictive alert.
Prevent — the CHW outreach happening in response to the
Magenta forecast is itself the Prevent action.
Detect → Respond interval is what this scenario optimizes.
Scenario D — Agency-side cluster review
Persona: ADHS Vector-Borne & Zoonotic Diseases epidemiologist,
end of week.
Channel: Agency web dashboard.
Steps
The Cluster Detection Agent ran nightly and flagged a
small but unusual cluster: 4 hantavirus-compatible observations
in Coconino County over the last 10 days, mostly from CHW
check-ins in tribal communities.
Notification Agent wrote a row to the ADHS analyst queue
(no PHI exposed to the wider system).
Epi opens the dashboard — sees the cluster overlaid on
the MapLibre map (every pin carries
kg_node_id; clicking opens the underlying observation
with its source MCP receipts).
Dashboard queries (plain SQL against the kg):
-- Co-located WHISPers wildlife mortality eventsSELECTw.label,w.description,p.value_textASevent_dateFROMkg.nodewJOINkg.propertypONp.node_id=w.node_idANDp.key='event_date'WHEREw.node_type='whispers_event'ANDw.node_idIN(SELECTobject_idFROMkg.edgeWHEREsubject_idIN(SELECTnode_idFROMcluster_observations)ANDpredicate='colocatedWith');-- Recent NEON small-mammal data near the clusterSELECT*FROMkg.v_datasets_apisWHEREresource_id='program.neon_srer'ANDinforms_questionsLIKE'%hantavirus%';
Epi clicks “open Verify workflow” — the Knowledge Update Agent
pulls the latest ADHS hantavirus-case report via adhs-mcp and
the latest NEON rodent-pathogen testing via a future
neon-mcp, both edged to the cluster.
Decision: field investigation dispatched. The dispatch is
itself a Verify milestone event, written back to the kg via
knowledge-graph-mcp.
The Figure 3 timeliness clock for this potential outbreak
now has Detect (community check-in), Notify (agency queue), and
Verify (field dispatch) timestamps. Lab Confirmation and
Respond will follow.
Why this is a “living database” scenario
Nothing in this flow was bespoke. The cluster detector ran on the
same observation table community reports land in; the dashboard
queries were the same recursive-CTE patterns used by the
knowledge-graph viewer; the MCP pulls were the same
ones the Intake Agent uses live. The agency epi got an end-to-end
investigative workspace without anyone writing a new ETL job.