🧪 CONCEPT BUILD — unofficial redesign specimen. The artifacts below are synthetic illustrations of what the proof layer would look like in production. Federal dataset names are real; values are invented.
The proof layer

Don't take our word for it. Take the receipt.

Four artifacts, one thesis: in a market newly suspicious of AI that sounds confident, the vendor that shows its work wins. Speed is the demo. Auditability is the product.

1 The data dictionary

Every variable the engine can draw on — source, vintage, refresh cadence, geography. Published at /data; every receipt cites rows by ID.

IDVariableSourceVintageRefreshGeography
PLACES-COPDCOPD prevalence, % adultsCDC PLACES2024 release (2022 est.)AnnualCensus tract
PLACES-CHFCoronary heart disease prevalenceCDC PLACES2024 release (2022 est.)AnnualCensus tract
ACS-NOVEHHouseholds with no vehicle, %ACS 5-year2023 (2019–2023)AnnualCensus tract
SVI-COMPSocial Vulnerability Index, compositeCDC/ATSDR SVI2022AnnualCensus tract
EJS-PM25PM2.5 percentileEPA EJScreen2024AnnualBlock group → tract
GEO-HOSP-DISTDrive-time to nearest acute-care hospitalCMS Provider Data + OSM2025 computationOn updateCensus tract
CMS-HRRPExcess readmission ratio, HF cohortCMS Hospital CompareFY2025AnnualHospital service area

Shown: 7 of 200+. SYNTHETIC EXCERPT Full table ships with Dataset schema for machine consumption.

2 The decision receipt

Ships with every answer. The one-page artifact that turns the engine's judgment into a negotiable, auditable object.

RECEIPT VHT-8F42-2606 ENGINE MCDA v0.9.2 ISSUED 2026-04-12 DATA BUNDLE dd-2026.04-a7c3
Question as typed
“Where should we pilot the mobile heart-failure follow-up team?”
Parsed intent
ComponentInterpretationConfidence
Target populationAdults with elevated CHF burden0.97
Access barrier weightingTransportation + pharmacy distance emphasized0.88 — flagged for review
Variables & weights
IDVariableVintageWeightRationale (human-edited)
PLACES-CHFCHF prevalence2024 (2022)0.28Primary demand driver
CMS-HRRPExcess readmission ratioFY20250.22Penalty exposure anchor
ACS-NOVEHNo vehicle access20230.18Follow-up feasibility
GEO-RX-DISTDrive-time to pharmacy20250.12Adherence proxy
Stability: top three tracts stable under ±20% perturbation of any single weight. Positions 4–6 sensitive to the transportation weight — full sweep in appendix export.
ACTOR: ENGINE · proposed weights ACTOR: HUMAN (J. Ruiz, Pop. Health) · approved 2026-04-12 ACTOR: HUMAN · adjusted no-vehicle 0.15→0.18

SYNTHETIC In production: “Adjust weights → re-run” — customer overrides logged as actor events, printed on the next receipt.

3 The ground-truth ledger

Predictions frozen at issue; outcomes appended as they land. A ledger with a miss in it is ten times more credible than a ledger without one.

IssuedClaim (frozen at issue)Outcome (appended)Status
2024-11Tracts 12 & 14: projected 30-day HF readmission cluster12-mo actuals: ED revisit +18% vs +6% system avg in flagged tractsVALIDATED
2025-02Transport-gap flags: missed-visit concentration, 5 tracts2.4× system avg in 4 of 5; 5th within noisePARTIAL (4/5)
2025-06Allocation rec: 60/40 split over 50/50Ops chose 50/50; counterfactual not testable — logged transparentlyNOT TESTED
2025-09Site demand forecast, 18-mo horizonFacility opens 2026-Q4PENDING

SYNTHETIC Rules: hash-frozen at issue, append-only, never edited. The NOT TESTED row is deliberate — honesty as a feature.

4 The worked example

One question, walked end-to-end: asked, challenged, re-run, put in a board packet, and closed out in the ledger six months later.

1 2 3 Composite: CHF burden · readmission ratio · transport access · pharmacy distance — ~3,150 affected adults in top three tracts (counts, not percentiles)

SYNTHETIC In the product: click any tract → variable contributions, ranked. Trivial math (the composite is additive), enormous trust payoff.

Procurement & security pack

Posture

We don't want your PHI

  • Analysis starts from public federal datasets, not patient records.
  • No EHR integration = no new doors into your clinical stack.
Infrastructure

Google Cloud, signed BAA

  • Vertex AI, Cloud Run, Cloud SQL — enterprise infrastructure under a Business Associate Agreement.
  • Patented geospatial analytics at the core; 343+ engineering iterations.
Governance

Agents propose, humans dispose

  • Every recommendation stages through a human review queue; below confidence threshold the system says “Review Required” instead of guessing.
  • Audit trail logs every actor (HUMAN / AGENT / SYSTEM) with state, confidence, and reward weights. Sample schema available on request.
For vendor risk

Answers in writing, fast

  • Security reviews move faster when the founder answers directly — architecture, BAA, data flows, no sales layer.
  • Questionnaire responses, references, and insurance certs: request the pack.