ENTERPRISE DATA ECOSYSTEM • RESEARCH AND AI GOVERNANCE • OSONWANNE GROUP LLC
Cairn: one governed data estate for operations, research and AI
A concept demonstration with synthetic data, prepared in response to Roswell Park Comprehensive Cancer Center’s RFI #27-28, “Enterprise Data Ecosystem” (NYS Contract Reporter 2140335). It shows a certified data catalog with lineage, an honest-broker research workflow with study-specific IDs, and AI access that only reaches approved data.
Governance you can click through, on synthetic data
The catalog, the broker workflow and the AI policy checks are working logic with unit tests. Every asset, owner, count, MRN and query result is invented and labeled synthetic.
12
Data products in the synthetic catalog, each with an owner, steward, sensitivity label, lineage and quality score
6
Policy checks before an AI question becomes SQL, including a request for patient names that is blocked
30
Unit tests covering the catalog, lineage, honest-broker, policy and audit logic
0
Real patient records: the demo uses synthetic data only and calls no external service

The ideas behind it
Our response proposes a vendor-agnostic reference architecture that builds on what the cancer center already runs, in open formats, so no single proprietary layer becomes a dependency.
THE PROPOSED REFERENCE ARCHITECTURE: SEVEN LAYERS, EACH REPLACEABLE
7
Consumers
Power BI, financial planning, operational reports, research apps, custom apps, AI agents and EHR-launched apps
6
Research and AI zones
Honest-broker workspaces, the OMOP research model, study datasets, data science and the governed AI access layer
5
Governance and catalog
Glossary, lineage, quality, sensitivity labels, access requests, research restrictions and audit
4
Logical and semantic layer
Stable logical models (patient, oncology record, provider, encounter, claim, trial participant) and certified metrics
3
Medallion lakehouse BronzeSilverGold
Raw and immutable, then cleaned and PHI-classified, then data products and metrics, in open Delta tables
2
Access and ingestion
Virtualized access where data can stay at source; CDC, batch, HL7, FHIR and file ingestion where it must persist
1
Sources
The new EHR, legacy systems being retired, the existing warehouse, finance, research, registry, genomics and external data
Consumers bind to layer 4, not to sources, so systems can be retired or replaced underneath without breaking reports, research tools or AI.
THE QUESTION FOR EVERY DATASET
Does a performance, availability, governance or cost rule require a copy?
NO: VIRTUALIZE
Query it in place through the logical layer. No copy to keep in sync.
YES: PERSIST ON PURPOSE
Land it in the lakehouse, with the reason recorded on the dataset.
ARCHITECTURE
Fabric-native architecture
Bronze, silver and gold layers in open Delta tables on the organization’s existing Microsoft Fabric capacity, with Power BI semantic models on top. A separate federated engine is added only if discovery shows Fabric alone cannot meet a need.
DATA ACCESS
Virtualize first, persist on purpose
Data stays at its source unless a performance, availability, governance or cost rule says to persist it, and every persisted dataset carries its recorded reason. Consumers bind to stable logical models, so sources can change underneath.
DATA PRODUCTS
Certified data products
Each product has an owner, a steward, a schema contract, quality tests, documentation and an access policy before it is published, and moves through draft, review, certified, deprecated and retired.
RESEARCH
Honest-broker research workflow
IRB-scoped requests become study datasets with study-specific IDs: a keyed HMAC-SHA256 hash with a per-study secret. The same patient gets unlinkable IDs in different studies, and re-identification needs a logged approval.
ARTIFICIAL INTELLIGENCE
Governed AI access
AI applications and agents reach only allow-listed, certified products through one access layer that checks identity, policy and permitted use. It shows the generated SQL and provenance and logs every request.
CATALOG
Catalog with lineage
One catalog for data, metrics and research metadata: glossary terms, certification badges, sensitivity labels, lineage from source through bronze, silver and gold to use, and access requests from the asset page.

Research and AI on the same certified data

HONEST BROKERS
Release only what the IRB approved
The broker builds the cohort, generates study IDs, shifts dates by a per-patient offset of up to 30 days, and releases the most detailed shape that passes a k-anonymity check of five. Direct identifiers are never released.
ANALYSTS
Ask questions, not tickets
Natural-language questions are limited to certified semantic models. Each answer shows the data products used, the generated SQL and its provenance.
DATA GOVERNANCE
One audit trail
Every request, approval, query, identifier transformation, release and re-identification lands in one audit log that exports to CSV.
THE HONEST-BROKER FLOW
STEP 1
IRB request
Protocol, approved variables and criteria; anything outside the approval is refused.
STEP 2
Cohort
Criteria applied in order, with counts at every step, in a restricted broker workspace.
STEP 3
Study-specific IDs
HMAC-SHA256 with a per-study secret: the same patient is unlinkable across studies.
STEP 4
Audited release
Dates shifted, k-anonymity checked, released to the study workspace, every action logged.

What is real, and what is simulated
Real: the catalog, lineage, honest-broker, policy and audit logic (30 unit tests). Study IDs are genuine HMAC-SHA256 values computed in the browser with WebCrypto, and date shifts come from a second keyed hash. The k-anonymity check supports, but does not replace, a HIPAA Expert Determination, which an independent qualified expert makes.
Synthetic: every data product, owner role, column, count, MRN, study key, protocol and query result. Simulated: the AI step (fixed rules, no language model), access provisioning and the underlying data platform. The demo is a static web app that illustrates the proposed architecture; it does not run on Microsoft Fabric.
It is a concept built for a Request for Information, not a production system, not a product and not past performance.
See it running
No login: take the four-stop guided tour, trace any product’s lineage, run the honest-broker workflow, or ask the governed AI for something it should refuse. Press Ctrl+K or ⌘K to jump anywhere.
Boniface Osonwanne, CEO • bon@osonwanne.com • 602-632-9589
Cairn is an independent concept demonstration by Osonwanne Group LLC, prepared in response to Roswell Park Comprehensive Cancer Center’s RFI #27-28, “Enterprise Data Ecosystem” (NYS Contract Reporter 2140335). A Request for Information is market research only; no award or selection has been made. All demo data is synthetic and contains no real patient information, and AI output is simulated. It is not a Roswell Park system and not past performance. It is not affiliated with or endorsed by Roswell Park Comprehensive Cancer Center or any government agency.