CuRE Conduct

What sets it apart

• Configure a pipeline, not a services project — FHIR, CRO CSVs, registries, labs, and claims map to one canonical OMOP data plane.
• Vocabulary-as-a-service, and vocabulary-as-a-workbench: a governed CDISC Controlled Terminology register plus a concept API (search, hierarchy, multilingual synonyms, semantic search) that sibling apps read instead of forking a second copy of your standards — and a governed authoring layer where your terminologists write the extensions that register never had.
• Author terminology instead of filing a ticket for it: local concepts, typed properties, relationships, and source-code mappings are yours to write, each moving through draft → steward review → approval with a hash-chained audit trail, rather than waiting on a vendor's content team.
• Extend the standard without forking it: your assertions live in a tenant-scoped overlay that resolves alongside standard vocabulary in hierarchy closure, so upgrading to a new vocabulary release is a reconcile step against a release register — not a re-do of your local content.
• Your local extensions stay yours. Tenant-authored assertions are never published into the shared OMOP vocabulary and never reach cross-app views or the external WebAPI surface; only platform-curated and authority-ingested content is eligible for promotion into the canonical model.
• The standards/terminology plane of the platform Metadata Repository — CDISC CT, Biomedical Concepts, and the CDASH → SDTM → ADaM → define.xml traceability spine in one governed substrate.
• Governance is in the data plane, not an afterthought: row-level security + consent scope at every read.
• Research enclaves are the cross-org collaboration boundary — a first-class governed workspace (lifecycle-audited, snapshot-pinned) bound to a study or analysis project, where a collaborator reads a consent-scoped, DUA-gated slice of the record, never a copy of it.
• Versioned dataset releases, de-identification/PPRL, lakehouse deployment, and synthetic data run on the same governed plane.
• AI structural ETL authoring profiles an unmapped source and proposes the transform rules, grounded on the tenant's own accepted-mapping history and governed vocabulary.
• Mapping QA asks whether a concept is the RIGHT one, not just whether it resolves — a standing harness classifies every binding as unresolved, superseded, or semantically drifted, catching cases like a hemoglobin field bound to an arterial-blood LOINC that pass a resolve-and-validity check. Findings are reviewer-gated and never auto-applied.
• Migrate registry forms to OMOP progressively: parity-gated authority changes preserve field lineage and frozen historical report rows without an uncontrolled dual-write system.
• Workbench UI for data managers + service API for sibling apps — dual-nature by design.
• Cell-therapy products are first-class governed records: the graft itself — product identity, processing methods, and ISBT-128 identifiers — is a queryable OMOP-extension entity, not free text on a form (ADR-CON-061).
• CAR-T manufacturing and cold-chain data land on the patient's OMOP timeline and reconstruct back out losslessly, so release-checkpoint measurements, transit times, and temperature excursions align to outcomes (ADR-CON-038).
• Allogeneic donor↔recipient pairs are typed relationships on the EMPI substrate with their own DUA-gated disclosure scope — graft source, matching, and paired outcomes are queryable across the transplant, not a sidebar (ADR-CON-047).
• HLA is a first-class vocabulary plane, not free text: GL strings parse and canonicalize, ARS groups expand, serologic equivalents resolve, and every resolution pins to an IMGT/HLA release — transplant-registry machinery generic mappers don't carry.
• Registry crosswalks ship as governed, versioned mapping content — CIBMTR form families (2400/2402/2100, HLA-Save, AML/MM pathways) with concept-accuracy audits — not as a services engagement.
• Accuracy claims are scored, not asserted: a hash-sealed harness scores correct abstention as accuracy and over-reach as the costly error, so "no confident match" never silently becomes a wrong code.
• The accuracy evidence is visible in-product, not only in a report: steward-acceptance audits with a confidence interval, per-template quality rollups, and per-row confidence tiers sit beside the queue they describe — and a rate that has not been measured reads "not yet measured", never 0%.