Syn2Real won CDISC's 2026 synthetic-data use case on architecture; its 4.5× editing-time claim rests on an undisclosed internal benchmark
Asked (summary)
Gather all the research on Use Case #1 of the CDISC AI Innovation Challenge — AI-enabled synthetic data generation for automation testing — won by Novartis and Amazon Web Services with Syn2Real. The requester supplied a product description covering USDM/CDISC inputs, per-record AI agents, deterministic validation, Admiral-based ADaM derivation, end-to-end traceability, a ~10-vs-45-minute internal benchmark, and a plan to sell or license the tool.
This dossier curates 50 evidence entries across many independent hosts — 5 on Syn2Real itself, 15 on the CDISC standards it builds on, 20 independent research and evaluation papers, and 10 adjacent tools. The largest host supplies 18% of rows, and three Zenodo URLs are versions of one record, shown here as a single evidence family. It is a curated sweep of thousands of web results, not a census of every publication on synthetic health data.
The Syn2Real pipeline: standards in, traceable synthetic SDTM and ADaM out
Pipeline reconstructed from the public Syn2Real record (Zenodo) and the requester-supplied description. Benchmark bars: minutes of manual editing before the same test automation ran — internal benchmark, methodology not public.
What the evidence actually supports
■ Public record confirms the architecture: per-record AI agents over an evolving subject narrative, deterministic validation against metadata, Controlled Terminology and study rules, and Pinnacle 21 plus natural-language checks — zenodo.org
■ The headline number — about 10 minutes of manual editing versus about 45 for other open-source generators (a 35-minute, ~78% reduction) — is an internal benchmark with no public sample size, comparator versions, error mix or replication — zenodo.org
■ A secondary review of the 2026 challenge, dated 2026-10-05, confirms the Use Case 1 win and reports no public Syn2Real source-code repository — jaimeyan.com
■ The plan to sell or license Syn2Real is requester-supplied: no public pricing, licensing terms, launch date or product page was found; CDISC's winner announcement itself names only the use case and the entrants — xn--druniespaa-19a.es
Evidence strength, standards context, and open questions
Verified public facts
Winner of Use Case 1, 2026 CDISC AI Innovation Challenge (Novartis + AWS)
Standards-native design: USDM and CDISC specifications are the inputs
ADaM derived from synthetic SDTM with reusable Admiral R code, not generated directly
Value-level provenance links each output to a spec, terminology, reasoning, validation evidence or code
Vendor and internal claims
~10 min vs ~45 min of manual editing (4.5× faster, ~78% less editing) — internal, methodology not public
Per-subject narratives keep subjects coherent across domains
Custom conformance rules can be written in natural language
Requester-supplied status
Planned sale or licensing — not corroborated by any public pricing, terms or product page
Unresolved questions
Benchmark sample size, comparator names and versions, study designs, error mix, hardware, statistical uncertainty
No public source-code repository found (review dated 2026-10-05)
Privacy, scalability, cost and production readiness have no independent evidence
Why standards-native matters, and what to demand before buying
Standards stack
USDM gives a machine-readable study definition the generator can execute against
SDTM structures collected data for submission; CDISC publishes conformance rules and traceability examples for ADaM (15 standards entries in this dossier, mostly cdisc.org)
Admiral's reusable, tested R functions make ADaM derivations reproducible rather than hallucinated
Fit-for-purpose evaluation checklist
Independent research says to evaluate fidelity, utility, privacy, clinical and rule validity, longitudinal and cross-domain consistency, robustness, equity and reproducibility — most published work fails this bar (e.g. SynTrustBench: 4 of 30 reports pass its evaluability gate)
For automation testing specifically: standards conformity, controlled fault coverage, deterministic reproducibility and reduced manual remediation matter more than matching population distributions
Risks
Synthetic data are not automatically anonymous, unbiased, statistically faithful or clinically valid, and sit in regulatory limbo
Generators can propagate and amplify training-data bias; cross-domain consistency is the historical weak point Syn2Real claims to fix
Alternatives, and one distinction
Adjacent tools: Leova (USDM-driven EDC testing, challenge runner-up), pharmaverse admiral/pharmaversesdtm, PHUSE Test Data Factory, Merck's LLM+knowledge-graph ADaM pipeline, and several open GitHub simulators
Do not conflate Syn2Real with Novartis's open-source synadam: synadam starts from real ADaM and preserves structure for de-identification; Syn2Real starts from USDM/specs and derives ADaM from synthetic SDTM
Practical verdict: a compelling, standards-native architecture for test-data engineering — but public evidence is presently too thin to independently validate performance, privacy, scalability, cost or production readiness. Commercial diligence should demand the benchmark protocol, comparator versions, a reproducible run, and licensing terms in writing.
All 50 evidence entries
Section
Item
Key finding
Caveat
Date
Source
Method: 50 curated evidence rows from a research sweep of thousands of web results across standards bodies, regulators, repositories, peer-reviewed literature and industry sources, compiled through 2026-10; four sections (Syn2Real evidence 5, standards 15, independent research 20, adjacent tools 10). Largest single host contributes 18% of rows; three Zenodo URLs are versions of one Syn2Real record and count as one evidence family. Benchmark minutes measure average manual editing before the same test automation ran, per an internal Novartis/AWS benchmark whose methodology is not public. Long findings and caveats are truncated for space; follow each source link for full text.