The 2026 CDISC AI Innovation Challenge.
CDISC put a set of open problems to the clinical-data community and asked what AI could actually do about them. Verdatic entered the first one — synthetic data for testing. The entry is a six-minute video, and it's on this page.
What the challenge asked for.
The problem CDISC put forward is one most study teams will recognize: there is never enough clinical data available for testing, preparing it by hand takes real effort, and without it, validating systems and workflows is hard.
Entries were submitted as a video of six minutes or less and judged on three things: how meaningfully AI was applied to a real industry problem, whether the technical approach could survive actual adoption, and — weighted heaviest of the three — how deeply CDISC standards were integrated and how far traceability ran across the data lifecycle.
What Verdatic entered.
A simulation that teaches itself what good data looks like — from the standard, not from a checklist someone wrote once and filed away.
It learns quality from the standard
Conformance and quality rules don't just grade the data after the fact. Verdatic turns them into constraints on how the data gets made in the first place, and when a rule keeps being violated, an agent adjusts the simulation parameters behind it so the next run produces fewer of them.
The referee is the official engine
The check isn't our reading of the CDISC rules — it's the official CDISC CORE engine running on the generated datasets. Its findings become a worklist with a recorded decision on each one and run-over-run deltas, so nothing gets quietly re-litigated.
Every parameter change on the record
Each adjustment is logged with the field it touched, the value before and after, the rule that prompted it, and the agent's own plain-language reason — and any of them can be reverted. The AI proposes, the standard constrains, and people confirm.
The entry, in six minutes.
One study, followed end to end: a study design goes in, synthetic data comes out and gets checked, and the programs that transform it are generated and run — all on synthetic data, so there's no patient information anywhere in it.
What you're watching
- A version-locked catalog. Every decision downstream resolves against a dated snapshot of the CDISC library — pin the snapshot and the whole run reproduces.
- One measurement, followed the whole way. Systolic blood pressure, from the raw source field to a mapped SDTM result — it reappears in nearly every screen.
- A treatment effect set as a rule. Blood pressure drifts down on the active arm and holds flat on placebo, visible in the preview before a single record exists.
- Data built to fail on purpose. From one seed, a clean study and a deliberately defective twin — so you can score your pipeline on what it caught.
- Programs, generated and run. The mapping comes from the standard's own metadata rather than code hand-written for this one study, and the generated SAS, R, or Python bundle is executed on screen.
- The official engine, then the trace. CDISC CORE checks what those programs produced, and a traceability report walks one value back through its mapping method, its specialization, and its biomedical concept to the locked snapshot it came from — with the AI's proposal and the human decision side by side.
Where this stands.
The approach behind Verdatic's self-correcting simulation was selected for presentation at the 2026 CDISC Interchange.
That's an acceptance to present — not an award, a ranking, or a result, and we're not claiming one. What the video shows is what the product does today; nothing in it is a mockup or a roadmap item.
Want to see the same workflow on one of your own study designs? That's what guided access is for.
Request accessOr email [email protected] — and if you're weighing this up for a vendor assessment, Trust & security says plainly where every control stands.