The 2026 CDISC AI Challenge: Winners, Open Code, and a Synthetic ADaM Showdown

The 2026 CDISC AI Innovation Challenge results landed days ago, and the winners are presenting at the CDISC US Interchange in Denver today, October 5th. Fifty submissions from 30 organizations — more than double the 22 from 2025 — across three use cases: AI-enabled synthetic data for automation testing, AI-driven SAP generation, and AI-driven TFL generation. If you care about where AI and clinical data standards actually meet in production, this is the closest thing to a survey of the state of the art.

I did two things with the results: read everything public about the six winning and runner-up solutions, and then went looking for code. That second search turned up exactly one installable artifact adjacent to the winners — Novartis’s open-source synadam R package — so I read its source too. What follows is a practitioner read on all of it.

TL;DR — The 2026 winners validate directions we took in our PhUSE 2025 ML12 paper: standards-native generation, ARS as the intermediate layer, evidence records over agent counts. The open-source pickings are thin — none of the six solutions published code. The one package that exists, synadam, is not the winning Syn2Real; it’s a much simpler de-identification tool with a “glimpse-then-simulate” mechanism that preserves ranges but not distributions. synadam vs. our ML12 pipeline is an apples-to-oranges comparison — and making it precisely is the useful part.

The 2026 winners at a glance

Per the CDISC challenge page, judging weights standards integration and traceability most heavily — a signal worth remembering.

Table 1: 2026 CDISC AI Innovation Challenge results by use case.

TLFGenix’s architecture deserves one more sentence because it’s the most fully described: SAP digitization → ADaM profiling → ARS-aligned analysis spec → R driver scripts → network-isolated sandbox execution → schema/consistency checks. Statisticians approve the spec before any codegen and sign off each TLF; a lineage graph enables selective regeneration when the spec changes; every TLF ships an evidence record.

Where is the code?

Nowhere, mostly. I checked the GitHub API today: of the six 2026 solutions, none has public code. SAP-Genie, TLFGenix, Syn2Real, PAIR, Leova — no repos. Smart ClinSAP has only a 19 KB documentation repo.

The single installable artifact adjacent to the winners is Novartis/synadam: an open-source R package on CRAN, MIT-licensed, created June 2026, roughly two stars at time of writing. Important: synadam is not Syn2Real. Syn2Real is the USDM-driven, standards-native framework that won Use Case 1; synadam is a separate, much simpler Novartis package in the same problem family. Conflating them would be a mistake, so I’ll keep them strictly apart here.

One channel worth watching: COSA, the CDISC Open-Source Alliance. The 2025 open-source winners were invited to contribute through it, and CDISC’s own cdisc-org repos — including the Analysis Results Standard — are already open. If any 2026 code surfaces publicly, that’s where it will likely land.

synadam under the hood

Since synadam is the only thing you can actually install.packages() today, I read the R source. The mechanism is “glimpse-then-simulate”: a family of glimpse*() functions extracts per-column summaries from real ADaM data, and simulate*() draws synthetic values from those summaries.

library(synadam)

# Scan real ADaM datasets, build a YAML study config (seed included)
yaml_path <- generate_study_config(
  adam_dir   = "path/to/adam",
  output_dir = "./syn_data",
  seed       = 42
)

# Deterministic simulation; syn_*.rds files land in output_dir
simulate_study(yaml_path)

Listing 1: synadam quickstart. The zero-config scan plus seeded simulation is the whole workflow.

What it preserves:

• Structure and ranges. Numeric and date columns are sampled uniformly between the observed min and max.
• NA positions, if you ask it to mirror them; IDs are regenerated sequentially.
• Exact treatment + flag combination counts, and Y/N proportions on flag columns.
• User-declared joint combinations via ordered_col_sets (e.g. REGION1/REGION1N stay paired).
• A BDS spine: per-subject PARAM/visit profiles are resampled and crossed with synthetic ADSL rows — ADSL-first propagation, so cross-dataset joins work.

What it does not preserve:

• Marginal distributions. runif between min and max means a bimodal or skewed lab value becomes flat. Only the range survives.
• Category frequencies. Character columns are uniformly resampled over unique values — an 80/20 sex split drifts toward 50/50.
• Cross-PARAMCD plausibility. AVAL min/max is computed across all PARAMCDs mixed together, so a lab test can draw values from another test’s range.
• Correlations and clinical constraints. Nothing enforces AVAL vs. BASE vs. CHG coherence or AESTDY ≤ AEENDY unless you manually declare the relationships. There is no clinical-constraint layer at all.
• No built-in fidelity evaluation — no KS tests, chi-square, or divergence metrics to tell you how far off you drifted.
• A privacy caveat: the observed min/max are embedded in the summary, so the extremes of your real data leak into the synthetic output.

To be fair: for its actual use case — de-identification, when you have real ADaM and cannot share it — this trade-off is defensible. You’re not asking it to invent a plausible trial; you’re asking it to scramble one you already have. And the engineering is clean: S3 dispatch, checkmate input validation, seeds for determinism, YAML study configs, a version attribute on outputs, CRAN distribution. No LLM cost, no nondeterminism.

synadam vs. our ML12 pipeline

Our PhUSE US Connect 2025 paper ML12 (“A Novel Pipeline for Generating Realistic Synthetic CDISC ADaM Datasets Using Large Language Models and Knowledge Graphs”, Yan & Su, 2025; PDF) attacks the opposite problem: you have no real data yet — study-build testing, pipeline development before first patient in — and must generate from the spec alone. Spec → JSON schema → knowledge-graph enrichment (protocol/SAP/CRF queried block-by-block to inject ranges, derivations, allowed values) plus rule-based restructuring into ADSL/BDS/OCCDS → Faker-based, ADSL-first generation → formal evaluation via a composite Q score over ~13 sub-metrics (KS tests, chi-square, JS divergence on AEDECOD, AVAL–CHG correlation, temporal validity, cross-dataset USUBJID consistency).

Table 2: synadam vs. our ML12 pipeline. Same problem family, opposite starting conditions.

Two honest observations. First, the shared design: both systems independently arrived at ADSL-first generation with subject propagation into BDS/OCCDS. When two teams with opposite inputs converge on the same spine, that’s probably a real constraint of the data model, not a coincidence. Second, the candid side: synadam cannot do clinical realism or self-validation — it was never meant to. And ML12’s relationship score is its own weakest metric at every level (0.38 direct / 0.58 KG-enhanced / 0.65 template), temporal validity in OCCDS sat at 50–65%, and the whole thing degrades if the input documents are thin. Neither tool is a synthetic-data panacea; they bracket the problem from opposite ends.

What the challenge means for our pipeline

Four concrete takeaways for our own stack:

1. ARS is the emerging industry intermediate layer. SAP-Genie emits ARS-conformant SAPs; TLFGenix builds ARS-aligned analysis specs; PAIR consumes the same standards chain; and CDISC’s ARS repo is actually open. Our Mock Shell Generator already exports cards-style ARD and tfrmt JSON — but ARD is not ARS. An ARS-conformant export for our spec/binding layer is now on the roadmap, not the wishlist.
2. Borrow TLFGenix’s evidence record + lineage graph. Human approval of the spec before codegen, sign-off per TLF, and selective regeneration when the spec changes via a lineage graph — that is exactly the pattern a spec-change impact analysis needs, and it maps directly onto our deterministic-IR design.
3. Leova’s adversarial patients are the most transferable idea. Synthetic “rule-breaker” records pushed through an EDC to verify edit checks is closed-loop testing in its purest form. We can generate boundary-violating records — AESTDY after AEENDY, out-of-range AVALs, orphaned USUBJIDs — specifically to stress our own data-readiness checks. Offense testing your own defense.
4. Deterministic + human-gate remains defensible. TLFGenix has six agents; agent count is not the metric that matters. The judges weighted traceability highest, and the runner-up in every use case had explicit human gates. Our Mock Shell Generator’s confirm queue (BYO-key LLM suggests, never auto-writes) and the gxptlf engine’s locked-reference replay with independent recompute sit squarely on the right side of that line.

Key takeaways

• The 2026 CDISC AI Challenge drew 50 submissions from 30 organizations; winners present today in Denver. Standards integration and traceability were the top judging criteria.
• None of the six winning/runner-up solutions published code. The only installable artifact nearby is Novartis/synadam — which is not the winning Syn2Real.
• synadam’s glimpse-then-simulate mechanism preserves structure, ranges, NA patterns, declared joint combos, and exact treatment counts — but not marginal distributions, category frequencies, cross-PARAMCD plausibility, correlations, or clinical constraints, and it embeds real min/max values (a privacy caveat). For de-identification, that’s a defensible trade with clean engineering behind it.
• synadam and our ML12 pipeline solve opposite problems — de-identify data you have vs. generate data you don’t — yet both converged on ADSL-first subject propagation. Each is weakest exactly where the other doesn’t try: synadam has no fidelity metrics; ML12’s relationship score is its own lowest component.
• ARS is the de facto intermediate layer of the 2026 winners; TLFGenix’s evidence-record-plus-lineage-graph and Leova’s adversarial rule-breaker patients are the two patterns worth borrowing now.