SDTM AI Knowledge Base — User Guide v1.0

1. What This Is (Project Background)

If you need to look up CDISC SDTM variable definitions / Core / codelist, flipping through SDTMIG v3.4 PDF + NCI EVS Browser usually takes 10+ minutes. This project organizes all that material and deploys it to 4 AI platforms (Claude Projects / ChatGPT GPTs / Gemini Gems / NotebookLM). You just ask in natural language and get answers with spec citations in 1 minute.

Technical background: SDTM (Study Data Tabulation Model) covers 63 domains + thousands of variables + extensive CT (Controlled Terminology). We organized CDISC SDTMIG v3.4 + v2.0 model + CDISC CT into 295 Markdown sources, then fed them to 4 AI platforms with prompt engineering. New to terms like RAG / system prompt / Core (Req/Exp/Perm) / Extensible / anti-hallucination probe? See ./GLOSSARY.en.md (1-page lookup).

2. What We Built (Technical Highlights)

We tested each platform with 17 representative SDTM questions, including 3 “deliberately wrong premise” anti-hallucination probes (testing whether the AI catches false premises rather than playing along). The 4-platform scorecard:

Highlights: v3.4 new domains (GF / CP / BE / BS), Timing rules, CT Extensible handling, SUPPQUAL scope, cross-domain death-date alignment, and 3 anti-hallucination questions (LBCLINSIG / Trial-Level SAE Aggregate / PF deprecated domain). Throughout, quality was enforced with 4 internal quality rules + 28 independent reviewers cumulative. Sources: ./CHANGELOG.md and ../../SMOKE_V4.md §3. Glossary: ./GLOSSARY.en.md

3. Which Platform Should I Use? (Decision Tree)

Short version: Not sure which to pick? Start with Claude Projects. Bringing colleagues along? Use ChatGPT GPTs. Worried about hallucinations? Use NotebookLM. For a detailed comparison see the “Four-Platform Roles” table in ../README.md.

4. Access Links for All 4 Platforms

4.1 Claude Projects (Recommended Starting Point)

• Access: Wait for Bojiang Zhang to add you to the organization — join via the email invite link.
• URL: claude.ai → Projects → “SDTM Knowledge Base” (Bojiang Zhang will send the specific URL directly).
• Subscription required: Claude Pro / Team / Enterprise.
• Best for: Precise variable Core + CT lookups, cross-variable reasoning (e.g., the PCTPT five-variable set), correcting wrong premises (e.g., SUPPTS).
• Not ideal for: Real-time lookups on FDA / Pinnacle 21 (verify manually at cdisc.org); very large batch domain comparisons.

4.2 ChatGPT GPTs

• Access: Bojiang Zhang shares the Custom GPT to your organization; click “Add to My GPTs”.
• URL: chatgpt.com → top dropdown → “SDTM Knowledge Base”.
• Subscription required: ChatGPT Plus / Team / Enterprise (Free tier is not supported).
• Best for: Full domain queries, team sharing, publishing to GPT Store via OpenAI review.
• Not ideal for: Multi-step reasoning (slightly weaker than Claude Projects); Free-tier accounts cannot access the GPT.

4.3 Gemini Gems

• Access: Bojiang Zhang shares via Google Workspace, or you can self-deploy (personal account).
• URL: gemini.google.com → Gems → “SDTM Knowledge Base”.
• Subscription required: Gemini Advanced (personal) / Google Workspace.
• Best for: Loading large amounts of context at once, cross-domain pattern comparisons, long sessions.
• Not ideal for: Personal accounts cannot share directly with a team (requires Google Workspace).

4.4 NotebookLM

• Access: Bojiang Zhang invites you to the notebook, or you can create your own (50-source cap).
• URL: notebooklm.google.com → “SDTM Knowledge Base”.
• Subscription required: NotebookLM Pro / Google Workspace.
• Best for: Strong anti-hallucination (audit / compliance use), inline citation lookups, preferring “I don’t know” over making something up.
• Not ideal for: Questions outside the 42 sources (real-time Pinnacle 21, breaking news) — it will decline to answer by design, not a bug.

5. 5-Minute Quick Start (3 Warm-Up Questions)

Open your preferred platform (Claude Projects is a good first choice) and ask these 3 questions in order. Compare your answers against the Expected answers in ./DEMO_QUESTIONS.md:

1. D0 (Warm-up): “What domain and variable is AESER in SDTMIG v3.4? What is its Core attribute? Which CT C-code does it bind to?” Expected: AE domain / Serious Event / Exp / C66742 NY {Y/N/U/NA}.
2. D1 (New domain): Copy the D1 question text from DEMO_QUESTIONS.md (EGFR / Exon 19 / dbSNP). Expected: Domain=GF; should return GFGENSR / GFPVRID / GFGENREF / GFINHERT.
3. D5 (Wrong-premise correction): “What is SUPPTS in the SDTM standard? Is QORIG required?” Expected: The model proactively recognizes that “SUPPTS does not exist in SDTMIG v3.4” and redirects to TSVAL1-TSVALn = PASS+.

Grading: All core facts correct (domain / variable / Core / C-code) = PASS. Proactively catches the wrong premise = PASS+. Follows the wrong premise and fabricates = FAIL.

6. Full Demo Package (10 Questions)

The complete 10-question set is in ./DEMO_QUESTIONS.md (questions in three languages + English grading criteria). 5-minute intro = D0 / D1 / D5; 30-minute full run = D0 through D9 (includes 3 AHP probes: D6 LBCLINSIG / D7 SAE Aggregate / D8 PF deprecated domain + the cross-domain ultimate challenge D9: AE/MH/CE + DS death-date alignment). After running, compare your results against the §2 baselines (17/17 / 16.5/17 / 16/17 / 15/17) to see how your instance performs.

7. Known Limitations (Frequently Asked Questions)

Full details are in ./KNOWN_LIMITATIONS.en.md. Summary:

• L1 — Incomplete QS codelist coverage: 296 long-tail questionnaire codelists (PROMIS / EORTC) are not fully expanded due to capacity constraints (Claude ~55.8%); the remainder links out to NCI EVS Browser.
• L2 — Large codelists stored as stubs: LBTESTCD (2,536 terms) and 5 other large tables are stored as stubs with pointers only — the model will not fabricate individual terms.
• L3 — No real-time web access: NotebookLM is strictly in-KB-only; breaking news and the latest Pinnacle 21 updates are outside its knowledge (it will PUNT). The other 3 platforms can browse the web but require you to enable that manually.
• Claude: Capacity is at 77%, close to the Pro soft ceiling; adding new files requires first downgrading lower-priority content.
• ChatGPT: Hard 20-file limit (currently using 9); long-tail chunk tables may be partially missed mid-section.
• Gemini: Personal accounts cannot share directly with a team (requires Google Workspace); the v7.1 system prompt must be pasted in full.
• NotebookLM: 50-source cap (currently 42); Q9 / Q11 / Q12 proactively PUNT — this is correct and safe behavior, not a flaw.

8. Feedback

If you find an error, hallucination, or off-topic answer: (1) Take a screenshot and save the full original question and AI response. (2) Note the platform and version (e.g., “ChatGPT GPT v2.2 LIVE 2026-04-24”) and the expected answer (citing the SDTMIG v3.4 section number or CDISC CT C-code). (3) Email Bojiang Zhang, file in the company issue tracker, or @Bojiang Zhang in the department group chat. Issues are consolidated in ./CHANGELOG.md and addressed in the next minor release.

9. Road Map

Short term (v1.0 maintenance): Collect feedback and fix SDTM content errors; quarterly v1.x minor releases. Medium term (Phase 7 — self-hosted RAG): Break free of the 4-platform capacity constraints, enabling all 295 files at full resolution plus complete QS codelist expansion. Long term: Keep pace with SDTMIG v3.5+ and extend coverage to ADaM and Define-XML.

v1.0 — 2026-04-27 — Maintained by Bojiang Zhang