CHANGELOG — SDTM AI Knowledge Base Release

v1.0 — 2026-04-27 (Company Release)

First public release for company-internal use.

Added

• 4-platform LIVE deployments: Claude Projects v2.6 / ChatGPT GPTs v2.2 / Gemini Gems v7.1 / NotebookLM (Custom mode instructions 8,925 chars)
• Trilingual user documentation (zh / en / ja): top README, USER_GUIDE, self_deploy/README, 4 × UPLOAD_TUTORIAL
• 10-question DEMO pack with multilingual prompts
• KNOWN_LIMITATIONS reference (en)
• 17-question smoke test framework v4.0 with 3 anti-hallucination probes (AHP)
• Per-platform decision tree (“which platform for what”)

Quality baselines (17-question evaluation, R1+R2)

• Claude Projects: 17/17 (100%) — all PASS
• ChatGPT GPTs: 16.5/17 (97%) — post v2.2 Q1 GFINHERT spelling fix
• Gemini Gems: 16/17 (94%) — post v7.1 Q10 SUPP— Core anchor; R2 dual-gate PASS
• NotebookLM: 15/17 (88%) — Q11/Q12 supplemental topics PUNT (correct), Q9 safety-correct PUNT

Methodology references

• 4 prevention rules (Rules A/B/C/D) governing all AI-assisted content work
• 28-slot independent reviewer chain (Rule D) across 17-question evaluation + V5C regression
• Per-platform Phase 0-5 lifecycle with sign-off (PHASE5_RETROSPECTIVE.md)

Source Knowledge Base

• 295 Markdown files derived from CDISC SDTMIG v3.4 + CDISC CT (NCI EVS) + SDTM v2.0 model
• 63 SDTM domains (full spec + assumptions + examples per domain)
• ~2,400K total tokens across all source files

Release artifacts

• ./ (this folder) — self-contained release: public-facing docs + per-platform deploy bundles under self_deploy/{claude,chatgpt,gemini,notebooklm}/, each with system_prompt.md (or instructions.md) + uploads/ + trilingual tutorial.{zh,en,ja}.md
• git tag: v1.0-company-release (created at Phase E end)

Known limitations

See KNOWN_LIMITATIONS.en.md for full list (L1-L3 cross-platform + L4 per-platform).

Pre-release (work history, abbreviated)

• 2026-04-23 to 2026-04-24: Phase 5 sign-off across 4 platforms (28th independent reviewer ack-ready)
• 2026-04-24: Gemini v7.1 LIVE (post Q10 PASS+) + ChatGPT v2.2 LIVE (post Q1 PASS)
• 2026-04-22: smoke v4.0 question battery design (3 + 14 + 3 = 20 evaluation files per platform)
• 2026-04-21: NotebookLM v2 architecture pivot (3-notebook → 1-notebook × 42 sources)
• 2026-04-20: Claude Projects v2.6 final (24/24 PASS, 1.29M tokens / 77% capacity)
• 2026-04-18 onwards: 4-platform parallel deployment program kickoff

Full development history (28-slot Rule D reviewer chain, R1/R2/PHASE5/V5C retrospectives, per-platform retros) is maintained internally and available from Bojiang Zhang on request.