PIF AI Whitepaper

Appendix D: Changelog

Complete version history. Format follows Keep a Changelog.

v0.3 — 2026-07-06

Added

• Chapter 14 “Toxicology Safety Engine” (bilingual zh-TW + en) — NOAEL six-tier fallback cascade (numeric → read-across → structural similarity via RDKit Tanimoto ≥ 0.50 → qualitative authority → TTC → data_gap), Margin-of-Safety computation, dermal-absorption (DAp) correction (fixed oils 50%→5%, essential oils / small molecules as negative controls), fail-safe asymmetry doctrine (prefer false-positive over false-negative, hard gates override, reviews must state a reason), and the display-layer rule that every ingredient must show an explicit safety basis.
• Chapter 15 “Regulatory Correctness” (bilingual) — EU disclosure threshold vs concentration limit (positional parser stripping the disclosure clause; 62 rows recomputed preserving 7 true limits), TFDA/EU/CIR authority hierarchy (TFDA limit met → EU downgraded to advisory; TFDA silent → EU binding to prevent false-negatives), EPA ToxValDB backfill (sanity ceiling 10000, flag verdict fail-closed), ECHA C&L structured harvesting (1125 classifications → 421 genotoxic + 684 CMR CAS auto-blocked), heavy-metal-salt gap fix, and the “never create a false-negative when relaxing a limit” doctrine.
• Chapter 16 “Self-Driving Evolution & Computation-Basis Provenance” (bilingual) — agreement_rate → 1.0 definition, asymmetric learning (false-safe auto-tightens / over-flag requires human review), an honest boundary on autonomy without an SA sign-off feature, three genuinely autonomous blocks (deterministic engine + resolved-cache + active re-grounding), the tox_reference computation-basis SSOT, and adversarial red-teaming.

Fixed

• Chapter 13 footer “next” navigation link changed from Appendix A to Chapter 14.
• Top-level / zh-TW / en README table of contents, chapter count (13→16), Schema.org about / keywords, and citation version synced to v0.3.0.

v0.2 — 2026-04-30

Added

• Chapter 13 “Compliance Engine Deep Dive (Phase 22–23)” (zh-TW + en, both languages) — covers compliance capabilities accumulated across 11 sub-phases:
  • §13.1 Lifecycle 5-stage reorganization (pre_launch / development / manufacturing / post_launch / signing)
  • §13.1.2 Business-type × 16-item responsibility matrix (4 × 16 = 64 cells)
  • §13.1.3 Auto-derived 7-step workflow + outsourcing recommendations
  • §13.2 PIF version management V0/V1/V2/V3 snapshots
  • §13.2.2 Change detection (formula / process / packaging fingerprints, SHA-256)
  • §13.2.3 One-click V2/V3 drafting
  • §13.2.4 Document expiry auto-tracking (GMP / test reports / §12 reporting countdown)
  • §13.3 Cross-item lint engine — 14 rules R1–R14 (including R1 advanced filtering)
  • §13.4 Penalty mapping table (§22–§25 fine ranges, NT$30K to NT$5M)
  • §13.5 SA signature metadata (method / hash / ip / cert ref)
  • §13.6 14-page regulatory PIF PDF one-click generation (WeasyPrint A4 + Noto Sans CJK)
  • §13.6.6 E2E real-data dual-path validation (by item number / by stage)
• Appendix B new §B.1.9 PIF Compliance Engine endpoints (14 new endpoints):
  • Responsibility matrix: 2
  • Workflow & outsourcing: 4
  • V0–V3 version snapshots: 5
  • Cross-item lint: 1
  • Regulatory PDF: 2

Changed

• Chapter 12 footer “Nav” link from “Appendix A” → “Chapter 13”
• Appendix A header “Nav” link from “Chapter 12” → “Chapter 13”
• README TOC adds Chapter 13
• CITATION.cff version bumped from 0.1 to 0.2; date-released updated to 2026-04-30

Corresponding Commits

• Parent repo VincentLinB/pif: Phase 22.0 → Phase 23E (commits ac44591, 7618936, 8755f8d, 43d649e, 91190a2, …)
• Whitepaper repo: this v0.2 commit

Known Limitations

• Performance numbers cited in Chapter 13 (PDF 14 pages in 2-3 sec, PDF size 640-690 KB) are measured; latency of other Phase 22-23 endpoints has not been benchmarked
• Cross-item lint R1–R14 rule table is PIF AI’s interpretation of ITRI curriculum; other SAs may interpret slightly differently — discussion welcome via GitHub issues

v0.1 — 2026-04-19

Added

• First public draft, structurally compatible with baiyuan-tech/geo-whitepaper
• 12 main chapters:
  • Ch.01 Abstract and Core Propositions
  • Ch.02 Regulatory Background (Taiwan Cosmetic PIF)
  • Ch.03 The 16 PIF Items (deep analysis)
  • Ch.04 System Architecture
  • Ch.05 Frontend Stack
  • Ch.06 Backend Stack
  • Ch.07 AI Engine (including Claude Code co-development practice)
  • Ch.08 Database and Multi-Tenant Isolation
  • Ch.09 Toxicology Data Pipeline
  • Ch.10 Central RAG Integration (Scheme C+, with L1 Wiki + L2 Vector RAG)
  • Ch.11 Security Model
  • Ch.12 Roadmap, Deployment, and Open-Source Strategy
• 4 appendices:
  • Appendix A — Glossary and Acronyms (50+ entries)
  • Appendix B — Full API Endpoint List
  • Appendix C — References
  • Appendix D — Changelog (this file)
• 15+ Mermaid diagrams: ER, flowchart, sequence, state, gantt, pie
• AI-friendly structured data: YAML frontmatter per chapter + Schema.org TechArticle JSON-LD per README
• Chinese edition (zh-TW/) available in parallel with full chapter correspondence

CI and Tooling

• .github/workflows/build-pdf.yml: Pandoc + XeLaTeX + Noto CJK bilingual PDF build (matrix: zh-TW, en)
• .github/workflows/lint.yml: markdownlint + lychee link check
• .markdownlint.jsonc: linting rules aligned with geo-whitepaper
• assets/pdf/concat.sh: chapter concatenation script (strips frontmatter, JSON-LD, nav footer)
• CITATION.cff: GitHub auto-recognized citation info

Licensing

• Whitepaper: CC BY-NC 4.0
• Underlying software (PIF AI code): AGPL-3.0

Known Limitations

• Some PIF AI performance metrics remain marked as “target values”; formal Phase 1 GA benchmarks pending (§1.4.1)
• English chapters are human-reviewed machine-assisted translation; regulatory terms may still require local legal review (especially §2.5 international comparison)
• Mermaid diagrams are stripped in PDF build and replaced with “[See online version]” placeholders because XeLaTeX does not natively render Mermaid

Corresponding Commits (in baiyuan-tech/pif and baiyuan-tech/pif-whitepaper)

• f33392e — feat(i18n): extend locales to Japanese, Korean, French with language dropdown
• (RAG integration commit, pending) — feat(rag): central RAG integration (Scheme C+) backend
• (whitepaper commit, pending) — docs: initial whitepaper v0.1 draft with 12 chapters + 4 appendices

Future Plans

v0.2 (expected 2026-09)

• Replace “target values” with actual Phase 1 GA benchmark data
• Extended SA electronic-signature chapter
• New chapter: “Regulatory-Change Monitoring and Notification Mechanism” (aligned with Phase 3)
• Additional multilingual review for ja, ko, fr translations (originally UI-only; whitepaper body not yet translated)

v1.0 (expected 2026-12)

• Released alongside Phase 2 GA
• Includes first-batch anonymized customer case studies
• Formally peer-reviewed edition in partnership with academia

---

Nav ← Appendix C: References · Back to en README →

This site is open source. Improve this page.