PIF AI Whitepaper

Chapter 3: The 16 PIF Items — Deep Analysis

This is the most specification-dense chapter in the whitepaper. We analyze each of the 16 PIF items with: (a) what the regulation requires, (b) the data source, (c) PIF AI’s handling strategy, and (d) the database columns and code modules involved. At the end we generalize a “regulation-to-engineering mapping method” usable for extending to other jurisdictions.

📌 Key Takeaways

3.1 Classification Overview

By source, the 16 items cluster into five categories:

pie title PIF 16 Items — Source Classification
    "Form input" : 2
    "File upload" : 7
    "System scraped" : 1
    "Database query" : 2
    "Hybrid" : 3
    "Human signature" : 1

Figure 3.1: File uploads are the largest category (7 items, covering GMP and test reports), followed by hybrids (3, like usage instructions that combine user input with AI cross-checking). Only Item 16 is human-signed, but its status depends on the completion of items 1–15.

3.2 Item-by-Item

Each item follows the same template. The DB, AI module, and strategy fields are verifiable in the code.

3.2.1 Item 1 — Product Basic Data

Aspect Detail
Regulation Product name, category, dosage form, intended use, manufacturer info, TFDA registration number
Source Form input
AI strategy LLM structured validation (name zh/en consistency, TFDA number format)
DB columns products.{name, name_en, category, dosage_form, intended_use, manufacturer_name, manufacturer_address, registration_id}
Module app/api/v1/products.py, app/schemas/product.py

3.2.2 Item 2 — Product Registration Evidence

Aspect Detail
Regulation TFDA registration platform screenshot, registration date, validity period
Source System scraped (TFDA platform)
AI strategy LLM parses TFDA query response, extracts status
DB products.registration_id + uploaded_files (screenshot)
Module app/ai/tfda_registration_checker.py (planned)

3.2.3 Item 3 — Full Ingredient Names and Amounts

One of the most engineering-intensive items.

Aspect Detail
Regulation All ingredients in INCI nomenclature, with CAS, concentration %, function; sorted descending by concentration
Source User-uploaded formulation (PDF / Excel / image)
AI strategy Claude Vision parses document → INCI normalization → CAS validation → concentration total check (±2%)
DB columns product_ingredients.{ingredient_id, concentration_pct, function, sort_order} + ingredients.{inci_name, inci_name_normalized, cas_number}
Module app/ai/document_parser.py + app/ai/ingredient_validator.py

See §7.2 for the Claude Vision + INCI prompt design.

3.2.4 Item 4 — Labels / Packaging

Aspect Detail
Regulation Labels and packaging design proofs; must include legally mandated elements (ingredients, use, manufacturer, etc.)
Source User-uploaded design
AI strategy OCR + Claude Vision → 11-element regulatory checklist
DB uploaded_files (pif_item_number=4)
Module app/ai/label_checker.py (planned)

3.2.5 Item 5 — GMP Certification

Aspect Detail
Regulation Manufacturing facility GMP certificate; contract-manufacturers need OEM GMP
Source User upload
AI strategy Claude identifies certificate type (ISO 22716 / MOHW GMP / other), extracts validity, cross-references manufacturer
DB uploaded_files (pif_item_number=5, file_type='gmp')
Module app/ai/document_parser.py + pif_builder.py

3.2.6 Item 6 — Manufacturing Method / Process

Aspect Detail
Regulation Detailed manufacturing steps, operational parameters (T, t), equipment used, QC checkpoints
Source Manufacturer or contract manufacturer
AI strategy AI structures the narrative into a standard template; flags missing fields as [pending]
DB uploaded_files + pif_documents.ai_draft_url
Module app/ai/pif_generator.py

3.2.7 Item 7 — Usage Instructions

Aspect Detail
Regulation Method, recommended dose, frequency, applicable areas, precautions, warnings
Source User-provided
AI strategy AI cross-checks claim consistency — medical-efficacy claims (e.g., “improves acne scars”) are flagged as prohibited
DB pif_documents (item 7)
Module app/ai/usage_claim_validator.py (planned)

3.2.8 Item 8 — Adverse-Reaction Data

Aspect Detail
Regulation Historical adverse reactions; if none, a statement is required
Source User-provided
AI strategy AI classifies reactions (irritation / allergy / other) + risk grade
DB pif_documents
Module app/ai/adverse_reaction_classifier.py (planned)

3.2.9 Item 9 — Substance Characterization Data

Aspect Detail
Regulation Chemical name, molecular formula, MW, physical props (appearance, pH, solubility), chemical stability
Source PubChem automatic query
AI strategy LLM Tool Use → pubchem.query(cas_or_inci) → structured extraction
DB toxicology_cache.{data_json, source='pubchem'}
Module app/ai/toxicology_engine.py + app/mcp_servers/

3.2.10 Item 10 — Toxicological Data

Aspect Detail
Regulation Acute toxicity, skin/eye irritation, sensitization, genotoxicity, reproductive toxicity, carcinogenicity per ingredient
Source Cross-query across databases (PubChem, TFDA, ECHA, SCCS)
AI strategy AI synthesizes multi-source data → risk-summary table with citations
DB toxicology_cache.{data_json, risk_level, ai_summary, fetched_at, expires_at}
Module app/ai/toxicology_engine.py (Claude Sonnet + Tool Use)

See §9 for the full pipeline.

3.2.11 Item 11 — Stability Testing

Aspect Detail
Regulation Accelerated + long-term test reports; conditions, test items (appearance, pH, viscosity), conclusions, recommended shelf life
Source User-uploaded report
AI strategy Claude parses data tables + compliance judgment against ICH Q1A
DB uploaded_files (pif_item_number=11)
Module app/ai/test_report_parser.py (unified for §3.2.11–13)

3.2.12 Item 12 — Microbial Testing

Aspect Detail
Regulation Total plate count (TPC), E. coli, Staph. aureus, P. aeruginosa, etc.
Source User upload
AI strategy Parses test data + compares to TFDA cosmetic microbiology baselines
DB uploaded_files
Module Same as §3.2.11

3.2.13 Item 13 — Preservative Efficacy Testing

Aspect Detail
Regulation Challenge Test report (inoculation strains, survival counts, log reduction, Grade A/B)
Source User upload
AI strategy Parses data + Pass/Fail judgment against ISO 11930
DB uploaded_files
Module Same

3.2.14 Item 14 — Functional Evidence

Aspect Detail
Regulation Evidence supporting efficacy claims (clinical, in-vitro, literature). Medical-efficacy claims prohibited.
Source User upload
AI strategy Matches “claim” vs “evidence” consistency; flags medical-efficacy violations
DB uploaded_files
Module app/ai/claim_evidence_checker.py (planned)

3.2.15 Item 15 — Packaging Material Report

Aspect Detail
Regulation Packaging specs, heavy-metal migration tests (Pb, Hg, As, Cd), plasticizer tests, container compatibility
Source User upload
AI strategy Parses test report + compares to regulatory thresholds
DB uploaded_files
Module Same as §3.2.11

3.2.16 Item 16 — SA Safety-Assessment Signature

Aspect Detail
Regulation A qualified SA reviews items 1–15 and signs a safety-assessment report
Source SA online review
AI strategy AI produces a draft (does not substitute for SA judgment); SA revises + signs electronically
DB sa_reviews.{ai_draft_assessment, sa_final_assessment, sa_comments, signature_url, signed_at}
Module app/services/sa_workflow.py + app/api/v1/sa_review.py

See §11 (SA workflow) for details.

3.3 State Machine

Each pif_documents record has a status field. The state machine:

stateDiagram-v2
    [*] --> missing: item created
    missing --> uploaded: user uploads file
    uploaded --> ai_processing: trigger AI analysis
    ai_processing --> ai_draft_ready: AI success
    ai_processing --> uploaded: AI failure (fail-soft)
    ai_draft_ready --> human_reviewed: user review
    human_reviewed --> approved: SA final sign-off
    approved --> [*]

Figure 3.2: Six legal states. The ai_processing → uploaded fail-soft edge lets users retry after transient AI failures. Only an SA may trigger approved (enforced at the application layer), mapping directly to the Act’s SA signature requirement.

3.4 The Regulation-to-Engineering Mapping Method

This whitepaper proposes a five-column mapping method — applicable to extending to other jurisdictions (EU CPNP, US MoCRA):

Column Content
① Regulation Text, article number, date
② Source Who (user / system / SA) provides? What format?
③ AI operation LLM Tool Use verb (parse / validate / compare / summarize)
④ DB column Table.Column or JSONB path
⑤ State-machine state Position in the 6-state model

Any new regulatory obligation can be systematically translated into engineering work, avoiding ad-hoc “where does this belong?” chaos. See §8 for Schema detail.

📚 References

📝 Revision History

Version Date Summary
v0.1 2026-04-19 First draft. 16 items deep analysis, state machine, five-column mapping method

© 2026 Baiyuan Tech. Licensed under CC BY-NC 4.0.

Nav ← Chapter 2: Regulatory Background · Chapter 4: System Architecture →