Medical Claims Verification

Rules

Rule 1: Every PubMed Query MUST Be Tested

Before any PubMed scan code is merged, the query builder MUST have tests covering:

• _name_clause() correctly handles single names, multi-word names, and OR-separated synonyms
• Gene synonym expansion produces valid PubMed boolean (not quoted phrases)
• No duplicate genes per pathogen (test: len(genes) == len(set(genes)))
• No synonym pairs listed as separate genes (test: known synonym pairs must not coexist)
• No non-pathogen entries in pathogen lists (e.g., Human GRCh38)

Test file: tests/test_pubmed_queries.py — run with python3 -m pytest tests/test_pubmed_queries.py -v

Rule 2: Never Claim "Zero Research Exists"

The correct phrasing is always:

"Zero results returned by our PubMed query [exact query string] as of [date]"

NEVER say:

• "No research exists"
• "Completely unexplored"
• "No one has studied this"
• "This has never been done"

Why: PubMed is one database. It misses:

• Chinese databases (CNKI, Wanfang, VIP)
• Russian databases (eLIBRARY.ru, CyberLeninka)
• Arabic/Persian databases
• Preprints not indexed in PubMed (bioRxiv, medRxiv partial coverage)
• Patents (Google Patents, Espacenet)
• Conference proceedings not indexed
• Dissertations and theses
• Papers with different terminology than our query

Rule 3: Every New Gene MUST Have Synonyms Checked

When adding a gene to PATHOGEN_GENES:

1. Search UniProt for the gene name → note all synonyms
2. Search PubMed manually with the primary name AND each synonym
3. Add all synonyms that return different results to GENE_SYNONYMS
4. Verify the gene is not a synonym of an already-listed gene

Synonym trap examples:

• NP = nucleoprotein (same protein)
• S protein = spike = spike protein (same protein)
• F protein = fusion protein (same protein)
• G protein = attachment protein (RSV context)
• inhA = enoyl-ACP reductase = Rv1484 (same enzyme)

Rule 4: Every Scan Result MUST Be Spot-Checked

After running pubmed_scan.py, manually verify at least:

• 3 claims that return zero (are they genuinely zero, or query bugs?)
• 3 claims that return low counts (1-5; could be synonym issues)
• All newly-added pathogens or genes (first scan of a new entity)

Use scripts/spot_check_claims.py or manual PubMed queries.

Rule 5: Numbers Flow One Direction

The single source of truth for gap counts is web/data/pubmed-scan-results.json. When the scan runs, all documents MUST be updated from this file:

pubmed_scan.py → pubmed-scan-results.json → all papers/briefs

NEVER manually edit a number in a paper. Always re-run the scan and propagate.

After updating, grep for stale numbers:

grep -rn "OLD_NUMBER" publications/ docs/ --include="*.md"

Rule 6: Pathogens ≠ Reference Genomes

Human GRCh38 is in the CRISPR target database for off-target screening. It is NOT a pathogen. It MUST be excluded from gap analysis. The scan script MUST filter out non-pathogen entries before generating claims.

Rule 7: Document Non-English Database Limitations

Every paper MUST contain a limitation statement about PubMed's English-language bias. Template:

Limitation: English-language database bias. Our literature analysis queries PubMed exclusively, which has limited coverage of research published in Chinese (CNKI, Wanfang), Russian (eLIBRARY.ru), Arabic, and other non-English databases. Gap claims reflect PubMed coverage as of [date], not the totality of global CRISPR research. Researchers in non-English-speaking countries may have published relevant work not captured by our scan.

Verification Checklist

Use this checklist before publishing any claim:

□ Query string logged with exact PubMed syntax
□ Date of query recorded
□ OR-separated names produce boolean, not phrases
□ Gene synonyms expanded (checked against UniProt)
□ No duplicate genes in gene list
□ Non-pathogen entries excluded
□ At least 3 zero-result claims manually spot-checked
□ Phrasing says "our query returned zero" not "no research exists"
□ Non-English database limitation documented
□ All documents updated from single source (JSON)
□ grep confirms no stale numbers remain
□ Tests pass: pytest tests/test_pubmed_queries.py

Recovery Procedure

If a false claim is discovered:

1. Stop. Do not publish, present, or share anything until fixed.
2. Run scripts/spot_check_claims.py on the specific claim
3. If confirmed false: fix the query bug in pubmed_scan.py
4. Add a regression test to tests/test_pubmed_queries.py
5. Re-run full scan: python3 scripts/pubmed_scan.py
6. Update ALL documents (grep for old numbers)
7. Rebuild ALL PDFs: bash scripts/build-preprints.sh
8. Redeploy: bash scripts/deploy-azure.sh --site-only
9. Record lesson in docs/lessons/
10. Commit with fix: prefix explaining the false claim