Retraction check — supporting material

Supporting material for the Retracted Literature project.

Contents

• check_retractions.py — the checker. Walks every genes/*/*/*-ai-review.yaml,
  collects each PMID: citation with its citation site (reference,
  annotation = an annotation's original_reference_id, or supporting_text =
  a supported_by[].reference_id), queries PubMed in batches of 200 through NCBI
  E-utilities efetch, and reads the PublicationType list and the
  CommentsCorrections RefTypes. Records that are themselves retraction /
  expression-of-concern / erratum notices are excluded — citing the notice is
  good practice, not a defect.
• retraction-check.tsv — generated: one row per flagged PMID, with severity,
  the raw PubMed signals, the notice PMID(s), which genes cite it, how they cite
  it, and whether the review already records is_invalid: true.
• retraction-check.json — generated: the same data plus run totals and the
  PMIDs PubMed did not resolve.
• retraction-register.md — generated register (do not edit by hand):
  retractions first, then expressions of concern, then errata in a quiet
  informational section.

Regenerate

uv run --no-dev python projects/RETRACTIONS/check_retractions.py

A full pass is ~126 efetch requests over ~25,000 PMIDs and takes the better part
of an hour; set NCBI_API_KEY to raise the rate limit. Useful variants:

# check specific PMIDs (still reports where they are cited in the reviews).
# A partial run needs its own --out-dir so it cannot overwrite the full-scan
# reports with a handful of rows; --pmids also accepts a comma-separated list.
uv run --no-dev python projects/RETRACTIONS/check_retractions.py \
    --pmids 19225519,16616141 --out-dir /tmp/spot

# rebuild the register from the last run, no network
uv run --no-dev python projects/RETRACTIONS/check_retractions.py \
    --from-json projects/RETRACTIONS/retraction-check.json

Severity tiers

The citation_sites column is what makes a row actionable: a retracted paper in a
references: list is bibliographic cleanup, whereas the same paper as an
annotation's original_reference_id is a curation question.