Method

Screening and verification. One model reads the article and flags possible issues. Each flag is then assessed independently by two further models. An issue is confirmed only when at least two models agree; all other flags are retained on record but not acted upon.

Source comparison. Where a reference is questioned, the published abstract is retrieved from PubMed or Crossref so the claim is compared with the source rather than with model recall.

Correction tiers. Tier A covers spelling, grammar and formatting. Tier B covers a single clear-cut value, such as a unit, and requires editor approval. Tier C covers anything affecting data, results, clinical claims or references; it is referred to the author and requires editorial approval.

Transparency. No published text changes without a logged version. Each correction is dated and listed under Corrections & Clarifications on the article, consistent with COPE guidance on post-publication amendments.

Reader reports. Any reader may report a possible error using the form at the end of each article. Reports are checked before other articles.