What a Crossref record shows when a paper is retracted: one month of retraction deposits (v3)

Abstract

# What a Crossref record shows when a paper is retracted **A field-level audit of one month of retraction deposits (v3)** Author: Trafalgar Law (independent). Data pulled from the public Crossref REST API on 2026-09-23. Sample and reproduction query included. This is a metadata audit, not an accusation against any publisher or author. ## Why On 2026-09-19 the record for a fisetin/renal-fibrosis paper in *Applied Biological Chemistry* (10.1186/s13765-025-01072-z) was updated in Crossref. The title field had been rewritten to `RETRACTED ARTICLE: ...` and the record carried a new `updated-by` link to a retraction note. The abstract field still carried the original study text, word for word, with nothing in it about a retraction. That is one record. v1 of this audit looked at one day (100 notices). This version looks at one month, and the question is the same: when a paper is retracted, which parts of the record actually move, and what does a reader who only sees the metadata learn? ## Method Query, 2026-09-23: ``` GET https://api.crossref.org/works ?filter=update-type:retraction,from-deposit-date:2026-09-01 &rows=100&cursor=* &select=DOI,title,update-to,deposited,container-title,type,abstract,published,updated-by ``` Cursor-paginated to exhaustion. This returns every Crossref record carrying a retraction-type update deposited from **2026-09-01T00:44:37Z to 2026-09-23T04:50:07Z**: **1,303 notices**. For every `update-to` entry of type `retraction`, I recorded the target article DOI. Where the target DOI differs from the notice DOI, I fetched the target record (`GET /works/{DOI}`) and read `title`, `abstract`, `updated-by`, `published`. Where the target DOI equals the notice DOI, the publisher rewrote the article record in place; those records are read from the notice set itself. Counts: - 1,303 retraction-type notices. - 1,391 notice -> article links. - 1,293 unique article DOIs. - **839 of those 1,293 are the same DOI as the notice** (in-place rewrites). That is 65%, much higher than the 28% seen in the one-day v1 sample; a single month mixes in bulk in-place remediation of older articles. - **454 distinct article records are separate from their notice** and were fetched individually (454/454 retrieved, no gaps). "Retraction word" means the case-insensitive stem `retract` or `withdraw` appearing in the field. Every raw value is in the attached CSV (1,293 rows). ## Findings ### A. The 454 separately-noticed article records | Field | Result | |---|---| | `updated-by` back-link to the retraction | **454/454 (100%)** | | title contains a retraction word | 374/454 (82%) | | abstract present in Crossref at all | 190/454 (42%) | | abstract mentions the retraction | **4/454 (1%)** | | publication date rewritten to the retraction date | 0/454 | | silent in both title and abstract | **79/454 (17%)** | | of those 79, an abstract is present and still serves the original text | 49/79 (62%) | ### B. The 839 in-place rewrites (notice DOI == article DOI) | Field | Result | |---|---| | title contains a retraction word | 828/839 (99%) | | abstract present in Crossref at all | 3/839 | | silent in both title and abstract | 11/839 (1%) | ### C. All 1,293 distinct article records | Field | Result | |---|---| | title contains a retraction word | 1,202/1,293 (93%) | | silent in both title and abstract | **90/1,293 (7%)** | ## What this means The machine-readable side of retraction is close to perfect. All 454 separately-noticed article records carry an `updated-by` link to their retraction. Any system that follows links finds it. The human-readable side is not. Over one month, **90 of 1,293 retracted article records say nothing about the retraction in the two fields a person skims**. In 49 of those 90 the record still serves the original study abstract verbatim, with the retraction unmentioned; in the remaining 41 there is no abstract in Crossref and the title carries no retraction word either. Three patterns worth naming: - **The abstract almost never moves.** Of the 190 records that carry an abstract, 4 mention the retraction. The abstract field is effectively write-once: publishers flag the title and leave the abstract as deposited. That is the fisetin case, and it is the norm, not an exception. - **In-place rewrites are the majority of notices (65%) and they are clean.** When a publisher rewrites the article record rather than depositing a separate notice, the title flags the retraction 99% of the time. The silent cases concentrate in the separately-noticed set. - **No single publisher or journal explains it.** Of journals with 10 or more separately-noticed records in the month (Scientific Reports 16, Food Science & Nutrition 12, Thinking Skills and Creativity 11), none had a silent record. The 90 silent records are spread thin across dozens of journals. ## Who this matters to Anyone whose tooling reads Crossref metadata without following `updated-by`: reference managers, citation-alert and literature-monitoring services, aggregators, and the retrieval layer of AI literature tools. A link-following system is safe. A text-reading system that shows title and abstract is told nothing in 7% of cases, and in 49 of those it is shown the original abstract of a retracted paper as if it were live. ## What this does not show - Crossref abstracts are optional and publisher-supplied. "No abstract" is not evidence that a publisher hid anything; it is evidence that Crossref cannot help a reader who needs one. - This is metadata only. A publisher's own landing page usually does display a retraction banner; this audit does not measure that and does not claim otherwise. - The window is one month of deposits, not all retractions. Older records remediated in bulk inside this window are counted where they were deposited. ## Update (v3): does OpenAlex see the same silence? v3 adds one cross-check, because Crossref is not the only index a reader's tool may sit on. For the **90 records that are silent in both title and abstract in Crossref**, I asked OpenAlex (`api.openalex.org/works/doi:{DOI}`) the same two questions: does it flag the work as retracted, and does it carry an abstract? Result, all 90 rows in the attached `openalex_crosscheck_90.csv`: | Question | Result | |---|---| | OpenAlex marks the work `is_retracted` | **90/90 (100%)** | | OpenAlex carries an abstract at all | 75/90 (83%) | | of the 50 Crossref records still serving the original abstract, OpenAlex also carries an abstract | **50/50 (100%)** | So the silence is not a Crossref-specific rendering bug: where Crossref still shows the pre-retraction abstract, OpenAlex holds an abstract for the same work, and OpenAlex's retraction flag rides on the work independently of the title text. Retraction status propagates through index-to-index links almost without loss. What does not propagate is a human-readable signal inside the record itself. This is a small cross-check, not a second study. It exists so a reader can check the claim: follow either index, and the retraction is there in the machine fields; read the record, and in 7% of cases it is not. Related work: the same question is being measured from the Retraction Watch side in a preregistered study by Veronika Lux, *Retraction status propagation into Crossref and OpenAlex* (OSF preregistration, DOI 10.17605/osf.io/grnv8). This audit works the same question from the deposit-field side. ## Reproduce `notices.json` is the raw Crossref response set; `sep_audit_rows.csv` has one row per distinct article record with the field flags; the pull, extract and analysis scripts are short and deterministic. Re-running the query on a later date will return a different window, which is the point: this is a rate, not a fixed list.