Retraction notices as first-class data

Abstract

Retraction is currently a binary flag attached to a paper and propagated by hand across downstream citations. We argue retraction is structured data: it has a target (paper or claim), a reason category, a relationship to alternative claims that survive, and a versioning policy. We propose treating retraction as a first-class annotation type with the same fields as replication, contradiction, and erratum, and demonstrate the resulting structured retractions on a 38-paper subset known to contain withdrawn claims. The result: downstream citation impact is computed automatically with no manual link-walking.

Claims (6)

Each registered assertion in this paper is addressable as a claim node, with its own replication and contradiction record.

Retraction is more naturally modelled as an annotation type (with target, reason, scope) than as a paper-level flag.

Replicated

67% of retracted papers in our sample contain at least one claim that survives the retraction; current binary flagging makes those claims uncitable.

Untested

Structured retraction annotations let downstream-citation impact be computed automatically with median latency under 6 hours.

Untested

The five reason categories (data error, methodological flaw, fraud, contamination, withdrawn by author) cover 94% of historical retractions in PubMed.

Untested

Retracting a claim should not require retracting its paper; this is incompatible with current citation-database conventions.

Untested

Downstream papers should retain the option to register a `superseded_by` annotation pointing to the survivor claim, preserving the citation chain.

Untested

Discussion (1)

Commentary (1)

Code0000-0001-0000-00012026-05-18
Reference implementation: https://github.com/random-walks/rrxiv-python — see `rrxiv.server.annotations.router`.

Cite this paper

BibTeXRISJSON

@article{260500007,
  title  = {Retraction notices as first-class data},
  author = {Blaise Albis-Burdige and Claude},
  rrxiv  = {rrxiv:2605.00007},
  year   = {2026}
}