Public retraction log

Verdict diff — arXiv:2301.12597

The verdict on BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models was retracted on 2026-05-13. Below is the structured before / after: what we said, what we say now, and why the original verdict was incorrect.

Before / after

Before — original verdict
WRONG
agent_version
v0.1.0-blip2-flickr30k-n20-greedy
After — current verdict
PARTIAL
The reproduction was re-run under a corrected protocol and now sits at PARTIAL. Click through to the paper page for the live evidence + claim citation.

Why the original verdict was incorrect

The driver decoded with greedy search, n = 20 examples, and no prompt prefix. The paper uses beam search width 5 with a "a photo of" prompt prefix on the full Flickr30k Karpathy split. The retraction replaces the WRONG row with a PARTIAL captured under the paper's published decoding protocol.

Evidence trail

  • Audit thread — long-form post-mortem covering all seven 2026-05-13 retractions, including this one.
  • Rollback PR — the GitHub pull request that landed the corrected driver and flipped the verdict row to is_current=false.
  • All public retractions — the append-only retraction log under PRD §17.X.8(d).
  • Verdict Validator — the C1/C2 gates that prevent this class of mistake from shipping again.

What changed structurally

The 2026-05-13 retraction rollup landed two structural fixes so the citation-side failure that caused the original incorrect verdict cannot ship the same way again:

  1. Typed claim citation per verdict. Every reproduction driver now declares a structured CLAIM_CITATION (Table, row, column, reported value, quoted text, PDF page) before its Modal job runs. The original verdict on this paper was published against a non-citable headline — that path is now closed by the build-failing validator-wiring lint.
  2. PDF-verified textual gate. The Verdict Validator fetches the cited paper's PDF and checks that the quoted text appears within ±200 characters of the cited reported value. Made-up, mis-cited, or category-confused citations fail this gate and the verdict is auto-downgraded.