Retracted10.1186/s13046-018-0845-9
- Duplication of/in image
- Investigation by third party
- Paper mill
Graph Hacks 2026 · Track 01 · built on FalkorDB
What is your bibliography standing on?
Paste your references. Bedrock follows every citation path from your sources to retracted research, and shows you the chain.
- Your reference10.1016/j.neubiorev.2019.09.040
- Intermediate source10.1111/ejn.13603
- Retracted 2014-10-0110.1016/j.jpsychires.2014.06.018
This path comes from the audited corpus. Reference managers flag a paper that cites a retracted one directly. Bedrock also finds the ones that reach it through another source.
61,299retracted works indexed, from the Retraction Watch database
251,645articles from 213 journals checked for exposure
7,162of them (2.8%) cite a work that has since been retracted
How it works
From a pasted list to a decision you control
Paste your references
Any text that contains DOIs works, up to 300 at a time. Bedrock finds them and ignores the rest.
The graph finds the paths
FalkorDB follows citations up to two hops and returns every path that reaches a retracted work, with the retraction date and the recorded reasons. In our tests, 150 references took between 0.15 and 0.4 seconds.
The agent explains, you decide
A language model summarizes the findings and proposes edits. It can only read. Nothing changes until you approve each action.
Example output
What a finding looks like
These cards use real DOIs from the audited corpus. Try the approval panel: it shows how the revised list changes only for the actions you accept.
Cites a retracted work10.1016/j.phrs.2019.104313
10.1016/j.phrs.2019.104313 → 10.1186/1476-4598-12-147 (retracted 2022-12-15)
Two hops from a retracted work10.1016/j.jpdc.2021.07.009
10.1016/j.jpdc.2021.07.009 → 10.1016/j.jpdc.2017.05.009 → 10.1016/j.procs.2014.05.039 (retracted 2014-06-06)
No exposure found10.1111/iej.13488
Built to be checked
The agent is on a short leash
It can read, never write
The agent has three tools, all read-only. None of them changes a bibliography. Edits happen only when you approve them on the page, and the server applies only the actions it proposed itself.
Every DOI is checked against the data
After the agent answers, each DOI it mentions is compared with what the tools returned. An invented or mistyped one shows up as a warning.
The findings do not depend on the model
The finding cards come straight from graph queries. If the model is slow, rate-limited or wrong, you still get the full audit.
It never accuses people
Findings concern papers and citation paths. The public demo holds no author data. A flag is a prompt for human review.
What we tested
We tried to predict retractions. It did not work, so Bedrock does not claim to.
Before looking at any result, we wrote down the success test: a graph method had to beat the best simple baseline, and the gap had to survive removing the largest journal. We then ran it on 134,798 articles with 140 later retractions.
| Method | Score | Average precision | Top 1% recall |
|---|---|---|---|
| Random ranking | 0.0012 | 1.6% | |
| Journal's past retraction rate | 0.0066 | 0% | |
| Cites a retracted work (one hop) | 0.0066 | 10.0% | |
| Two citation hops | 0.0021 | 7.3% | |
| All graph signals combined | 0.0041 | 7.1% | |
| Journal history plus graph | 0.0074 | 15.7% |
The result: no reliable improvement. The best combination beat the journal baseline by a margin well inside the uncertainty, the multi-hop signals did not beat their one-hop versions, and most of the 140 retractions came from a handful of journals. One signal did stand out: a paper that cites an already retracted work was 5 to 12 times more likely than average to be retracted later, depending on the cutoff. That is a lookup, so Bedrock reports exact provenance instead of forecasts. Method, intervals and caveats are in docs/RESULTS.md.
Try it
Three ways in
Live audit
Paste a reference list and get findings, paths and proposed edits in seconds. The live demo runs on a small AWS server. If it is ever slow or down, run it yourself with one command (see the repository).
Link added at launchDemo video
A two-minute walk through an audit, the approval step and the backtest result.
Link added at launchSource and setup
The full pipeline, the Cypher queries, the tests and the decision log, with setup steps in the README.
Open the repository