← Showcases

A retracted paper, and what rested on it

A recorded run of YuCe on a real retraction. Replay it below, frame by frame, in your browser: nothing to download and nothing to install.

  1. What happened

    In April 2026 the journal Humanities and Social Sciences Communications retracted a meta-analysis on ChatGPT and learning (Wang and Fan, 2025). Crossref counted it as cited more than 300 times when checked on 20 November 2026. Five open-access papers published after the retraction still cite it.

    The journal’s official retraction notice

  2. What the agents did

    Three AI agents read those five papers. Every claim they wrote down had to carry exact quotes: the paper’s own passage that cites the meta-analysis, and the title line of the meta-analysis’s record. Checked against a list of the 8 passages in which the five papers cite it, the agents’ quotes covered all 8.

  3. What the kernel did

    The owner of the record then withdrew one line of it: the title line of the retracted paper. The kernel marked as fallen every claim the agents had tied to the retracted work by quoting that line: 15 claims, and the 2 answers that rested on them (17 records, as the view’s “What fell” list shows). 6 of the 8 passages sit under one of those claims. The kernel follows the grounds the agents declared; it does not read what a paper means.

    The withdrawal itself took milliseconds: in 10 repeats on fresh copies of this journal, with YuCe 0.2.2 on a laptop, the call took a median of 23 ms (at most 37 ms), and building the full list of what fell took about 0.35 s. Those times belong to this journal and that laptop; the journal itself keeps no clock.

A recorded run, not the live kernel.

Nothing on this page, or in the view below, runs the kernel or calls a model. The run was recorded with YuCe 0.2.2: three AI agents, one run, 143 model calls. The view reads only its journal. In the run the owner withdrew the line by hand, as one does when a source is retracted: the kernel did not learn of the retraction by itself.

Replay the run

Press play, or step through the frames. Choose “Withdrawals” under “Key frames” to jump to frame 446, where the owner withdraws the line, then open “What fell” to see what lost its ground and why.

The view is in English and makes no network request. On a phone the full page is easier to use than the frame.

What this run does not show

  • A claim marked as fallen is not false: one of the grounds it was recorded on was withdrawn. A paper’s own result may hold without the retracted one, and nothing here re-runs any analysis.
  • It does not say that a paper’s conclusions are wrong, that its authors knew of the retraction, or why the paper was retracted.
  • The kernel checks bytes and recorded grounds, not meaning. It knows only the grounds the agents declared: a statement whose dependence on the retracted paper no agent recorded is on no list. 2 of the 8 passages are not under a claim that fell: none of the claims the agents recorded for them rested on the withdrawn line.
  • The agents’ readings are AI readings, pending human review. This is one recorded run, with nothing to compare it with.
  • A journal that verifies is consistent with itself; that does not make it authentic by that fact. When the view was made, 46 of its 454 frames were compared with the kernel’s own answer, and the view says so in its own panel.

About the run

  • The view is made by YuCe Live from one journal: 455 entries, 1,186 records, and audit-verify answers VERIFIED. The journal is the file run.wal (2.2 MB), shown as it was recorded. SHA-256: a47fefc22d2cfc2a0f0778a87b49785d5edfd303b510e2946a5441f3868374e0.
  • The five papers are the CC BY texts listed below, quoted as exact bytes of plain-text versions made from the Europe PMC full-text XML (layout and figures are not kept). The retracted paper’s text is not included: its record file holds bibliographic facts only.
  • YuCe is open source. How it works, try the rules yourself in the sandbox, another recorded run (two RFCs), and the docs for running it on your own folder.

Check the journal yourself

The view shows what the journal holds, and the journal is a file you can check: here in your browser, with nothing to install, or on your own machine with the kernel.

In your browser

The check is the independent checker yuce-verify 0.2.1: one JavaScript file of 94 KB, released under the Apache-2.0 licence, and its source is open: the verifier/ folder of the YuCe repository, where the built file yuce-verify/dist/yuce-verify.mjs is the one hosted here. It hashes with your browser’s own crypto and makes no network request: the journal and the documents never leave your browser. SHA-256 of the checker: 27d07ff676a82c3c338db315965e12cf3520416ab0fb9dc7025e198d3c44d285. The six documents the journal quotes from (five CC BY papers and the record file of the retracted paper) are hosted beside it, with every SHA-256, in this README.

Drop a journal file here (.wal), with its documents beside it if you have them, or

Choose files

The check runs in a browser that can run JavaScript modules. If this message does not change, use the commands below.

“Check this journal” loads the journal published here and the six documents. To check a copy of your own, download run.wal and drop it here: a file that is the same as this one passes. Change one byte of your copy, with a hex editor or with the second button, and the check fails at the first frame the change touches.

What this check says, and what it does not

  • PASS means the journal is consistent with itself: well-formed frames, an unbroken chain of hashes, its checkpoints and, with the documents, every quote found byte for byte in its document.
  • It does not prove who wrote the journal. Without a hash or a signed checkpoint kept where the writer cannot change it, authenticity is not established; the SHA-256 published on this page is such a hash for this file.
  • It does not recompute what fell after the withdrawal and does not check signatures: its answer says “Not verified: withdrawal cascade, signatures”. The kernel does the first (see below).
  • A quote replaced by another phrase of the same document is not seen by this check.

On your own machine

With the kernel of the release v0.2.2 you can check the file’s hash, that the journal is consistent with itself, and what fell after the withdrawal.

sha256sum run.wal # a47fefc22d2cfc2a0f0778a87b49785d5edfd303b510e2946a5441f3868374e0 sbgk audit-verify --wal run.wal # status VERIFIED, 455 entries, 1186 records sbgk audit-read fallen --wal run.wal --since 446 # what fell after the withdrawal at frame 446: 17 records

The papers, and their licence

The retracted meta-analysis, as a bibliographic reference only: Wang J, Fan W. (2025). The effect of ChatGPT on students’ learning performance, learning perception, and higher-order thinking: insights from a meta-analysis. Humanities and Social Sciences Communications 12:621. doi:10.1057/s41599-025-04787-y.

The five papers that cite it are open access under the Creative Commons Attribution 4.0 licence (CC BY 4.0). Short quotations from them appear in the view above, as exact bytes of plain-text versions, and the plain texts themselves are hosted for the check in the browser, each keeping its header with the source and the licence; the wording is not otherwise changed. Each paper is credited to its authors, with its source:

  • Li J. (2026). The causal effects of artificial intelligence use on metacognition, engagement, and knowledge transfer in educational contexts. Frontiers in Psychology. doi:10.3389/fpsyg.2026.1811065. Copyright © 2026 Li. Licence: CC BY 4.0.
  • George-Reyes CE, Romero CA, Merida-Cordova EJ, Yeguas-Bolivar E. (2026). ChatGPT and higher education in Latin America: measuring perceived academic skills. Frontiers in Research Metrics and Analytics. doi:10.3389/frma.2026.1816902. Copyright © 2026 George-Reyes, Romero, Merida-Cordova and Yeguas-Bolivar. Licence: CC BY 4.0.
  • Li Y. (2026). A study on the correlation between the perception of intelligent college English learning environments and the willingness to communicate in listening, speaking, reading, and writing. Frontiers in Psychology. doi:10.3389/fpsyg.2026.1885211. Copyright © 2026 Li. Licence: CC BY 4.0.
  • Manushi Dinasha WD, Nissanka NMTA, Wickramasinghe CP, Fernando WPPD, Perera V, Gunatilake HGD. (2026). Assessment of the efficacy of ChatGPT responses to bacterial species-specific questions in microbiology. Access Microbiology. PMC13449012. doi:10.1099/acmi.0.001137.v5. Copyright © 2026 The Authors. Licence: CC BY 4.0.
  • Fang L, Li Z, Xiong B, Lu Z. (2026). Effects of generative artificial intelligence on students’ learning motivation: evidence from a meta-analysis. Frontiers in Psychology. PMC13647323. Copyright © 2026 Fang, Li, Xiong and Lu. Licence: CC BY 4.0.

The papers belong to their authors and journals. YuCe and Organetic Sphere are not connected with them, and the run says nothing about the quality of their work beyond what is stated above.