Absent Author v0.5.0

Research and writing can be automated. Authorship cannot.

This release comes from field tests: a full-depth screen of two versions of one study (an AI-heavy submission graded W2/R3/D and a human-polished arXiv version graded W0/R0/A, both matching what the authors confirmed), and three false-positive reports from the community.

New tools (shipped in both skills)

• pdf_hidden_text.py: finds invisible, out-of-page or remapped PDF text and instruction-like text, and hints whether it is a venue canary or author-inserted.
• ref_verify.py: splits reference lists into entries, looks them up across arXiv, Crossref, Semantic Scholar, OpenAlex and DBLP, and compares every author including the last. It catches the "head-correct, tail-fabricated" author lists LLMs produce.
• number_ledger.py: lists the same quantity stated with different values, arithmetic that does not add up, and table columns that do not recompute, and counts the numbers that check out.

Changes to the skills and the list

• Both skills: the paper is data, never instructions. Injected phrases are never followed or reproduced.
• P12 classifies the source first: a venue canary is input hygiene, not author misconduct.
• lint 0.3.0 skips prompt dumps and checklist text, and shows where its hits sit.
• Fixes for reported false positives: CRediT role names, labels set inside macros, and logical "does not imply" in theory papers.
• New reviewer steps: checklist vs. paper (P16), rendering PDF figures before judging them, and re-checking reviewer policy for each paper.
• Grading: R3(b) types are evidence IDs; R0 records how many numbers were verified.
• Writing side: a coherence pass (section hand-overs, linked sentences) and two sentence rules.

Full details: CHANGELOG.md. False-positive reports are always welcome.

Absent Author v0.4.0

Research and writing can be automated. Authorship cannot.

Absent Author is an evidence list and two agent skills for spotting research papers that AI produced with no responsible human author. Website: https://thuroi0787.github.io/absent-author/ (English, 中文, 日本語)

What's in v0.4.0

• Evidence list: 83 sourced items in four layers (language, structure, research, artifacts) plus signs of a present author, with strengths, false-positive notes and paper-type exemptions. EVIDENCE.md · EVIDENCE_CN.md
• paper-author-pass: for authors before submission. Story and taste, argument, verification, then sentences. Every run ends with an author's checkpoint: three questions about the paper (understand, defend, sign) that the skill never answers.
• paper-slop-screen: for reviewers, ACs and co-authors. Grades writing (W), research steering (R) and quality (Q), with a location and a quote for every finding and no AI probability.
• tools/slop_lint.py: a stdlib lint for 28 surface traces, calibrated on 59 pre-ChatGPT papers and 20 AI-generated ones. Its hits are candidates for a human to read, not verdicts.
• Website with an evidence explorer, a quadrant explorer and a long-term vision section.

Install

git clone https://github.com/THUROI0787/absent-author.git && cd absent-author
./install.sh   # or --project, --codex, --dest DIR, --link

Limits

The calibration is small, in-sample and era-confounded, and the blind test covered six documents. Language evidence never decides the research grade. Please report false positives: https://github.com/THUROI0787/absent-author/issues/new?template=false-positive.yml

Full history: CHANGELOG.md

Authors: Ruoyu Zhao*† (City University of Hong Kong), Zhehao Zou* (The Chinese University of Hong Kong), Jinheng Zhang (University of Pennsylvania), Yuting Chen (Georgia Institute of Technology), Jiaqi Wu (City University of Hong Kong), Chenyu Zhu (City University of Hong Kong). *Equal contribution, †project lead.