Sift

Two pre-submission checks for academic papers, in one Streamlit app:

• Reference checker — flags hallucinated citations and wrong author names in any PDF.
• LaTeX diff — colored, strikethrough diff between two .zip versions of a paper, with cite/ref numbers preserved from the old version.

Run locally

pip install -r requirements.txt
OPENALEX_KEY=... S2_API_KEY=... streamlit run app.py

Both keys are optional but raise rate limits and unlock cross-checking.

Offline databases (recommended)

Drop dblp.db and/or acl.db into databases/. Sift auto-detects them. Without them, queries hit the public APIs and are slower / rate-limited.

curl -sSf https://hallucinator.science/install-cli.sh | sh
hallucinator-cli update-dblp databases/dblp.db   # ~2.5 GB, 20–30 min
hallucinator-cli update-acl  databases/acl.db    # ~60 MB

Layout

slate/
├── app.py            # Streamlit entry, sidebar, tabs
├── check_ref.py      # extract + validate + audit + cross-check
├── diff.py           # zip extract + latexdiff + pdflatex pipeline
├── databases/        # offline DB dumps
├── requirements.txt
└── README.md

Inspired from

• hallucinator — reference validation across academic databases
• git-latexdiff-web and latexdiff.cn — robust LaTeX diff