Limitations

1. The builds are unsigned. macOS and Windows will warn you; Install shows how to open them.
2. Windows and Linux are preview, not yet tested. The installers exist; the app has not been launched on those systems by the vendor. See Platform support.
3. No OCR for scanned PDFs. A PDF with no text layer is refused with a clear message.
4. Local speech-to-text is not available on Intel Macs. Use Gemini or OpenRouter instead.
5. Table and figure detection is imperfect and slow. Detection is 12–54 s per paper; extraction adds a model call per element (about 92 s for a 2-page paper, 278 s for three charts). Chart values are always estimates.
6. Zenodo deposits are drafts, and the live path is untested. The app never publishes; the user publishes on Zenodo. All tests ran against a local mock.
7. The AI-use statement is optional and never pushed. It is never inserted automatically and never required to export.
8. You bring your own AI engine. No model is included; free OpenRouter models have daily limits.
9. There is no log file or diagnostics export yet. Run the app from a terminal to capture output.

Smaller things: importing an existing single-file paper cannot be drafted section by section; dblp is link-out only and bioRxiv/medRxiv search runs through Europe PMC; CORE needs a key and Unpaywall an email; Tectonic's first compile is fast only after a background preload that needs a network connection the first time; Kaggle and warehouse imports are not built; there is no collaboration, cloud sync, hosted version, auto-update or mobile app.

For technical users

Vault Vision's evaluation and held-out sets cover 79 + 60 + 40 pages; the arXiv set is optimistic; held-out tables 92.6% precision / 100% recall, figures 85.3% / 96.7%. These are research measurements, not product benchmarks, and are not quoted as marketing numbers.