AI disclosure
A record of AI-assisted mathematics, kept with AI assistance. Here is exactly where.
Where models are used
In reviewing submissions: reading the source, checking that a disclosure is in the paper and not only in the form, scanning for duplicates, drafting the verification, result and AI-role notes a curator then edits and signs. In scoring significance, where the whole ladder is AI-assigned against a published prompt. In finding candidates, where a script harvests arXiv, trackers and repositories and flags disclosures for a human to triage.
Which models
Anthropic's Claude, through Claude Code, is the working model for review notes and curator prose; the significance prompt names the model current at each rescoring. When a note quotes a paper, a tracker or an expert, the words are theirs and the source is linked. Per-field provenance - which model, when, from which source - is recorded from 2 September 2026 on and shown as a marker on the field; curator prose written before that date was AI-assisted without a per-field record, and is not marked.
What this is not
Not expert peer review. A model reading a paper and a curator checking the model is AI-assisted curation, and the verification ladder says so on every entry: Unreviewed means nobody independent has checked the mathematics, and most entries sit there. Claims that would be a landmark are held rather than listed until an expert or a formal proof exists, whatever a model thought of them.
How it is corrected
In the open. Every field is editable by signed-in readers with the change in a public changelog; every entry has a discussion thread and a report button; and readers have downgraded, relabelled and unpublished entries after showing a model or a curator was wrong. If something here is wrong, the fastest fix is to say so.
The standard the site holds itself to
Entries have been declined because a paper's disclosure named no model and attributed no step. The site should meet the bar it applies: this page says which models, the provenance markers say which field, and the reviewing checklist says what a human checks before anything is published. The source code is public, including the scripts that record each review.