Submit an entry
Know of a problem proved or disproved with an AI model in the loop that isn't tracked here? Propose it. A reviewer checks the source before it goes live, and the published entry credits you by pseudonym.
What belongs here: a precisely stated open question whose answer is now a proved or disproved theorem, with AI in the loop - any field of mathematics, theoretical computer science included. Not in scope: formalizations of known human results, empirical attacks, or improved heuristics. Every entry needs a real, checkable source - an announcement, an arXiv preprint or an article, not a summary of a summary. If a result is contested, unreviewed or partial, say so in the status and verification fields rather than leaving it out. Submissions are reviewed before they appear. You can submit up to three entries per day. Full criteria in the methodology.
The result
What was solved, and where it is written down. Only you can supply this.
Compact label used on chart axes and narrow layouts.
Plain-language statement. Math works: $inline$ or $$display$$.
YYYY, YYYY-MM or YYYY-MM-DD. For a range, the completion date.
The AI's part
Which model, what it actually contributed, and how checked the mathematics is.
How much of the mathematics the model contributed, going by the authors' own disclosure - pick the lower tier when it is vague. Writing or proofreading alone is out of scope entirely.
Math works here too.
How checked the mathematics is. Changing this requires updating the note below in the same edit, so the reason is on record.
Where the claim lives: a bare announcement, a preprint, or a peer-reviewed venue. Changing this also requires updating the verification note.
The prose explaining how strongly this result is checked.
Context
Helpful but not essential - a reviewer can fill or correct any of it.
The bucket the field filter groups by.
Free-text subfield shown on the card, e.g. Additive combinatorics.
What happened to the problem. Changing this requires updating the verification note in the same edit.
Four-digit year, or blank if unknown.
Comma-separated names.
The decisive step: an explicit object, a finite certificate or case analysis, or a conceptual proof. Classify by what the result hinged on.
Short caveat for results that aren't cleanly proved/disproved.
Beyond the primary source: Lean repositories, independent proofs of the same theorem, verifiers, community records.
Rarely neededSkip unless one genuinely applies. Most entries leave every one of these blank.
If this is a numbered Erdős problem. Blank otherwise.
Only when a documented issue exists with the claim (refuted lemma, misformalized statement). Renders as a visible flag.
Whole dollars, only when a source actually states what the result cost. Never estimate it from model pricing - leave blank if undisclosed.
What the figure covers and where it came from, e.g. an aggregate across several results.
Looked-up count, or blank.