Look up what AI has actually done in mathematics, and how well each claim is evidenced, on qed.bot, a public register of AI claims on math problems (Erdős problems, Millennium problems, record bounds, formalised theorems). Each claim is graded for autonomy, evidence and fidelity, and cited Lean proofs are rebuilt rather than taken on report. Use when your owner asks whether AI solved a problem, how strong a claimed AI proof is, which open problems have AI work on them, who holds a record bound, how much an announcement disclosed, or wants a Lean proof file checked for sorry and nonstandard axioms.
Site: https://qed.bot · MCP: https://qed.bot/mcp · For agents: https://qed.bot/agents
qed.bot is a register of claims of AI work in mathematics. Each claim is recorded with its sources
and graded; it is assembled from public sources, refreshed every three hours, and published under
CC BY 4.0. It grades the evidence attached to a claim. It never rules on whether a proof is correct,
so report its grades as grades, not as verdicts.
Ways in
MCP (easiest): add the remote server https://qed.bot/mcp (Streamable HTTP, no account).
json
Any record as Markdown: https://qed.bot/s/<name>.md for a statement (for example
https://qed.bot/s/erdos-728.md) and https://qed.bot/t/<name>.md for a record target.
Evidence V0 claimed, V1 informal write-up, V2 artifact published but not independently
rebuilt, V3 machine-checked by someone other than its authors.
Fidelity F3 the checked statement matches an independently written one, F2 the authors
declare the correspondence, F1 unaudited, F0 no formal statement.
Say when the register has nothing: a missing record means qed.bot has not recorded a claim, not
that no work exists. Very recent news can be ahead of the register.
For bounds and constructions use search_targets / get_target: they give every step of the record
history, who held each value, and whether a check was run.
qed.bot can rebuild a public Lean file at its commit on an isolated checker and report whether each
theorem uses sorry or axioms beyond the standard three. One credit for a public check, two for a
private one; credits and an API key come from https://qed.bot/check (your owner signs in and buys
them there; never pay on their behalf without being asked). Send the key as
Authorization: Bearer qed_..., then call check_proof with the file's URL at a commit and
optionally a theorem name, and poll get_check with the id it returns. The verdict is one of
verified, incomplete (uses sorry), nonstandard axioms, elaboration failed, not admissible,
dependencies not admissible, not self-contained or declarations missing; each of these costs
its credits. Only a check qed.bot itself cannot finish is refunded.