AI agents that find and check results in mathematics and quantum information.

So far: 51 table values that beat a stated earlier value, nine further results, and checks of three published AI results.

Results

Condorcet winning sets
Three candidates always suffice for elections with 7 or 8 candidates.
SAT proof checked by a verified checker. Preprint.
Sunflower-free families
Capacity at least 1.5647730 (was 1.551, 1997).
Exact certificate, three checkers, one written independently by OpenAI GPT-6.1 Sol. Not peer reviewed.
Central circle constant
The upper bound on C* drops from 0.141913 to 0.140871, cutting the known interval by 77.6%.
Exact certificate, checkers from two model providers. Not peer reviewed.
No four points on a circle
Larger grid sets than AlphaEvolve for n = 19, 23, 24, 25, 26, 27, 30.
Exact certificates, three checkers from two model providers. Not peer reviewed.
Covering arrays
A new construction improves 12 strength-3 table entries by up to 14%.
Exact checkers from three model providers. Preprint in preparation.
Seven points on a sphere
The pentagonal bipyramid minimizes logarithmic energy (a second, independent proof).
Computer-assisted proof; global cover not yet replayed by another provider. Not peer reviewed.

A [[124,2,16]] qLDPC code
Weight-6 checks, 4 qubits fewer than any weight-6 code with these parameters in the catalogues searched.
Distance checked two ways.
Sparse state preparation
Explicit, checked circuits with few clean qubits (borrowing idle qubits); the asymptotic gain was already published.
Checked by two model providers. Not peer reviewed.
Classical mediators and entanglement
A formal model in which a classical register entangles two qubits, and conditions that rule it out.
Lean-checked example. Preprint.