Nook is NukkAI's constraint solver. We measured it against the best solutions ever published, and against the leading academic and commercial solvers on the same machine.
Routing, scheduling and network problems from the research benchmarks that academic and commercial solvers measure themselves against.
Most large problems have never been solved to a proven optimum. So every solver is scored against the best solution anyone has published: the gap is how far its answer sits above that, in percent.
Lower is better. 0% matches the best known solution, and a negative gap beats it: a new record.
Every comparison ran on one machine, with the same time limit for each solver.
FrontierCO (2025) sets eight classic optimisation problems at frontier scale, up to millions of nodes. The published methods had one hour per instance; Nook had five minutes on one core.
With 3,000 to 30,000 customers and the same 300 seconds on the same machine, nook's routes average 2.1% more distance than the best known. HGS-CVRP's average 7.6%.
On FrontierCO's 16 hard independent-set instances, nook ran beside KaMIS, the specialised solver, on the same machine for the same 300 seconds.
RG300 is the hardest standard set for project scheduling: 480 projects of 300 activities, 347 never solved to proven optimality. Project length after 60 seconds; shorter is better.
