Kernellum is an independent AI research lab building systems for machine reasoning, scientific discovery, AI systems & architecture, and evaluation & reliability.
K1 preserved workload rankings across 9/9 routed workloads after final-route correction.
SIMILARITY passed 13/13 preregistered criteria across 96/96 routed implementations.
Current Architecture Evidence
Final-route feedback improved the routed optimum on 12/12 Transformer GEMMs; equal-budget random search improved 0/12.
A preregistered 96-route ECP5 study predicted held-out timing tax with 0.3039 ns MAE and 0.8921 correlation.
A two-candidate local portfolio beat broadcast-only in 36/36 unseen deployments, with 23.56% mean latency improvement and 1.44% mean oracle regret.
