Kepler Gliese investigates how models learn, where they fail, and how to build reliable AI systems. We distinguish a trained neural capability from a retrieval result, a proposed experiment from an executed run, and a compelling explanation from a verified one.
Some work is publicly released; other work is active research. Labels describe evidence, not marketing readiness.
KIFF v0.1 is a 462,208-parameter neural model trained from scratch for limited structured sports-evidence questions. Its 360-question synthetic evaluation returned 175 correct answers (48.6%). The separate deterministic comparison answered 360 correctly. This is not demonstrated open-ended sports reasoning or screenshot comprehension.
Read the model card and limitations →An independent experimental research effort. More will be shared publicly when there are reviewed results worth discussing.
Follow Along for updates →Experiments investigating factual grounding, citation behavior, and content discoverability in AI-generated answers. The distinction between observed evidence and recommended changes remains central.
Get in touch →Data acquisition, model training, retrieval, evaluation, and application presentation are separate stages. A retrieved sports record does not mean a neural network learned it. A correct answer on a familiar example does not establish generalization. We preserve those distinctions across model versions.
Private training data, applicant information, and internal knowledge repositories are not hosted on this public research page. Publications have their own model cards, rights notices, source acknowledgments, and evaluation limits.
Kepler Gliese combines engineering discipline with long-term curiosity about model intelligence and applied scientific questions. Learn more about the company and our separate access initiative.
About Kepler Gliese → Follow Along on Substack → AFROTECH 2026 Access Initiative →