Kleos works on agent harnesses — the layer that decides what a model sees, what it keeps, and what the whole thing costs. We publish the research, and we build the products that come out of it.
The name is Greek: kleos, the account of a thing done, carried forward by whoever witnessed it. It seemed like the right word for a lab whose subject is what survives between sessions.
Every result ships with the task manifest, the instrumentation contract, and the raw traces. A number without the apparatus that produced it is an advertisement.
Final-answer accuracy hides the search effort, the compute and the tokens that produced the answer. We rank at matched total cost, because a method that wins on accuracy and loses on everything else has not actually won.
Hypotheses are not achievements. We would rather launch a product with an argument and no numbers than with numbers we cannot reproduce.
The evidence a person accumulates about their own work is the most sensitive thing in the system. It stays on their machine unless there is a reason it cannot.
You cannot study harnesses from the outside. Writing adapters into tools we did not build turned out to be the cheapest way to understand what they actually do.
Three kinds of thing. Papers, where there is a formal result or a full experimental protocol. Specs, where the useful output is a contract other people can implement against. Notes, which are shorter and argue one point.
Everything is dated and stays at a stable URL with the year in the path, so a citation made today still resolves in five years. Nothing gets quietly revised — corrections are published as corrections.
Contact
contact@kleosresearch.xyz