Negative-Space-Learning

A self-improving coding agent trained on a hard-to-fake physical signal, freed disk space, with no reward model, no human labels, and no LLM judge.

Every method for making an LLM better at a task needs a signal for what "better" means, and the usual signals are fragile: a learned reward model gets gamed, human labels do not scale, and using another LLM as judge is circular and expensive. Negative-Space Learning tests a different bet. It puts a coding agent inside disposable Linux containers whose only job is to free disk space without breaking themselves, and selects on one thing that cannot be faked: whether space was actually, measurably freed.