Scholar Loop
An autonomous, multi-agent AI research loop that reads real literature, forms a hypothesis, runs real machine-learning experiments on a single GPU, reflects on the results, and drafts a peer-reviewed write-up, with guards against reward-hacking.
Eight typed agents run the loop: a Director sets the next direction, topic and budget from a ledger and literature trends; a Lit Scout pulls citation-ranked papers from arXiv and OpenAlex into cited findings; a Reasoner proposes the next experiment from constraints, literature and past lessons; a Debate Panel of three personas votes on whether an idea is worth a GPU; a funnel screens ideas cheaply through smoke, verify and full tiers; a Runner executes a real PyTorch experiment scored against a frozen ground-truth metric; a Reflector turns the outcome into a lesson in a decaying skill library; an Advisor steers the loop to proceed, refine or pivot; and a Writer and Reviewer turn confirmed findings into a number-grounded draft and its peer review.
Two experiment domains ship today, digit classification and diabetes regression, and a new domain is added as a YAML profile plus an engine pair with no orchestrator changes. Ideas climb a parallel population funnel under a self-stopping governor that halts on budget, round cap or convergence, so a bad idea dies after one cheap run and a marginal one never burns a full run.