RSI — Capacity Growth in Verified Self-Improvement

RSI asks whether function-preserving model expansion can restore learning after a language model reaches a plateau in verified self-training. The intended setting is a model that proposes tasks and learns from responses accepted by an objective verifier. Before testing expansion, I need a self-training process that reliably produces useful targets and can recover its full state after interruption.

I implemented the task-proposal and verification pipeline, accepted-response training, held-out evaluation, compute accounting, and experiment-state recovery. Initial controls demonstrated supervised-training feasibility and agreement across saved-state restarts within the tested conditions.

The autonomous pilot did not generate usable training tasks, and a skipped training update prevented the full recovery check from passing. This exposed problems in the training process, rather than evidence that the model had reached a learning plateau.

I am working on reliable verified-target production and fresh-session recovery before testing a plateau or expanding the model. The eventual comparison will need to distinguish capacity growth from continued adaptation. The growth experiment has not begun, so recursive learning benefits and improved held-out performance remain hypotheses.