Empirical neural scaling laws describe how model capability grows as a power-law function of parameters, data, and compute. They are accurate within the regime they measure, but that regime assumes an expanding supply of high-quality training data. As the stock of usable human text approaches exhaustion, the marginal return on acquiring further data declines. This note proposes an alternative axis of improvement, which I term scaling in: extracting greater capability from a fixed corpus by applying structured multi-perspective reflection during training, such that the resulting depth is written into the model weights. I argue, on information-theoretic grounds, that reflection does not add information to a corpus; rather, it elicits latent capability more efficiently than re-reading the corpus or padding it with unstructured generated text. I distinguish this training-time claim from inference-time perspective methods, identify the specific contribution as the combination of multi-perspective structure with held-out depth probes and a volume-matched unstructured control, and propose a falsifiable experiment to test it. A working system, Second River, is presented as an existence proof that sequential integration of reflected experience is implementable.
Dmitry Negai (Mon,) studied this question.