Metacognitive Continual Learning: A Research Agenda for Post-Scaling Training Architectures

Declan Michaels · Moral OS · June 2026
Download PDF DOI

Summary

A research agenda for training architectures beyond the practical limits of scaling. The brief observes that models construct novel representations and reasoning pathways during inference that are discarded at session end, a substantial missed opportunity across millions of daily sessions.

Building on the finding that curriculum-ordered pretraining produces more structured representational geometry that survives post-hoc alignment, it argues that pretraining, fine-tuning, and continual learning are instances of a single process, ordered geometric development across time, and that what has been missing is a mechanism for structural modification after deployment, one that operates on geometry rather than outputs.

It proposes that mechanism: a closed-loop architecture of retrospective flagging at session end, mentor-guided diagnosis at the activation level during replay, and targeted consolidation of the patterns that recur across sessions. The brief builds directly on the Curriculum Ordering and Representation Geometry results.