Somnium: A Recurrent World Model that Learns to Dream Before It Wakes

Preprint v1.0. Paper PDF attached; LaTeX source in the source archive.
Code, full run registry, and per-run records:
https://github.com/wilberforce1900/somnium

We train a small non-Transformer world model, Somnium, under a developmental
curriculum: chaos -> dream -> wake. A single recurrent substrate carries two
dynamics -- fast streaming expansion (yang) and energy-relaxation contraction
(yin) -- softly mixed by a gate; prediction happens in latent space (JEPA-style
next-latent), and training begins with a dream phase of self-generated rollouts
before any real environment interaction. Dream-first pretraining reliably cuts
the real-interaction steps needed to reach a fixed accuracy by ~60%: replicated
across five independent runs at 16K parameters (against a pre-registered
threshold). Three further lines of evidence -- a scale ladder, a dose-scaling
control, and 871-round autonomous growth runs -- converge on the same
conclusion: the advantage is capacity-bound. Alongside the positive result we
report a registry of pre-registered negative results (a VQ situation codebook
that hurts training; a learned gate with no accuracy gain whose polarity is a
one-way developmental ratchet; dream-statistics feedback with zero calibration
gain). All hypotheses, thresholds, and outcomes are logged in a public run
registry.

Writing note: the scientific work (experiments, code, 90+ run records) was led
by the author; text drafting was AI-assisted; the author takes full
responsibility for all content.

Keywords: world models; pretraining; self-supervised learning; continual learning; pre-registration; negative results