TL;DR: Small models aren't dumb, they're shallow. I designed a cross-domain, blind, visual experiment to see if a large model can compress its "planning discipline" into a reusable scaffold that makes a small model deeper — with zero fine-tuning. Three.js is the testbed because you can't fake structure with verbose text; the render exposes everything. I’ve been spending a lot of time testing smaller models (like 9B parameters), and I’ve noticed something: they aren’t exactly dumb, ...