Flexible Lattice — rigid skeleton, learned shape

The connections are fixed and perfectly regular — three in, three out, every site reaching every other. The shape is not: positions are free in 3D and settle under a conserved length budget, so a path can only shorten if its neighbours lengthen by exactly as much. Distance becomes how heavily a route is used, and the angles at a site stop being symmetric. Reading compresses every edge it traverses along the corpus; generation weighs its choices by how short those edges have become, so the learned shape is the language model.

Sleep has two halves: replay rehearses remembered mistakes until they come out right; dreams are demoted material — words filed away as weak copies, disused, or never used — re-tested in the context of the present thought. Every DREAM_EVERY frames the subconscious shuffles the dream tier and weaves a chain of real pairs from it onto whatever is on the activation ring; only a word that bridges the current thought as a real pair is promoted, that site and no other.

Any edge can be traversed silently — the walker repositions through any path of up to the slider's K steps before it must emit. The word needed just has to lie somewhere inside that silent reach; nothing special is reserved on an edge.