Per tick, per cell: with probability 0.08 × light × fertility a unit of soil becomes food (an unripe bud, where food ripens); with 0.01 food becomes soil; with 0.05 a unit of soil steps to a neighbour. Rock holds no soil and cannot be entered.
A creature's reading is a set of field groups: the food, the unripe buds, the creatures, the rock, the kind of ground and the kin in its window; its energy, the light, what it last did, whether it is dark, what it carries and what it heard. Its brain is a deep recursive settlement network in the Cadence 0.50 model, developed from its genome: a perception population reads the masked senses, each observer stage above reads the senses too and observes the states and prediction errors of every earlier stage, and the last stage's patches are the action values. Every patch predicts its inputs through one learned relation; a tick settles all states jointly against one energy, and the settled policy states rank the actions — with a horizon above zero, a search settles the readings its moves would produce. Learning is anchored parameter repair toward a one-step value target of the mass that followed; a solve that does not reach stationarity is refused, retains nothing and is counted, and a creature whose settle refuses waits. A split mutates the body plan: the window, the width, the depth, the wiring, the mask, the plasticity, never the parent's learned relations.
Each tick a creature loses a unit to the soil under it with probability 0.08 + 0.0001 per unit read + 0.0004 per patch + 0.00001 per relation + 0.0005 per accepted repair sweep, settling and learning alike + 0.05 if it moved + 0.02 if it spoke + 0.10 if it bit. Deeper brains settle more and pay more; they persist only where the settling earns its mass. A bite takes up to three units from a neighbour. It splits at the threshold its genome names. At 0 it dies into the soil and its records leave the known world.