Life-sim
An interactive evolution playground. Neural-net swimmers that learn within their own lifetimes, under purely economic selection, with no fitness function. WebGL2.
An evolution playground. Neural-net creatures swim, eat, learn, and die in a browser tank, and selection is economic, energy in and energy out, not a fitness function you wrote for them. Behaviour emerges instead of being scripted.
Creatures learn within their own lifetimes through reward-modulated Hebbian plasticity (the Baldwin effect made operational), so useful reactions can appear in a single life and then get fixed by selection across generations. The world is a 2D toroidal ocean of heterogeneous food pockets cut by a divergence-free flow field, with walls and environmental sensing. The demo world is tuned as a grazer-predator system: food concentrated into deep, isolated pockets so local density makes predation geometrically possible, with a diet-axis disruptive-selection knob so a bimodal grazer/predator split persists past twenty thousand ticks instead of collapsing into one type.
It is built feel-first. You can reach in while it runs to tune any knob, drop food, and follow one creature while watching its brain fire. Population-level evolution is readable without opening a panel. Dot colour is diet by default, a vivid green-grazer to red-predator ramp, and the population graph carries a stacked grazer/omnivore/predator band plus a diet-diversity line. New mechanisms ship as off-by-default knobs that are byte-identical when off, so the demo, the headless harness, and the tests all share one source of truth.