| Goal | Steps | Coverage | Excess retrace |
|---|---|---|---|
| — | — | — | — |
| Goal | Steps | Coverage | Excess retrace |
|---|---|---|---|
| — | — | — | — |
The goal is visible to you but is never supplied to either controller. Both see the same local walls and previous movement. The organ starts blank and can attach branch marks to random place-codes; the liquid-only controller receives zeros in that socket. Hebbian evolution, when enabled, changes only inherited mutations between generations—never weights during a maze.
| Condition | Solved | Excess retrace | Mean steps |
|---|---|---|---|
| Evolve champions first. | |||
There is no guided pass. On a new tree, braided, or loopy maze, the creature explores until it finds a hidden goal. The key waste metric counts every corridor traversal after its first outward and first return traversal.
Both populations evolve the same mixed liquid-and-spiking recurrent controller with a direct four-direction action head. Twelve-dimensional place-codes reduce collisions in larger mazes. Organ capacity is explicit: when slots run out, the least recently used location is replaced.
Focus changes only after held-out thresholds are crossed: first valid exploration, then dead-end discipline, then goal solving. If the new focus destroys validated ability, the population is repopulated from its best checkpoint. A small branch-marking scaffold is present only before the final solving focus.
The liquid-only champion receives the same maze budget and evolves independently. Biopsy can reset its recurrent state every move, hide its previous-direction signal, or combine those knockouts. The organ lesion leaves direct action intact.
A credible memory claim requires solutions on held-out mazes, few excess-retrace steps, and a selective collapse under the relevant memory knockout. Final results report raw solutions and raw excess retracing separately instead of hiding them inside one reward score.
Hebbian evolution records activity but never changes weights during a maze. Between generations, every active parameter family in both populations receives more frequent, gentler inherited mutations; inactive circuitry retains background exploration, and the total expected mutation energy is matched to uniform evolution. One run is still not proof.