Echo prefers LEFT.

Echo supplies move preferences. Search simulates every offered move and reweights each preference by exp(child value), then normalizes. The weighted average of the child values becomes the parent value. A death has value −(score + 10), here −30. A surviving leaf with no food gained has value 0. Approximately-zero parent values are shown as ≈0. Search weights are not probabilities of survival.

This is seed 9000, move 176 from the Jev project. The explanation extracts the two levels that reveal the trap. After the original DOWN, Echo prefers LEFT at 79.7%; that move collides. DOWN and RIGHT survive, leaving revised weights of 91.9% and 8.1%.

The camera examines one branch at a time. The actual implementation expands all offered moves breadth-first, evaluates living boards in one Echo batch per level, and backs up values after growing the tree. The complete recorded search reached nine levels and evaluated 482 boards. This is a teaching sequence, not a measured timing replay.

Recorded trace. No new API calls were needed.