Can better inputs fix Snake?

Inspect the controlled experiment. Watch every route decision.

12 × 12Classic · solid boundaries

Loading recorded games…Head Food

Load a run and watch every move.

DifficultyControllerRun

Replay speed

0 / 0 moves

—Food / 8

—Moves

—Decision latency

—Input / decision

Every replay comes from a saved game.

Last decision

Model proposal and executed move.

Left—

Straight—

Right—

Advance the replay to inspect a decision.

Held-out Classic games

Eight-food target · 16 seeds per profile

Controller	Targets reached	Shortest proposals	Code overrides

What does “efficient” mean here?

Code uses exact A* search over the moving snake body to find a shortest legal route to the current food. In assisted mode Jev receives those computed route lengths. Verified mode additionally replaces any non-shortest proposal. Planner-only mode makes no model calls. This is explicit algorithmic assistance, not evidence that Jev independently plans perfectly.

Eight foods is a capped task, not filling the board. Shortest routes to the current food do not prove the best route across future food or prevent every future trap. Fresh live games are outside the frozen study. Each live model move is a paid request; the board waits for the reply.

Local live server: python -m snake_demo.research.server. API key stays on the server. Original four-difficulty experiment · Full diagnosis, hypotheses and evidence