Jev / UnassistedAI decision-making in a simple world.

Better inputs. No answers supplied.

Watch Jev choose every move. Compare the held-out tests.

12 × 12Classic · solid boundaries
Loading recorded games…Head Food
Load a run and watch every move.
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
ControllerTargets reachedShortest proposalsMean food

What does “unassisted” mean?

Jev receives observable board state, rules and action meanings. Code supplies no safe-move flags, distance scores, route plans, ranked options or corrected answers. All three turn commands stay available. The model’s choice executes unchanged, including collisions.

Head-relative coordinates rotate and translate all geometry so forward is always positive y. That is explicit observation preprocessing, not a route solution. Prose changes the description format. The original prompt is the control. Offline shortest-path labels are computed only after games finish and are never sent to Jev.

Eight foods is the capped task. Outcomes include failed games. Live calls are separate from study results, and each move makes a paid request. Local server: python -m snake_demo.unassisted.server. Keys stay server-side.

Full findings, screening variants and limitations · Earlier assisted study