Inside the maze.

Watch four models navigate the same walls, keys and doors. Decoice completed 3 of 10 development episodes. Replay any run to see the decisions that led to its outcome.

Explore all four games
Episodes completed
TaskDecoiceSarvam PrismJev 1.14Jev 1.13
Maze3 / 101 / 101 / 102 / 10

RECORDED API DECISIONS

Follow every decision.

Choose a task and step through the moves returned by the model.

One decision every 0.65 s for inspection. Animation speed is not model speed.

● current position · ↗ exit · ⚿ key · ▣ locked door · dark cells: walls / obstacles

Decoice

READY

●
↗

Starting state · key not collected

0 / 4 recorded decisions

Recorded result & timing

Did not complete · 4 decisions · 2.707 s full-task wall time.

Prism

READY

●
↗

Starting state · key not collected

0 / 3 recorded decisions

Recorded result & timing

Did not complete · 3 decisions · 0.261 s summed API time.

Jev 1.14

READY

●
↗

Starting state · key not collected

0 / 4 recorded decisions

Recorded result & timing

Did not complete · 4 decisions · 2.092 s full-task wall time.

Jev 1.13

READY

●
↗

Starting state · key not collected

0 / 3 recorded decisions

Recorded result & timing

Did not complete · 3 decisions · 1.912 s full-task wall time.

Progress shows how many recorded moves you have watched. API playback uses the measured response time of each move.

How to read a run

Each model starts with the same board and rules. After each choice the simulator sends that model its new state. A run ends at the goal, an illegal move, a repeated state or the move limit. The same development tasks were used for each model. These episodes are separate from the held-out language batches.

Use readable steps to compare decisions, or switch to recorded API durations. Decoice used a research HTTP endpoint, Prism used a local socket, and the Jev calls used hosted HTTPS. Transport contributes to these times.

See the four-model language evaluation