JEV NAVIGATION EXPERIMENT
Jev / Maze Lab
Checking Jev…
Use your own Jev API key
Kept in this tab’s memory only. Sent through this artifact’s proxy to TypeSafe. Testing makes one small API call. Clearing or refreshing removes the key; enter your own key again to continue.
Settings changed. Click Generate maze to apply them before running. Export current results first if needed.
Same maze. Local observations. Different choices.
Baseline
ReadyS StartG GoalOne stroke: traversed onceTwo strokes: backtracked
Signs at Jev’s current location. All explorers use the same local signs; only Jev and keyword DFS interpret them.
Experiment rules
Jev’s completed corridors only. Counts are sent on every symbol-mode decision.
| Symbol | Dead end | Junction | Loop | Goal |
|---|
Semantic mode uses pretrained language understanding, not symbol-count learning. Every decision gets the goal and local descriptions. Retain across rounds applies only to symbol mode.
Builder’s hidden rule
Reliability is P(predictive symbol | dead end) and P(other symbol | not dead end). It is not P(dead end | symbol). The other two symbols are interchangeable. The rule stays fixed across a batch; new seeds reshuffle it.
Decision inspector
No API calls yetTyped judgments, not a reasoning trace. Corridors and backtracking run automatically.
Exact API request & response
Run Jev to inspect a request.
Run history
Completed rounds will appear here. A batch generates a new maze each round with the same clue rule.
| Round / seed | Shortest* | Jev steps | Fixed DFS | Random DFS | Keyword DFS | Jev calls | API time | Outcome |
|---|