AN OPEN RETRIEVAL EXPERIMENT
Follow an edge.
Carry the evidence.
A small judgment at each hop. Several paths kept alive. Watch what happens when the next decision can use what the graph just revealed.
Recorded Jev calls · Synthetic graphs · No key needed
Loading recordings…
Where the walk goes
RetrievedCurrent hopNot yet visitedSelect a node to inspect its sources
0 / 0
SAME GRAPH, DIFFERENT WALKS
What changes?
| Strategy | Required evidence | Nodes read | Model calls | Input tokens | Stopped because |
|---|
Required evidence is scored against separate authored labels. It measures retrieval, not answer correctness. Token counts include reused recordings. Select a row to replay that strategy.
THE MECHANISM
The context moves
with the path.
- Read
Fetch an entity's source records. Keep their IDs and assertion status.
- Judge
Score each outgoing edge against the question and that path's evidence.
- Branch
Keep useful alternatives within a fixed budget. Repeat with new evidence.