NewJevTree
Compose local Jev choices into multi-step probability trees for finite decision tasks.
Added to Jevfast
NewHow it uses Jev
JevTree enumerates a finite task's possible states and legal actions, then batches typed Choice questions asking which action is most likely to complete the goal from each state. It combines those local action distributions with transition and terminal-outcome logic to compare full paths and derive a policy.
What Jev decides
Three-step release workflow choice (illustrative)Example answers · not a recorded Jev response · Source ↗
Question 1 · d0 s0
YOUR APP
INSTRUCTION
Choose the action most likely to complete a reliable release.
STATE
The source-defined adapter starts at step 0 with risk 0 and 0 minutes. At that step, legal actions are `fresh_build` and `reuse_cache`; their next-state risk and time are included in criteria. The chart uses the source instruction and string action IDs. The enumerator names the depth-zero first-state request d0_s0.
JEV · CHOICE
- fresh_build
- reuse_cache
App workflow
- Define task adapter
Describe the workflow state, legal actions, action transitions, and executable terminal success check.
- Enumerate decision tree
Build reachable state/action paths to the configured depth and ask the backend for batched local Choice distributions.
- Compose path probabilities
Combine Jev action distributions with transitions and terminal outcomes, including the configured exploration mixture when enabled.
- Derive policy and report
Select a policy path and emit outcome probabilities, probability-mass error, and provider request usage.
Why it is interesting
It separates local judgments from long-horizon planning. Jev supplies per-state action probabilities, while the adapter defines legal actions, state transitions, and terminal verification; the tree search composes future path quality instead of choosing only the next locally highest-probability action.