Small decisions. Interesting possibilities.Submit contentSubmit
NVIDIA NeMo Switchyard preview

NVIDIA NeMo Switchyard

Route agent requests between efficient and capable models using a typed Jev judgment.

Added to Jevfast

How it uses Jev

Switchyard can use Jev as a decision judge for model routing. It sends the task, anonymous candidate labels, capable/efficient comparison and supplied evidence. A Choice asks whether the capable candidate alone succeeds. Code routes to the capable model only when the advantage probability exceeds the configured cutoff.

What Jev decides

Source-defined request (illustrative)Example answers · not a recorded Jev response · Source ↗
Question 1 · route
YOUR APP
INSTRUCTION

Does the capable candidate succeed while the efficient candidate fails?

STATE

Task messages, anonymous candidate labels, a comparison identifying capable and efficient candidates, and configured reference evidence. Candidate labels are configuration-dependent; no particular model pair is asserted.

JEV · CHOICE
  1. advantage
  2. no_advantage

App workflow

  1. Configure routing

    The operator selects judge and completion targets, anonymous candidate mappings, evidence and cutoff.

  2. Judge the task

    The decision classifier sends the task and two candidate roles to the configured judge.

  3. Select the target

    Code compares the advantage probability with the cutoff and routes to capable or efficient completion.

Why it is interesting

Routing depends on relative advantage rather than the selected label alone. Shared failure belongs to no_advantage, and the code explicitly warns that this score is not a calibrated probability of task success.