Would the user accept this turn?

Yes/no and rating about the agent's own final message, before handing back.

Call the local server

curl -s localhost:21118/v1/systemone -d '{"state": "User request: fix the failing test\nAgent reply: I fixed it and ran go test, all pass", "questions": {"accept": {"type": "noul", "instructions": "Would the user accept this reply as done?"}, "more": {"type": "noul", "instructions": "Would the user want more than this reply gives?"}, "satisfaction": {"type": "score", "instructions": "How satisfied would the user be with this reply?", "criteria": ["unhappy", "neutral", "happy"]}}}' | jq -c .answers

Real answer (openjevx v0.4.0 8-bit model, CPU):

{"accept":{"action":{"act_probability":1},"answer_confidence":0.6548,"confidence":0.6548,"noul":0.6548,"probabilities":{"false":0.3452,"true":0.6548},"type":"noul"},"more":{"action":{"act_probability":1},"answer_confidence":0.7412,"confidence":0.7412,"noul":0.2588,"probabilities":{"false":0.7412,"true":0.2588},"type":"noul"},"satisfaction":{"action":{"act_probability":1},"answer_confidence":0.7421,"confidence":0.7421,"probabilities":{"0":0.1084,"1":0.1495,"2":0.7421},"score":1.6336,"type":"score"}}

jevx judge sends its own built-in question set; the curl above is a hand-written equivalent with three plain questions, not the exact questions jevx judge sends.

Same thing with jevx

jevx judge --profile openjevx --no-context --request "fix the failing test" --proposal "I fixed it and ran go test, all pass"

Real output:

accept        0.67
wanted more   0.54
reaction      accept (0.53, confidence 0.53)
satisfaction  2.8 / 4  (openjevx)

What to do with the answer

If accept is low, add evidence (the test output); if wanting more is high, finish the missing part.

How the local model did

Leans accept both ways: 0.65 on the curl questions and 0.67 from jevx judge, but neither is confident.