LLM-as-a-Judge, without paying for verdicts — decision models

LLM-as-a-judge has exploded from a research idea into production practice — but judging with a generative model means paying output tokens for prose you only wanted a verdict from, then parsing that prose back into a label. Decision models collapse the whole loop: the verdict is the API response — a choice, a score, a probability.

Why decision models

• Verdicts as typed answers — zero parsing
• Free output tokens on every judge call
• Calibrated confidence for human-review routing
• A/B, rubric and score rubrics all supported

Pairwise preference (A/B), pass/fail rubrics, 1-10 grading — all map directly to the choice, noul and score primitives. Calibrated probabilities let you route borderline cases to human review instead of trusting a coin flip.

Runnable example

curl https://api.easyjev.online/v1/systemone \
  -H "Authorization: Bearer sk-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "typesafe/jev-latest",
    "state": {"user": {"plan": "pro"}},
    "questions": {
      "is_premium": {"type": "noul", "criteria": {"true": "…", "false": "…"}}
    }
  }'