OPEN LINK ↗
// 1h agoPRODUCT UPDATE
MLflow 3.17.0 adds typed Jev decisions
MLflow 3.17.0 integrates TypeSafe’s Jev decision models into evaluation, tracing, and AI Gateway workflows, enabling typed yes/no and finite-category judgments with recorded probabilities. The release also adds fine-grained resource permissions and faster trace analytics. [MLflow release notes](https://www.mlflow.org/releases/3.17.0/)
// ANALYSIS
This is a meaningful LLMOps upgrade: MLflow treats evaluation judges as governed, observable infrastructure instead of opaque model calls.
- –Jev can reduce latency and cost for repetitive classification-style evaluations
- –Typed outputs avoid parsing generated text, while probability metadata exposes borderline decisions
- –AI Gateway support centralizes credentials, budgets, routing, and fallback behavior
- –Versioned Jev model IDs improve reproducibility; `jev-latest` is better suited to experimentation
- –The integration is specialized: it returns decisions rather than explanations, so teams still need general-purpose judges for open-ended evaluation [TypeSafe judge documentation](https://mlflow.org/docs/latest/genai/eval-monitor/scorers/llm-judge/custom-judges/typesafe/)
// TAGS
mlflowevaluationmlopsobservabilitystructured-outputopen-source
DISCOVERED
1h ago
2026-10-08
PUBLISHED
1h ago
2026-10-08
RELEVANCE
9/ 10
AUTHOR
MLflow