Why it exists
Model evaluations are often compressed into a score without the decisions that produced it. Decivo keeps those decisions close to the evidence: the task definition, the sample, the labels, and the limits of the result.
The first published study compares Jev 1.13 with GPT-4o-mini on 50 synthetic customer-support messages. Both models matched all 50 expected labels. The published evidence includes model outputs, end-to-end latency, reported usage costs, and exact served revisions.
Decivo is an independent project. It is not affiliated with TypeSafe AIor OpenRouter.
Scope of the evidence
The frozen rubric uses six labels, including other, and states how to treat negation and overlapping intents. The 50 cases are AI-authored synthetic text and AI-reviewed against that rubric; no human or expert labeling is claimed.
This is one run per model on a small, fixed sample. A perfect score on these cases does not establish general superiority or production accuracy. Expected labels and review rationales were excluded from the model inputs.