Jev MCP — arunav25

Use a decision model when you need a branchable probability. Use an LLM when you need wording, planning, or tool sequencing, then gate the risky step with a typed Jev question.

Install from the repository or link a clone, set the key before install so clients inherit it, then register for Claude Code, Desktop, or Codex. The bare registry name belongs to an unrelated project.

Put every fact in state, define vague words in the instructions, ask one judgment per question, add an escape option, and describe score levels as real situations. Keys and conclusions do not travel as evidence.

Small gaps need thousands of labels under paired testing, while latency and token differences show up on dozens of unlabelled runs. The harness reports proper scores, calibration, AUC, and paired tests with intervals.

One tool, evaluate. You pass a state object plus one or more typed questions and answers come back under the same keys. A noul takes optional true and false criteria and returns a probability; a choice requires a map of option to description or null and returns choice, probabilities and confidence; a score requires an ordered array of at least two level descriptions and returns a zero-indexed score, its legend, probabilities and confidence.

The label command never shows you a model's answer, and you add a second rater because two raters bound what is actually resolvable; agreement then reports Cohen's kappa between them. Both systems are asked the identical question, which lives in the dataset config rather than in the command line, and compare scores only items both systems covered so per-item difficulty cancels out.

Every call is timed and its token counts recorded, so perf reports p50, p90, p95 and p99 latency, tokens per call, and a paired median latency difference with a confidence interval. The argument is that a one or two point accuracy gap needs thousands of labelled items, while a system twice as slow is unmistakable across fifty unlabelled ones. Cost is reported only from rates you supply, since they change and are per account.