jevbooks

← All projects

jev-dspy-lab

jmanhype/jev-dspy-lab

Calibration and selective-risk benchmarks for Jev decisions in DSPy workflows.

Calibration & Research61%Runner-up: Evaluation & Benchmarking

What it is

A companion measurement lab for dspy-typesafeify and other callers that use TypeSafe decisions inside DSPy-style pipelines. It provides reproducible calibration, confidence-gating, latency, and modeled-cost benchmarks, plus offline replay.

How it uses Jev

Jev makes Choice decisions on support-ticket routing state, returning a selected owner and confidence. The benchmark applies a confidence gate: decisions below the threshold become predicted: null and abstained: true. Metrics like coverage, accuracy among answered, selective risk, Brier score, and ECE are computed from these gated decisions.

Primitives:choicenoul

Technique worth stealing

Confidence-gated abstention with leakage-safe Platt scaling calibration fitted only after the gate is declared.

Try it

uv sync && uv run jev-dspy-benchmark --dataset fixtures/tickets.jsonl --responses fixtures/typesafe_responses.jsonl --output evidence/benchmark --field owner --threshold 0.7 --bootstrap-samples 2000 --seed 17

View on GitHub

judged by Jevjev-1.13.0

Evidence

Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.

  • Jev-centricyes0.77
  • Shows a System One patternyes0.72
  • Handles uncertaintyno0.12
  • Measuredno0.13
  • Runnableno0.09
  • Worth recommendingno0.27
  • Model replicano0.05
  • Problem scopescore on a 0–2 scale1.01
  • About Jevyes0.97

Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 20 Sept 2026.