jevbooks

← All projects

jev-bench

TheWayWithin/jev-bench

Does the cited source actually say it? A 42-claim benchmark.

Evaluation & Benchmarking99%Runner-up: Verification & GuardrailsTiered thresholds

What it is

A 42-claim benchmark testing whether a cited passage supports a sentence as written, run against Jev and three frontier models, with raw outputs and scored results.

How it uses Jev

Jev answers one Choice question per claim: given a sentence and its cited passage, decide supported, unsupported, or not_addressed. Its typed answer and probability are used as a publishing-gate pre-filter, escalating low-confidence verdicts.

Primitives:choice

Technique worth stealing

Score calibration and risk-coverage separately from headline accuracy, since frontier models put all answers in one confidence bucket.

Try it

bash bench.sh --system jev --tier A, or python3 risk_coverage.py results/*.jsonl (no keys needed).

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-centricunclear0.47
  • Shows a System One patternno0.18
  • Handles uncertaintyno0.07
  • Measuredno0.39
  • Runnableno0.05
  • Worth recommendingno0.30
  • Model replicano0.05
  • Problem scopescore on a 0–2 scale1.06
  • About Jevyes0.98

Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 21 Sept 2026.