JevMade field notes / Reproducible evaluation guide
LeJudge: A cost module you program in English
A research guide that explains a programmable planning-cost method, then provides reproduction gates, probes, judge-only studies, start-selection comparisons, confidence gating, ablations, and integration guidance.
Original by AbdelStarkEvaluationRepository READMESource reviewed
Before you dive in
What you’ll find in the original
Separate the programmable judge from the planner so the cost signal can be probed before end-to-end integration.
Use both random and relevance-filtered expert starts to expose how candidate selection changes outcomes.
Treat confidence gating as an ablation target and require the reproduction gate before comparing reported studies.
Worth knowing
Study outcomes are author-reported and were not rerun for this review.