JevMade hello@JevMade.com
← Back to guides

JevMade field notes / Model and benchmark report

Von

Von documents a local decision model and evaluates option-order sensitivity on JevBench alongside eight-seed ViZDoom runs.

Original by wfzyxEvaluationGitHub repositorySource reviewed

Before you dive in

What you’ll find in the original

  1. Shuffle identical options and measure answer flips; typed output alone does not make a model order-invariant.
  2. Break results into easy, standard, and hard tiers rather than hiding difficult cases in one average.
  3. Report language-task accuracy and interactive-control outcomes separately because latency and loop design affect the latter.
Worth knowing

The author reports hard-tier option-order flips of 49.5% for the untrained baseline and 0.0% for Von. All comparisons, including hosted Jev rows, are repository-reported and were not rerun.