Benchmarks & research

Jev next-best-action benchmark

Dorian Smiley tests whether Jev can predict a program’s next state from its current partial state.

What it does

Across 25 cases run 20 times, Smiley reports 98.6% canonical accuracy and 41.5% generalization accuracy.

Maker-reported (not independently measured by JevMade): 98.6% canonical accuracy · 41.5% generalization accuracy · 25 cases run 20 times

Source ↗

Source checked 2026-09-25 — opened the primary source directly.