A field trial of Jev
github.com/NaluKicks-808/jev-field-trial
Jev is TypeSafe's fast judgment model: you hand it a small piece of text and a menu of answers, and it hands back a probability for each one. I tested it on real decisions inside my second brain and my Claude Code history, wrote the pass mark down before every run, and published what failed beside what passed.
Why I built it
Every build I could find reported its own numbers, and nearly all of them were wins. I wanted to know which decisions in my own system the model could really make, and where it beats plain code, before I built anything on it. About half of what I tried failed, and two of my own written predictions were wrong. That turned out to be the useful part.
Start here
- Open the repository and read the results table in the README. It is the whole story on one screen.
- Read PLAYBOOK.md, one page on how to ask the model well.
- Download the repository as a ZIP, or clone it, and run the examples in dry-run mode. Nothing is sent and nothing is spent.
- To test it on your own data, copy the plan template, write your pass mark first, then add your key and run.
What it does not do
It does not include my data or my results files, only the method and the tools. The blind judging in my trial was done by a Claude session under written rules, not by an independent person, and the README says so.