Agent tooling
shad0w records a model's fixed-choice answers, learns a small local lookup table and uses the upstream model for requests the table cannot confidently answer.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
It can sit in front of Jev, other System One services or chat models. Gaurav Chauhan reports its speed and disagreement figures. Matching the teacher's answer does not establish truth, and the README includes a dataset where disagreement exceeded the target. Changed traffic therefore needs fresh checks.
Use it for repeated choices such as assigning a message to a team, not open-ended writing. Let it collect answers before serving its own table. Watch the spot checks and retrain when your traffic changes; copying a model's answer can also copy its error.