Agent tooling
shad0w learns a small local lookup table from a model's fixed-choice answers and sends requests back to that model when the table is unsure.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
The proxy can front Jev, other System One services or chat models. Its speed and disagreement figures are Gaurav Chauhan's reports. Agreement with a teacher model is not proof of truth, and one README dataset exceeds the disagreement target. Traffic changes call for fresh checks.
Use shad0w for recurring choices, such as which team should receive a message, rather than open-ended writing. First let it collect the model's answers, then let it serve its table. Watch the spot checks and retrain as traffic changes: the table can learn the model's errors along with its answers.