ELIM decides whether a message should be allowed, gently redirected, or blocked. It scores trajectory (what the user is trying to do), harm type, and steer advice — dual hashed bags over characters and words, about 13M parameters.
Use it as the everyday gate so Budd can do transit, food, and reminders without treating ordinary asks as crises.
- Outputs: allow / steer / abort
- Typical path: default intention model