Drag-and-drop steering. Suppress feature A, amplify feature B, compose the intervention, hit play, watch the model regenerate. Save as a recipe URL. Export as a .pt that plugs into transformers in one line.
Current tools (Goodfire Ember, SAELens) support single-feature interventions. Sandbox supports composable stacks: amplify [honesty, hedging] + suppress [overconfidence, sycophancy], serialize the stack, reuse across prompts and models.
Every change re-runs a 4-token prefix forward pass under your subscription budget and updates the response below. 200ms latency target. This is only possible with our Feature Firehose infra (shared with Watchtower).
Your steering stack exports as: (1) a shareable URL anyone can open, (2) a `.pt` file for `transformers.apply_interventions()`, (3) a Goodfire-compatible `.json`, (4) a PyTorch hook you can paste into your own code.
We prioritize researchers, educators, and safety teams who will use it publicly. Tell us what you want to build; we'll reach out when the beta opens.