Beyond steering
Sandbox asks: what happens if I change this feature? Counterfactual Studio asks: what happens if I change this token, and how did that propagate through features? The two are complementary tools for two different debugging questions.
Back to Laboratory
Q3 2026
Counterfactual Studio
'What if this token were X?' Surgical replay. Edit one token in the middle of a generation; see how features downstream shift; compare full trajectories. Feels like a React time-travel debugger but for the residual stream.
Prefix replay
Pick any token in an existing trace. Edit its value. We re-run the forward pass from that point (cached prefix), show you the new features, the new output tokens, and — critically — the feature-level diff vs original. Cached prefixes make this fast enough to explore interactively.
Research uses
Chain-of-thought robustness studies, prompt-injection susceptibility mapping, "what is the minimum edit that flips the answer" questions. Q3 2026.
Request early access
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.