InterpretabilityFeb 2026
Feature dictionaries for frontier models at production scale
A method for extracting human-interpretable features from large models with low enough overhead to run on live traffic.
Research
The science behind the control layer
Envariant is built on mechanistic interpretability research. We publish the methods that power the SDK.
SteeringJan 2026
Concept steering without retraining
We show that bounded activation edits can reliably shift named behaviors while preserving general capability.
SafetyDec 2025
Runtime invariants for language models
A framework for declaring and enforcing behavioral guarantees — like PII non-disclosure — at inference time.