Native representation engineering

Understand what a model represents. Change its direction.

Ster turns hidden states into explicit, measured steering artifacts — and applies them inside a Rust decoder loop built for exact intervention points.

direction
layer 18margin +2.41
Rust nativeCandle runtimeVersioned JSONCPU · Metal · CUDA

One continuous path

From contrastive examples to controlled generation.

No framework handoff, no opaque pickle, no guessed intervention point.

01

Read

Capture final-token residual states after the transformer blocks you select.

02

Learn

Fit CAA, PCA, or logistic directions from matched contrastive examples.

03

Measure

Compare layer and method candidates on held-out pair-ordering accuracy and margin.

04

Steer

Add a validated direction to the residual stream during native autoregressive generation.

Fail-safe artifacts

A vector carries its identity.

Every artifact records product and schema identity, model revision, hidden width, method, layers, accuracy, and margin. Ster refuses the wrong model or residual stream.

Artifact contract →
truthful.ster.jsonschema 1
model
meta-llama/Llama-3.2-1B
method
caa
layer
18
accuracy
0.92
margin
2.41

Install

One command. One native binary.

shellcargo install --git https://github.com/wisent-ai/ster --locked
Choose a backend

Ster 0.12

Read. Learn. Measure. Steer.

Open documentation