Steerscope reads out what a real instruct model predicts at every one of its layers, so you can see the moment it makes up its mind. Then drag a slider to push a behaviour, refusal, sentiment, formality, into its residual stream and watch the sentence rewrite itself. The logit lens and the steering vectors are built from scratch.
The weather in Melbourne today is
Each slider adds a contrastive activation vector into the residual stream at a single mid-depth layer, scaled to the layer's natural norm. Zero is a provable no-op. Past the shaded band the output loses coherence.
At every layer the model holds a running guess of the next token in its residual stream. We apply the model's final norm and unembedding to each layer's hidden state, so you read out that guess directly. The math is implemented from scratch and checked against a reference to the last decimal.
nostalgebraist, 2020For each behaviour we compute a direction offline as the mean difference in activations between contrastive prompt pairs. At inference we add a scaled copy of that direction into the residual stream at one mid-depth layer, so a single slider nudges refusal, sentiment, or formality without touching a prompt.
Rimsky et al. 2023, Arditi et al. 2024Generation, the per-layer read, and the steering injection run entirely on your GPU through WebGPU. There is no inference server and no per-visit cost, which is only possible because the model is small and the heavy work stays on the device.
Qwen2.5-0.5B-InstructThe load-bearing engineering, the logit-lens math and the steering composition, is real and tested now. The browser demo drives those same visualisations with a labelled preview while the instrumented model, built by the pipeline in this repo, is wired in. Nothing here claims to be measured that is not.