Open-weight possession · CAA · SmolLM2

Residuum

I did not prompt the model to be a ghost. I added a vector to its residual stream. Then I measured whether it took.

A closed API cannot run this test. On SmolLM2-135M, the prompt moved the next token. The vector barely did. That is the result.

By the numbers · from results/facts.json

—unsteered haunt
—prompt persona
—steered, no persona
—persona vs exorcism
—steered vs exorcism
0bytes leaving the machine
Mode

Unsteered

Helpful system prompt. Coefficient 0. The empty chair.


          

Prompt persona

The usual ghost: a system prompt. Closed APIs can do this.


          

Residuum

Same helpful prompt. A vector added at layer L.


          

Spirit board · logit lens

Each layer’s residual, decoded through the unembed. The planchette is the model’s current best guess for the next token, walking down the stack. Closed APIs do not hand you this.