INDEPENDENT AI-SAFETY RESEARCH
SP Lense studies how internal representations can help detect and steer model behavior. We investigate shutdown-related contexts, conditional activation steering, and context-aware interventions through reproducible experiments.
Our aim is to understand when an intervention works, where it fails, and how to evaluate its effects without overstating what the evidence shows.
Four connected studies explore detection, steering, and the limits of behavioral control.
Completed. Tests a guarded activation-steering rule in shutdown-related contexts, with a published paper and reproducibility package.
Investigates adaptive approaches to activation steering. Original research and paper snapshots remain available for inspection.
Ongoing, with published snapshots. Examines semantic gating and the transfer of fine-tuned behavior to steering, from binary shutdown choices to simulated multi-step tool workflows.
Ongoing. Investigates semantic detection, persistent episode state, and separate behavioral interventions in a bounded simulation. Development progress and limitations are documented in the study record.
We prioritize clear evaluation protocols, preserved study snapshots, and reproducible evidence. Negative results and limits are part of the research record.
These studies evaluate specific models and controlled tasks. They do not establish a general self-preservation mechanism or evidence of subjective experience.