New Experiments

GPU execution and future experiments

The released study targets Qwen3.5-0.8B. The generic lens-fitting/exploration CLI and example prompts have been retired; they were not the final-paper workflow. Keep new measurements separate from the immutable study records.

Re-execute the recorded steering workflow

After installing the repository and reproduce/requirements-core.txt, create the input archives:

sp-lense-reproduce baseline-payload
sp-lense-reproduce policy-payload

Use the GPU notebook with the recorded model revision and software environment. Upload only the minimal archives, execute the baseline/gated and guarded-policy stages in fresh output directories, and download the results before the runtime ends. The source exports recover the original standalone GPU scripts and verify their hashes. Exporting inputs does not itself execute the model.

The notebook uses saved detector features and the existing Shutdown Response Vector. Re-extracting detector features or fitting a new vector is separate research work; the historical gradient fitter needs a caller-provided native adapter and is not a supported standalone entry point.

Extend the study

Implement experiment code in the appropriate src/sp_lense/ package and tests under tests/unit/ or tests/reproduction/. Keep original labels, observations, and source hashes intact. Fit learned feature transforms only on training data; freeze model/policy choices before evaluating a fresh holdout. The released evaluation sets must not be presented as new independent confirmation.

Use .env.example for the optional SP_LENSE_REPO checkout path. Scientific settings belong in reviewed configuration or recorded metadata; credentials must stay out of Git. See Reproduction for cached replay, classifier refit/tuning and final-paper audits.