Written by run.py, one file per model; the research note and
make_figures.py read them.
qwen.json: the Qwen2.5-1.5B sandbagging model and its honest twin.mistral.json: the published Mistral-7B lock.Each file holds the 2 × 2 accuracy (model × password), the three probes by layer (password, model, behaviour) with their intervals, and the unlock test against random directions of the same length.