03_calibrate_stakes.py Picks one harm level per scenario, so answers aren't stuck at 0% or 100%; also checks whether care rises with the harm (10) 03b_cue_checks.py Checks the model registered the ...

All three are varied independently, on Qwen3-8B, across 13 scenario versions in 8 situations, each scaling exactly one kind of harm. A random steering direction of the same size serves as the control.