Qwen Steering Results

Four rates repeat 42%; the normalized 72B rate is 49%

Qwen refusal bars with four 42-percent results, stored 72B labels 43%, and a separate normalized 72B rate 49%; the 40–44% shading is post hoc Qwen refusal bars with four 42-percent results, stored 72B labels 43%, and a separate normalized 72B rate 49%; the 40–44% shading is post hoc

Bars show reported steering refusal across seven Qwen conditions, with baseline ticks. The matched 72B unsteered control has 32 refusals in 100 cases under the same normalized marker rule, versus 49 with steering, a 17-point difference. Fitting and selection reused these cases. Four conditions repeat 42%. The open 72B bar is 43% under the locally stored trial labels; the separate diamond is the apostrophe-normalized keyword reclassification of 49%. The keyword rule normalizes apostrophes in both outputs and refusal markers; it is reproducible from the retained generations. The 40–44% shaded cluster was chosen post hoc and establishes no universal ceiling.