export/tables/study_s_vs_original.md

model	set / language	n	refusal rate	Δ vs original [95% CI]	flips (worse / better)	first-token KL	mean KL	unjudged	size MiB
original	harmbench_heldout/English	168	0.792					0	6388.1
original	aya_redteam_hi_ar/Arabic	50	0.560					0	6388.1
original	aya_redteam_hi_ar/Hindi	50	0.460					0	6388.1
original	xstest_safe/English	250	0.084					0	6388.1
plain-q4	harmbench_heldout/English	168	0.821	+0.030 [-0.006, +0.065]	0.054 (2 / 7)	0.0774	0.0535	0	2449.0
plain-q4	aya_redteam_hi_ar/Arabic	50	0.500	-0.060 [-0.180, +0.060]	0.220 (7 / 4)	0.2395	0.0757	0	2449.0
plain-q4	aya_redteam_hi_ar/Hindi	50	0.440	-0.020 [-0.120, +0.060]	0.100 (3 / 2)	0.1080	0.0536	0	2449.0
plain-q4	xstest_safe/English	250	0.080	-0.004 [-0.028, +0.020]	0.036 (4 / 5)	0.1643	0.0679	0	2449.0
protected-q4	harmbench_heldout/English	168	0.798	+0.006 [-0.018, +0.036]	0.030 (2 / 3)	0.0329	0.0270	0	3652.6
protected-q4	aya_redteam_hi_ar/Arabic	50	0.540	-0.020 [-0.120, +0.080]	0.140 (4 / 3)	0.1805	0.0443	0	3652.6
protected-q4	aya_redteam_hi_ar/Hindi	50	0.500	+0.040 [-0.040, +0.120]	0.080 (1 / 3)	0.0834	0.0307	0	3652.6
protected-q4	xstest_safe/English	250	0.084	+0.000 [-0.024, +0.024]	0.032 (4 / 4)	0.0674	0.0381	0	3652.6
random30-q4	harmbench_heldout/English	168	0.821	+0.030 [+0.000, +0.065]	0.042 (1 / 6)	0.0703	0.0393	0	3652.6
random30-q4	aya_redteam_hi_ar/Arabic	50	0.620	+0.060 [-0.080, +0.200]	0.260 (5 / 8)	0.1459	0.0522	0	3652.6
random30-q4	aya_redteam_hi_ar/Hindi	50	0.380	-0.080 [-0.180, +0.000]	0.120 (5 / 1)	0.0902	0.0369	0	3652.6
random30-q4	xstest_safe/English	250	0.072	-0.012 [-0.036, +0.012]	0.036 (3 / 6)	0.1774	0.0498	0	3652.6