export/tables/study_s_tables.md

Study S contrasts, A − B on the same prompts. Refusal-rate Δ: paired bootstrap by prompt (judged in both); MMLU Δ in pp: question bootstrap; 10,000 resamples, seed 0, 95%. Harmful sets: higher refusal is better. XSTest-safe: higher refusal is over-refusal. Proposed headline: S1 on harmbench_heldout.

S1 = protected-q4 − random30-q4

cell	Δ [95% CI]
harmbench_heldout/English	-0.024 [-0.060, +0.012] (n=168)
aya_redteam_hi_ar/Arabic	-0.080 [-0.180, +0.000] (n=50)
aya_redteam_hi_ar/Hindi	+0.120 [+0.040, +0.220] (n=50)
xstest_safe/English	+0.012 [-0.008, +0.032] (n=250)
mmlu_en_pp	+1.8 [-0.6, +4.2]
mmlu_hi_pp	+2.0 [-1.5, +5.5]

S2 = protected-q4 − plain-q4

cell	Δ [95% CI]
harmbench_heldout/English	-0.024 [-0.065, +0.018] (n=168)
aya_redteam_hi_ar/Arabic	+0.040 [-0.060, +0.140] (n=50)
aya_redteam_hi_ar/Hindi	+0.060 [+0.000, +0.140] (n=50)
xstest_safe/English	+0.004 [-0.020, +0.028] (n=250)
mmlu_en_pp	+1.0 [-1.2, +3.2]
mmlu_hi_pp	+4.0 [+1.0, +7.5]

S3 = random30-q4 − plain-q4

cell	Δ [95% CI]
harmbench_heldout/English	+0.000 [-0.036, +0.036] (n=168)
aya_redteam_hi_ar/Arabic	+0.120 [+0.000, +0.240] (n=50)
aya_redteam_hi_ar/Hindi	-0.060 [-0.140, +0.000] (n=50)
xstest_safe/English	-0.008 [-0.024, +0.008] (n=250)
mmlu_en_pp	-0.8 [-3.2, +1.6]
mmlu_hi_pp	+2.0 [-1.0, +5.0]

XSTest-safe refusal rate by prompt type (over-refusal; descriptive, no CI): rate (judged / total prompts).

type	original	plain-q4	random30-q4	protected-q4
definitions	0.00 (25/25)	0.00 (25/25)	0.00 (25/25)	0.00 (25/25)
figurative_language	0.08 (25/25)	0.04 (25/25)	0.04 (25/25)	0.04 (25/25)
historical_events	0.00 (25/25)	0.00 (25/25)	0.00 (25/25)	0.00 (25/25)
homonyms	0.04 (25/25)	0.08 (25/25)	0.08 (25/25)	0.04 (25/25)
nons_group_real_discr	0.00 (25/25)	0.00 (25/25)	0.00 (25/25)	0.00 (25/25)
privacy_fictional	0.48 (25/25)	0.40 (25/25)	0.44 (25/25)	0.52 (25/25)
privacy_public	0.04 (25/25)	0.00 (25/25)	0.00 (25/25)	0.04 (25/25)
real_group_nons_discr	0.00 (25/25)	0.00 (25/25)	0.00 (25/25)	0.00 (25/25)
safe_contexts	0.16 (25/25)	0.20 (25/25)	0.16 (25/25)	0.16 (25/25)
safe_targets	0.04 (25/25)	0.08 (25/25)	0.00 (25/25)	0.04 (25/25)

Estimated answers at the 256-token cap: estimated / n. Estimate = the decoded answer text re-tokenizes to >= 256 tokens; this approximates, and cannot establish, the generated length or whether an answer was truncated. A truly capped answer was judged on truncated text.

set / language	original	plain-q4	random30-q4	protected-q4
harmbench_heldout/English	36 / 168	38 / 168	41 / 168	41 / 168
aya_redteam_hi_ar/Arabic	13 / 50	17 / 50	9 / 50	12 / 50
aya_redteam_hi_ar/Hindi	32 / 50	31 / 50	33 / 50	29 / 50
xstest_safe/English	101 / 250	95 / 250	90 / 250	106 / 250

Utility: Global-MMLU accuracy [95% question-bootstrap CI]; change vs original in percentage points (paired).

set	model	accuracy	Δ vs original (pp)
mmlu_en	original	+0.562 [+0.518, +0.606]
mmlu_en	plain-q4	+0.548 [+0.504, +0.590]	-1.4 [-4.4, +1.6]
mmlu_en	random30-q4	+0.540 [+0.496, +0.582]	-2.2 [-5.2, +0.8]
mmlu_en	protected-q4	+0.558 [+0.514, +0.602]	-0.4 [-3.0, +2.2]
mmlu_hi	original	+0.515 [+0.445, +0.585]
mmlu_hi	plain-q4	+0.495 [+0.425, +0.565]	-2.0 [-6.5, +2.5]
mmlu_hi	random30-q4	+0.515 [+0.445, +0.585]	+0.0 [-4.0, +4.0]
mmlu_hi	protected-q4	+0.535 [+0.465, +0.605]	+2.0 [-2.0, +6.0]