benchmark/tables/release_tables/baton/k4_input_effect.md

Stage K4in: causal vs oldin — paired device-cluster bootstrap (eval/paired_compare_joint.py, n_boot 1000), test split, both regimes

Both models are scored on the CAUSAL targets; A = causal-ROI inputs, B = old route-median inputs (input-isolation control) — the K4 input effect.

Seeds with rows: [np.int64(0), np.int64(1), np.int64(2)] (kill rule needs 3: ≥ 2/3 seeds CI excluding 0, same sign, |mean Δ| > seed sd). Source: results/paired_compare.csv tags K4in_s<k>_causal_vs_oldin. Cells: own endpoints (onset PR-AUC@3s), cross-leg endpoints (C1 / C2, PR-AUC + AUROC), hazard NLL, continuous emissions, nr / cooccur Brier. Driver-leg cells use the route-median (non-causal) ROI inputs (diagnostic, K4 pending).

Main comparison: A = causal, B = oldin

regime	cell	metric	A s0	A s1	A s2	Δ s0 [CI]	Δ s1 [CI]	Δ s2 [CI]	mean Δ ± sd	verdict
manual	drv_onset/hand_leave/3s	pr_auc	0.4238	0.4172	0.4385	+0.1695 [+0.1410,+0.2230]	+0.1637 [+0.1340,+0.2164]	+0.2303 [+0.1711,+0.2514]	+0.1878 ± 0.0369 (n=3, pos=36693)	BEATS
manual	drv_onset/hand_return/3s	pr_auc	0.7533	0.7294	0.7456	+0.3468 [+0.2355,+0.4473]	+0.2617 [+0.1592,+0.4003]	+0.3378 [+0.2450,+0.4583]	+0.3155 ± 0.0468 (n=3, pos=20868)	BEATS
manual	drv_onset/head_turn_onset/3s	pr_auc	0.4568	0.4720	0.4761	+0.1331 [+0.0905,+0.1472]	+0.1244 [+0.0734,+0.1423]	+0.1413 [+0.0643,+0.1672]	+0.1329 ± 0.0085 (n=3, pos=61049)	BEATS
manual	hazard_nll/hand_leave	nll (lower better)	0.1349	0.1378	0.1384	-0.0409 [-0.0797,-0.0299]	-0.0375 [-0.0740,-0.0271]	-0.0371 [-0.0727,-0.0271]	-0.0385 ± 0.0021 (n=3, pos=69578)	BEATS
manual	hazard_nll/hand_return	nll (lower better)	0.6260	0.6480	0.6368	-0.1583 [-0.2050,-0.0960]	-0.1340 [-0.1711,-0.0810]	-0.1035 [-0.1340,-0.0741]	-0.1319 ± 0.0274 (n=3, pos=68711)	BEATS
manual	hazard_nll/head_turn_onset	nll (lower better)	0.1724	0.1766	0.1732	-0.0342 [-0.0584,-0.0262]	-0.0228 [-0.0403,-0.0180]	-0.0189 [-0.0367,-0.0140]	-0.0253 ± 0.0079 (n=3, pos=114694)	BEATS
engaged	drv_onset/hand_leave/3s	pr_auc	0.3955	0.4018	0.4097	+0.2478 [+0.2029,+0.2867]	+0.2525 [+0.2015,+0.2926]	+0.2566 [+0.2056,+0.2969]	+0.2523 ± 0.0044 (n=3, pos=21986)	BEATS
engaged	drv_onset/hand_return/3s	pr_auc	0.6963	0.6928	0.6972	+0.3999 [+0.3139,+0.4578]	+0.3418 [+0.2553,+0.4146]	+0.3549 [+0.2668,+0.4453]	+0.3655 ± 0.0305 (n=3, pos=16138)	BEATS
engaged	drv_onset/head_turn_onset/3s	pr_auc	0.2987	0.3042	0.3088	+0.1452 [+0.1215,+0.1609]	+0.1312 [+0.1113,+0.1465]	+0.1194 [+0.0882,+0.1419]	+0.1319 ± 0.0129 (n=3, pos=47095)	BEATS
engaged	hazard_nll/hand_leave	nll (lower better)	0.1183	0.1186	0.1219	-0.0623 [-0.0936,-0.0390]	-0.0614 [-0.0929,-0.0381]	-0.0521 [-0.0850,-0.0290]	-0.0586 ± 0.0057 (n=3, pos=44787)	BEATS
engaged	hazard_nll/hand_return	nll (lower better)	0.5078	0.5021	0.5056	-0.1419 [-0.2080,-0.0998]	-0.1247 [-0.1861,-0.0857]	-0.0712 [-0.1026,-0.0504]	-0.1126 ± 0.0368 (n=3, pos=44490)	BEATS
engaged	hazard_nll/head_turn_onset	nll (lower better)	0.1542	0.1551	0.1535	-0.0321 [-0.0416,-0.0246]	-0.0282 [-0.0375,-0.0210]	-0.0222 [-0.0306,-0.0153]	-0.0275 ± 0.0050 (n=3, pos=85504)	BEATS

K4 (input causality) verdict: PASS

• own endpoints WORSE: none
• cross-leg endpoints BEATS: ['manual|drv_onset|hand_leave|3s|pr_auc', 'manual|drv_onset|hand_return|3s|pr_auc', 'manual|drv_onset|head_turn_onset|3s|pr_auc', 'engaged|drv_onset|hand_leave|3s|pr_auc', 'engaged|drv_onset|hand_return|3s|pr_auc', 'engaged|drv_onset|head_turn_onset|3s|pr_auc']
• own endpoints BEATS: none
• INCONCLUSIVE own endpoints (mean Δ): none