benchmark/tables/release_tables/baton/dms_coverage.md

DMS baselines — score coverage per (regime, split): finite rows, decoded routes, exposure hours (baselines_dms/eval_dms.py; append-only sections)

coverage — abl_dms_l2cs_gaze_learned — 2026-09-15T00:49:59Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tier A only (the 8 s windows before the 979 TRAIN label anchors, the clean gold anchors and the 299 newtest clips); rows outside those windows are undecoded by design, so the exposure coverage below is a lower bound until tiers B / C / D are extracted (the gold AUROC does not depend on exposure rows).

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_l2cs_gaze_learned	0	dms_l2cs_gaze_learned_sel_t15_w10_n979	manual	test	512,934	14,335	0.028	129 / 158	28.50	0.80
abl_dms_l2cs_gaze_learned	0	dms_l2cs_gaze_learned_sel_t15_w10_n979	manual	val	70,330	19,271	0.274	32 / 46	3.91	1.07
abl_dms_l2cs_gaze_learned	0	dms_l2cs_gaze_learned_sel_t15_w10_n979	engaged	test	451,840	10,645	0.024	69 / 101	25.10	0.59
abl_dms_l2cs_gaze_learned	0	dms_l2cs_gaze_learned_sel_t15_w10_n979	engaged	val	110,948	19,421	0.175	25 / 31	6.16	1.08
abl_dms_l2cs_gaze_learned	1	dms_l2cs_gaze_learned_sel_t15_w10_n979	manual	test	512,934	14,335	0.028	129 / 158	28.50	0.80
abl_dms_l2cs_gaze_learned	1	dms_l2cs_gaze_learned_sel_t15_w10_n979	manual	val	70,330	19,271	0.274	32 / 46	3.91	1.07
abl_dms_l2cs_gaze_learned	1	dms_l2cs_gaze_learned_sel_t15_w10_n979	engaged	test	451,840	10,645	0.024	69 / 101	25.10	0.59
abl_dms_l2cs_gaze_learned	1	dms_l2cs_gaze_learned_sel_t15_w10_n979	engaged	val	110,948	19,421	0.175	25 / 31	6.16	1.08
abl_dms_l2cs_gaze_learned	2	dms_l2cs_gaze_learned_sel_t15_w10_n979	manual	test	512,934	14,335	0.028	129 / 158	28.50	0.80
abl_dms_l2cs_gaze_learned	2	dms_l2cs_gaze_learned_sel_t15_w10_n979	manual	val	70,330	19,271	0.274	32 / 46	3.91	1.07
abl_dms_l2cs_gaze_learned	2	dms_l2cs_gaze_learned_sel_t15_w10_n979	engaged	test	451,840	10,645	0.024	69 / 101	25.10	0.59
abl_dms_l2cs_gaze_learned	2	dms_l2cs_gaze_learned_sel_t15_w10_n979	engaged	val	110,948	19,421	0.175	25 / 31	6.16	1.08

coverage — abl_dms_l2cs_gaze_ctx — 2026-09-15T00:52:56Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tier A only (the 8 s windows before the 979 TRAIN label anchors, the clean gold anchors and the 299 newtest clips); rows outside those windows are undecoded by design, so the exposure coverage below is a lower bound until tiers B / C / D are extracted (the gold AUROC does not depend on exposure rows).

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_l2cs_gaze_ctx	0	dms_l2cs_gaze_ctx_sel_t15_w10_n979	manual	test	512,934	14,335	0.028	129 / 158	28.50	0.80
abl_dms_l2cs_gaze_ctx	0	dms_l2cs_gaze_ctx_sel_t15_w10_n979	manual	val	70,330	19,271	0.274	32 / 46	3.91	1.07
abl_dms_l2cs_gaze_ctx	0	dms_l2cs_gaze_ctx_sel_t15_w10_n979	engaged	test	451,840	10,645	0.024	69 / 101	25.10	0.59
abl_dms_l2cs_gaze_ctx	0	dms_l2cs_gaze_ctx_sel_t15_w10_n979	engaged	val	110,948	19,421	0.175	25 / 31	6.16	1.08
abl_dms_l2cs_gaze_ctx	1	dms_l2cs_gaze_ctx_sel_t15_w10_n979	manual	test	512,934	14,335	0.028	129 / 158	28.50	0.80
abl_dms_l2cs_gaze_ctx	1	dms_l2cs_gaze_ctx_sel_t15_w10_n979	manual	val	70,330	19,271	0.274	32 / 46	3.91	1.07
abl_dms_l2cs_gaze_ctx	1	dms_l2cs_gaze_ctx_sel_t15_w10_n979	engaged	test	451,840	10,645	0.024	69 / 101	25.10	0.59
abl_dms_l2cs_gaze_ctx	1	dms_l2cs_gaze_ctx_sel_t15_w10_n979	engaged	val	110,948	19,421	0.175	25 / 31	6.16	1.08
abl_dms_l2cs_gaze_ctx	2	dms_l2cs_gaze_ctx_sel_t15_w10_n979	manual	test	512,934	14,335	0.028	129 / 158	28.50	0.80
abl_dms_l2cs_gaze_ctx	2	dms_l2cs_gaze_ctx_sel_t15_w10_n979	manual	val	70,330	19,271	0.274	32 / 46	3.91	1.07
abl_dms_l2cs_gaze_ctx	2	dms_l2cs_gaze_ctx_sel_t15_w10_n979	engaged	test	451,840	10,645	0.024	69 / 101	25.10	0.59
abl_dms_l2cs_gaze_ctx	2	dms_l2cs_gaze_ctx_sel_t15_w10_n979	engaged	val	110,948	19,421	0.175	25 / 31	6.16	1.08

coverage — abl_dms_videomae2_learned — 2026-09-15T01:04:22Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tier A only (the 8 s windows before the 979 TRAIN label anchors, the clean gold anchors and the 299 newtest clips); rows outside those windows are undecoded by design, so the exposure coverage below is a lower bound until tiers B / C / D are extracted (the gold AUROC does not depend on exposure rows).

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_videomae2_learned	0	dms_videomae2_learned_sel_t15_w10_n979	manual	test	512,934	20,204	0.039	130 / 158	28.50	1.12
abl_dms_videomae2_learned	0	dms_videomae2_learned_sel_t15_w10_n979	manual	val	70,330	19,921	0.283	33 / 46	3.91	1.11
abl_dms_videomae2_learned	0	dms_videomae2_learned_sel_t15_w10_n979	engaged	test	451,840	15,259	0.034	69 / 101	25.10	0.85
abl_dms_videomae2_learned	0	dms_videomae2_learned_sel_t15_w10_n979	engaged	val	110,948	20,022	0.180	25 / 31	6.16	1.11
abl_dms_videomae2_learned	1	dms_videomae2_learned_sel_t15_w10_n979	manual	test	512,934	20,204	0.039	130 / 158	28.50	1.12
abl_dms_videomae2_learned	1	dms_videomae2_learned_sel_t15_w10_n979	manual	val	70,330	19,921	0.283	33 / 46	3.91	1.11
abl_dms_videomae2_learned	1	dms_videomae2_learned_sel_t15_w10_n979	engaged	test	451,840	15,259	0.034	69 / 101	25.10	0.85
abl_dms_videomae2_learned	1	dms_videomae2_learned_sel_t15_w10_n979	engaged	val	110,948	20,022	0.180	25 / 31	6.16	1.11
abl_dms_videomae2_learned	2	dms_videomae2_learned_sel_t15_w10_n979	manual	test	512,934	20,204	0.039	130 / 158	28.50	1.12
abl_dms_videomae2_learned	2	dms_videomae2_learned_sel_t15_w10_n979	manual	val	70,330	19,921	0.283	33 / 46	3.91	1.11
abl_dms_videomae2_learned	2	dms_videomae2_learned_sel_t15_w10_n979	engaged	test	451,840	15,259	0.034	69 / 101	25.10	0.85
abl_dms_videomae2_learned	2	dms_videomae2_learned_sel_t15_w10_n979	engaged	val	110,948	20,022	0.180	25 / 31	6.16	1.11

coverage — abl_dms_videomae2_ctx — 2026-09-15T01:07:23Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tier A only (the 8 s windows before the 979 TRAIN label anchors, the clean gold anchors and the 299 newtest clips); rows outside those windows are undecoded by design, so the exposure coverage below is a lower bound until tiers B / C / D are extracted (the gold AUROC does not depend on exposure rows).

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_videomae2_ctx	0	dms_videomae2_ctx_sel_t15_w10_n979	manual	test	512,934	20,204	0.039	130 / 158	28.50	1.12
abl_dms_videomae2_ctx	0	dms_videomae2_ctx_sel_t15_w10_n979	manual	val	70,330	19,921	0.283	33 / 46	3.91	1.11
abl_dms_videomae2_ctx	0	dms_videomae2_ctx_sel_t15_w10_n979	engaged	test	451,840	15,259	0.034	69 / 101	25.10	0.85
abl_dms_videomae2_ctx	0	dms_videomae2_ctx_sel_t15_w10_n979	engaged	val	110,948	20,022	0.180	25 / 31	6.16	1.11
abl_dms_videomae2_ctx	1	dms_videomae2_ctx_sel_t15_w10_n979	manual	test	512,934	20,204	0.039	130 / 158	28.50	1.12
abl_dms_videomae2_ctx	1	dms_videomae2_ctx_sel_t15_w10_n979	manual	val	70,330	19,921	0.283	33 / 46	3.91	1.11
abl_dms_videomae2_ctx	1	dms_videomae2_ctx_sel_t15_w10_n979	engaged	test	451,840	15,259	0.034	69 / 101	25.10	0.85
abl_dms_videomae2_ctx	1	dms_videomae2_ctx_sel_t15_w10_n979	engaged	val	110,948	20,022	0.180	25 / 31	6.16	1.11
abl_dms_videomae2_ctx	2	dms_videomae2_ctx_sel_t15_w10_n979	manual	test	512,934	20,204	0.039	130 / 158	28.50	1.12
abl_dms_videomae2_ctx	2	dms_videomae2_ctx_sel_t15_w10_n979	manual	val	70,330	19,921	0.283	33 / 46	3.91	1.11
abl_dms_videomae2_ctx	2	dms_videomae2_ctx_sel_t15_w10_n979	engaged	test	451,840	15,259	0.034	69 / 101	25.10	0.85
abl_dms_videomae2_ctx	2	dms_videomae2_ctx_sel_t15_w10_n979	engaged	val	110,948	20,022	0.180	25 / 31	6.16	1.11

coverage — abl_dms_ov_gaze_learned — 2026-09-15T01:17:08Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tier A only (the 8 s windows before the 979 TRAIN label anchors, the clean gold anchors and the 299 newtest clips); rows outside those windows are undecoded by design, so the exposure coverage below is a lower bound until tiers B / C / D are extracted (the gold AUROC does not depend on exposure rows).

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_ov_gaze_learned	0	dms_ov_gaze_learned_sel_t15_w10_n979	manual	test	512,934	14,376	0.028	129 / 158	28.50	0.80
abl_dms_ov_gaze_learned	0	dms_ov_gaze_learned_sel_t15_w10_n979	manual	val	70,330	13,094	0.186	28 / 46	3.91	0.73
abl_dms_ov_gaze_learned	0	dms_ov_gaze_learned_sel_t15_w10_n979	engaged	test	451,840	10,651	0.024	69 / 101	25.10	0.59
abl_dms_ov_gaze_learned	0	dms_ov_gaze_learned_sel_t15_w10_n979	engaged	val	110,948	8,242	0.074	22 / 31	6.16	0.46
abl_dms_ov_gaze_learned	1	dms_ov_gaze_learned_sel_t15_w10_n979	manual	test	512,934	14,376	0.028	129 / 158	28.50	0.80
abl_dms_ov_gaze_learned	1	dms_ov_gaze_learned_sel_t15_w10_n979	manual	val	70,330	13,094	0.186	28 / 46	3.91	0.73
abl_dms_ov_gaze_learned	1	dms_ov_gaze_learned_sel_t15_w10_n979	engaged	test	451,840	10,651	0.024	69 / 101	25.10	0.59
abl_dms_ov_gaze_learned	1	dms_ov_gaze_learned_sel_t15_w10_n979	engaged	val	110,948	8,242	0.074	22 / 31	6.16	0.46
abl_dms_ov_gaze_learned	2	dms_ov_gaze_learned_sel_t15_w10_n979	manual	test	512,934	14,376	0.028	129 / 158	28.50	0.80
abl_dms_ov_gaze_learned	2	dms_ov_gaze_learned_sel_t15_w10_n979	manual	val	70,330	13,094	0.186	28 / 46	3.91	0.73
abl_dms_ov_gaze_learned	2	dms_ov_gaze_learned_sel_t15_w10_n979	engaged	test	451,840	10,651	0.024	69 / 101	25.10	0.59
abl_dms_ov_gaze_learned	2	dms_ov_gaze_learned_sel_t15_w10_n979	engaged	val	110,948	8,242	0.074	22 / 31	6.16	0.46

coverage — abl_dms_ov_gaze_ctx — 2026-09-15T01:20:06Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tier A only (the 8 s windows before the 979 TRAIN label anchors, the clean gold anchors and the 299 newtest clips); rows outside those windows are undecoded by design, so the exposure coverage below is a lower bound until tiers B / C / D are extracted (the gold AUROC does not depend on exposure rows).

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_ov_gaze_ctx	0	dms_ov_gaze_ctx_sel_t15_w10_n979	manual	test	512,934	14,376	0.028	129 / 158	28.50	0.80
abl_dms_ov_gaze_ctx	0	dms_ov_gaze_ctx_sel_t15_w10_n979	manual	val	70,330	13,094	0.186	28 / 46	3.91	0.73
abl_dms_ov_gaze_ctx	0	dms_ov_gaze_ctx_sel_t15_w10_n979	engaged	test	451,840	10,651	0.024	69 / 101	25.10	0.59
abl_dms_ov_gaze_ctx	0	dms_ov_gaze_ctx_sel_t15_w10_n979	engaged	val	110,948	8,242	0.074	22 / 31	6.16	0.46
abl_dms_ov_gaze_ctx	1	dms_ov_gaze_ctx_sel_t15_w10_n979	manual	test	512,934	14,376	0.028	129 / 158	28.50	0.80
abl_dms_ov_gaze_ctx	1	dms_ov_gaze_ctx_sel_t15_w10_n979	manual	val	70,330	13,094	0.186	28 / 46	3.91	0.73
abl_dms_ov_gaze_ctx	1	dms_ov_gaze_ctx_sel_t15_w10_n979	engaged	test	451,840	10,651	0.024	69 / 101	25.10	0.59
abl_dms_ov_gaze_ctx	1	dms_ov_gaze_ctx_sel_t15_w10_n979	engaged	val	110,948	8,242	0.074	22 / 31	6.16	0.46
abl_dms_ov_gaze_ctx	2	dms_ov_gaze_ctx_sel_t15_w10_n979	manual	test	512,934	14,376	0.028	129 / 158	28.50	0.80
abl_dms_ov_gaze_ctx	2	dms_ov_gaze_ctx_sel_t15_w10_n979	manual	val	70,330	13,094	0.186	28 / 46	3.91	0.73
abl_dms_ov_gaze_ctx	2	dms_ov_gaze_ctx_sel_t15_w10_n979	engaged	test	451,840	10,651	0.024	69 / 101	25.10	0.59
abl_dms_ov_gaze_ctx	2	dms_ov_gaze_ctx_sel_t15_w10_n979	engaged	val	110,948	8,242	0.074	22 / 31	6.16	0.46

coverage — abl_dms_mp_face_learned — 2026-09-15T01:29:49Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tier A only (the 8 s windows before the 979 TRAIN label anchors, the clean gold anchors and the 299 newtest clips); rows outside those windows are undecoded by design, so the exposure coverage below is a lower bound until tiers B / C / D are extracted (the gold AUROC does not depend on exposure rows).

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_mp_face_learned	0	dms_mp_face_learned_sel_t15_w10_n979	manual	test	512,934	13,539	0.026	127 / 158	28.50	0.75
abl_dms_mp_face_learned	0	dms_mp_face_learned_sel_t15_w10_n979	manual	val	70,330	16,130	0.229	32 / 46	3.91	0.90
abl_dms_mp_face_learned	0	dms_mp_face_learned_sel_t15_w10_n979	engaged	test	451,840	9,910	0.022	67 / 101	25.10	0.55
abl_dms_mp_face_learned	0	dms_mp_face_learned_sel_t15_w10_n979	engaged	val	110,948	10,313	0.093	25 / 31	6.16	0.57
abl_dms_mp_face_learned	1	dms_mp_face_learned_sel_t15_w10_n979	manual	test	512,934	13,539	0.026	127 / 158	28.50	0.75
abl_dms_mp_face_learned	1	dms_mp_face_learned_sel_t15_w10_n979	manual	val	70,330	16,130	0.229	32 / 46	3.91	0.90
abl_dms_mp_face_learned	1	dms_mp_face_learned_sel_t15_w10_n979	engaged	test	451,840	9,910	0.022	67 / 101	25.10	0.55
abl_dms_mp_face_learned	1	dms_mp_face_learned_sel_t15_w10_n979	engaged	val	110,948	10,313	0.093	25 / 31	6.16	0.57
abl_dms_mp_face_learned	2	dms_mp_face_learned_sel_t15_w10_n979	manual	test	512,934	13,539	0.026	127 / 158	28.50	0.75
abl_dms_mp_face_learned	2	dms_mp_face_learned_sel_t15_w10_n979	manual	val	70,330	16,130	0.229	32 / 46	3.91	0.90
abl_dms_mp_face_learned	2	dms_mp_face_learned_sel_t15_w10_n979	engaged	test	451,840	9,910	0.022	67 / 101	25.10	0.55
abl_dms_mp_face_learned	2	dms_mp_face_learned_sel_t15_w10_n979	engaged	val	110,948	10,313	0.093	25 / 31	6.16	0.57

coverage — abl_dms_mp_face_ctx — 2026-09-15T01:32:50Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tier A only (the 8 s windows before the 979 TRAIN label anchors, the clean gold anchors and the 299 newtest clips); rows outside those windows are undecoded by design, so the exposure coverage below is a lower bound until tiers B / C / D are extracted (the gold AUROC does not depend on exposure rows).

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_mp_face_ctx	0	dms_mp_face_ctx_sel_t15_w10_n979	manual	test	512,934	13,539	0.026	127 / 158	28.50	0.75
abl_dms_mp_face_ctx	0	dms_mp_face_ctx_sel_t15_w10_n979	manual	val	70,330	16,130	0.229	32 / 46	3.91	0.90
abl_dms_mp_face_ctx	0	dms_mp_face_ctx_sel_t15_w10_n979	engaged	test	451,840	9,910	0.022	67 / 101	25.10	0.55
abl_dms_mp_face_ctx	0	dms_mp_face_ctx_sel_t15_w10_n979	engaged	val	110,948	10,313	0.093	25 / 31	6.16	0.57
abl_dms_mp_face_ctx	1	dms_mp_face_ctx_sel_t15_w10_n979	manual	test	512,934	13,539	0.026	127 / 158	28.50	0.75
abl_dms_mp_face_ctx	1	dms_mp_face_ctx_sel_t15_w10_n979	manual	val	70,330	16,130	0.229	32 / 46	3.91	0.90
abl_dms_mp_face_ctx	1	dms_mp_face_ctx_sel_t15_w10_n979	engaged	test	451,840	9,910	0.022	67 / 101	25.10	0.55
abl_dms_mp_face_ctx	1	dms_mp_face_ctx_sel_t15_w10_n979	engaged	val	110,948	10,313	0.093	25 / 31	6.16	0.57
abl_dms_mp_face_ctx	2	dms_mp_face_ctx_sel_t15_w10_n979	manual	test	512,934	13,539	0.026	127 / 158	28.50	0.75
abl_dms_mp_face_ctx	2	dms_mp_face_ctx_sel_t15_w10_n979	manual	val	70,330	16,130	0.229	32 / 46	3.91	0.90
abl_dms_mp_face_ctx	2	dms_mp_face_ctx_sel_t15_w10_n979	engaged	test	451,840	9,910	0.022	67 / 101	25.10	0.55
abl_dms_mp_face_ctx	2	dms_mp_face_ctx_sel_t15_w10_n979	engaged	val	110,948	10,313	0.093	25 / 31	6.16	0.57

coverage — abl_dms_ov_action_learned — 2026-09-15T01:45:33Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tier A only (the 8 s windows before the 979 TRAIN label anchors, the clean gold anchors and the 299 newtest clips); rows outside those windows are undecoded by design, so the exposure coverage below is a lower bound until tiers B / C / D are extracted (the gold AUROC does not depend on exposure rows).

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_ov_action_learned	0	dms_ov_action_learned_sel_t15_w10_n979	manual	test	512,934	17,562	0.034	129 / 158	28.50	0.98
abl_dms_ov_action_learned	0	dms_ov_action_learned_sel_t15_w10_n979	manual	val	70,330	19,684	0.280	33 / 46	3.91	1.09
abl_dms_ov_action_learned	0	dms_ov_action_learned_sel_t15_w10_n979	engaged	test	451,840	13,175	0.029	69 / 101	25.10	0.73
abl_dms_ov_action_learned	0	dms_ov_action_learned_sel_t15_w10_n979	engaged	val	110,948	19,752	0.178	25 / 31	6.16	1.10
abl_dms_ov_action_learned	1	dms_ov_action_learned_sel_t15_w10_n979	manual	test	512,934	17,562	0.034	129 / 158	28.50	0.98
abl_dms_ov_action_learned	1	dms_ov_action_learned_sel_t15_w10_n979	manual	val	70,330	19,684	0.280	33 / 46	3.91	1.09
abl_dms_ov_action_learned	1	dms_ov_action_learned_sel_t15_w10_n979	engaged	test	451,840	13,175	0.029	69 / 101	25.10	0.73
abl_dms_ov_action_learned	1	dms_ov_action_learned_sel_t15_w10_n979	engaged	val	110,948	19,752	0.178	25 / 31	6.16	1.10
abl_dms_ov_action_learned	2	dms_ov_action_learned_sel_t15_w10_n979	manual	test	512,934	17,562	0.034	129 / 158	28.50	0.98
abl_dms_ov_action_learned	2	dms_ov_action_learned_sel_t15_w10_n979	manual	val	70,330	19,684	0.280	33 / 46	3.91	1.09
abl_dms_ov_action_learned	2	dms_ov_action_learned_sel_t15_w10_n979	engaged	test	451,840	13,175	0.029	69 / 101	25.10	0.73
abl_dms_ov_action_learned	2	dms_ov_action_learned_sel_t15_w10_n979	engaged	val	110,948	19,752	0.178	25 / 31	6.16	1.10

coverage — abl_dms_ov_action_ctx — 2026-09-15T01:48:27Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tier A only (the 8 s windows before the 979 TRAIN label anchors, the clean gold anchors and the 299 newtest clips); rows outside those windows are undecoded by design, so the exposure coverage below is a lower bound until tiers B / C / D are extracted (the gold AUROC does not depend on exposure rows).

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_ov_action_ctx	0	dms_ov_action_ctx_sel_t15_w10_n979	manual	test	512,934	17,562	0.034	129 / 158	28.50	0.98
abl_dms_ov_action_ctx	0	dms_ov_action_ctx_sel_t15_w10_n979	manual	val	70,330	19,684	0.280	33 / 46	3.91	1.09
abl_dms_ov_action_ctx	0	dms_ov_action_ctx_sel_t15_w10_n979	engaged	test	451,840	13,175	0.029	69 / 101	25.10	0.73
abl_dms_ov_action_ctx	0	dms_ov_action_ctx_sel_t15_w10_n979	engaged	val	110,948	19,752	0.178	25 / 31	6.16	1.10
abl_dms_ov_action_ctx	1	dms_ov_action_ctx_sel_t15_w10_n979	manual	test	512,934	17,562	0.034	129 / 158	28.50	0.98
abl_dms_ov_action_ctx	1	dms_ov_action_ctx_sel_t15_w10_n979	manual	val	70,330	19,684	0.280	33 / 46	3.91	1.09
abl_dms_ov_action_ctx	1	dms_ov_action_ctx_sel_t15_w10_n979	engaged	test	451,840	13,175	0.029	69 / 101	25.10	0.73
abl_dms_ov_action_ctx	1	dms_ov_action_ctx_sel_t15_w10_n979	engaged	val	110,948	19,752	0.178	25 / 31	6.16	1.10
abl_dms_ov_action_ctx	2	dms_ov_action_ctx_sel_t15_w10_n979	manual	test	512,934	17,562	0.034	129 / 158	28.50	0.98
abl_dms_ov_action_ctx	2	dms_ov_action_ctx_sel_t15_w10_n979	manual	val	70,330	19,684	0.280	33 / 46	3.91	1.09
abl_dms_ov_action_ctx	2	dms_ov_action_ctx_sel_t15_w10_n979	engaged	test	451,840	13,175	0.029	69 / 101	25.10	0.73
abl_dms_ov_action_ctx	2	dms_ov_action_ctx_sel_t15_w10_n979	engaged	val	110,948	19,752	0.178	25 / 31	6.16	1.10

coverage — abl_dms_openface3_learned — 2026-09-15T01:58:23Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tier A only (the 8 s windows before the 979 TRAIN label anchors, the clean gold anchors and the 299 newtest clips); rows outside those windows are undecoded by design, so the exposure coverage below is a lower bound until tiers B / C / D are extracted (the gold AUROC does not depend on exposure rows).

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_openface3_learned	0	dms_openface3_learned_sel_t15_w10_n979	manual	test	512,934	14,335	0.028	129 / 158	28.50	0.80
abl_dms_openface3_learned	0	dms_openface3_learned_sel_t15_w10_n979	manual	val	70,330	66,146	0.941	45 / 46	3.91	3.67
abl_dms_openface3_learned	0	dms_openface3_learned_sel_t15_w10_n979	engaged	test	451,840	10,645	0.024	69 / 101	25.10	0.59
abl_dms_openface3_learned	0	dms_openface3_learned_sel_t15_w10_n979	engaged	val	110,948	86,208	0.777	31 / 31	6.16	4.79
abl_dms_openface3_learned	1	dms_openface3_learned_sel_t15_w10_n979	manual	test	512,934	14,335	0.028	129 / 158	28.50	0.80
abl_dms_openface3_learned	1	dms_openface3_learned_sel_t15_w10_n979	manual	val	70,330	66,146	0.941	45 / 46	3.91	3.67
abl_dms_openface3_learned	1	dms_openface3_learned_sel_t15_w10_n979	engaged	test	451,840	10,645	0.024	69 / 101	25.10	0.59
abl_dms_openface3_learned	1	dms_openface3_learned_sel_t15_w10_n979	engaged	val	110,948	86,208	0.777	31 / 31	6.16	4.79
abl_dms_openface3_learned	2	dms_openface3_learned_sel_t15_w10_n979	manual	test	512,934	14,335	0.028	129 / 158	28.50	0.80
abl_dms_openface3_learned	2	dms_openface3_learned_sel_t15_w10_n979	manual	val	70,330	66,146	0.941	45 / 46	3.91	3.67
abl_dms_openface3_learned	2	dms_openface3_learned_sel_t15_w10_n979	engaged	test	451,840	10,645	0.024	69 / 101	25.10	0.59
abl_dms_openface3_learned	2	dms_openface3_learned_sel_t15_w10_n979	engaged	val	110,948	86,208	0.777	31 / 31	6.16	4.79

coverage — abl_dms_openface3_ctx — 2026-09-15T02:01:47Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tier A only (the 8 s windows before the 979 TRAIN label anchors, the clean gold anchors and the 299 newtest clips); rows outside those windows are undecoded by design, so the exposure coverage below is a lower bound until tiers B / C / D are extracted (the gold AUROC does not depend on exposure rows).

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_openface3_ctx	0	dms_openface3_ctx_sel_t15_w10_n979	manual	test	512,934	14,335	0.028	129 / 158	28.50	0.80
abl_dms_openface3_ctx	0	dms_openface3_ctx_sel_t15_w10_n979	manual	val	70,330	66,146	0.941	45 / 46	3.91	3.67
abl_dms_openface3_ctx	0	dms_openface3_ctx_sel_t15_w10_n979	engaged	test	451,840	10,645	0.024	69 / 101	25.10	0.59
abl_dms_openface3_ctx	0	dms_openface3_ctx_sel_t15_w10_n979	engaged	val	110,948	86,208	0.777	31 / 31	6.16	4.79
abl_dms_openface3_ctx	1	dms_openface3_ctx_sel_t15_w10_n979	manual	test	512,934	14,335	0.028	129 / 158	28.50	0.80
abl_dms_openface3_ctx	1	dms_openface3_ctx_sel_t15_w10_n979	manual	val	70,330	66,146	0.941	45 / 46	3.91	3.67
abl_dms_openface3_ctx	1	dms_openface3_ctx_sel_t15_w10_n979	engaged	test	451,840	10,645	0.024	69 / 101	25.10	0.59
abl_dms_openface3_ctx	1	dms_openface3_ctx_sel_t15_w10_n979	engaged	val	110,948	86,208	0.777	31 / 31	6.16	4.79
abl_dms_openface3_ctx	2	dms_openface3_ctx_sel_t15_w10_n979	manual	test	512,934	14,335	0.028	129 / 158	28.50	0.80
abl_dms_openface3_ctx	2	dms_openface3_ctx_sel_t15_w10_n979	manual	val	70,330	66,146	0.941	45 / 46	3.91	3.67
abl_dms_openface3_ctx	2	dms_openface3_ctx_sel_t15_w10_n979	engaged	test	451,840	10,645	0.024	69 / 101	25.10	0.59
abl_dms_openface3_ctx	2	dms_openface3_ctx_sel_t15_w10_n979	engaged	val	110,948	86,208	0.777	31 / 31	6.16	4.79

coverage — abl_dms_videomae2_learned_full — 2026-09-16T04:38:53Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tiers A + B + C (anchor windows + the full val and test exposure; tier D = train exposure is not needed for operating points); layouts = the tier-A read-outs applied only (fit_readout.py --apply-only, no refit) to the full-exposure summaries results/baselines_dms/summ_full//.

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_videomae2_learned_full	0	dms_videomae2_learned_sel_t15_w10_n979	manual	test	512,934	512,766	1.000	158 / 158	28.50	28.49
abl_dms_videomae2_learned_full	0	dms_videomae2_learned_sel_t15_w10_n979	manual	val	70,330	69,423	0.987	46 / 46	3.91	3.86
abl_dms_videomae2_learned_full	0	dms_videomae2_learned_sel_t15_w10_n979	engaged	test	451,840	451,834	1.000	101 / 101	25.10	25.10
abl_dms_videomae2_learned_full	0	dms_videomae2_learned_sel_t15_w10_n979	engaged	val	110,948	97,018	0.874	31 / 31	6.16	5.39
abl_dms_videomae2_learned_full	1	dms_videomae2_learned_sel_t15_w10_n979	manual	test	512,934	512,766	1.000	158 / 158	28.50	28.49
abl_dms_videomae2_learned_full	1	dms_videomae2_learned_sel_t15_w10_n979	manual	val	70,330	69,423	0.987	46 / 46	3.91	3.86
abl_dms_videomae2_learned_full	1	dms_videomae2_learned_sel_t15_w10_n979	engaged	test	451,840	451,834	1.000	101 / 101	25.10	25.10
abl_dms_videomae2_learned_full	1	dms_videomae2_learned_sel_t15_w10_n979	engaged	val	110,948	97,018	0.874	31 / 31	6.16	5.39
abl_dms_videomae2_learned_full	2	dms_videomae2_learned_sel_t15_w10_n979	manual	test	512,934	512,766	1.000	158 / 158	28.50	28.49
abl_dms_videomae2_learned_full	2	dms_videomae2_learned_sel_t15_w10_n979	manual	val	70,330	69,423	0.987	46 / 46	3.91	3.86
abl_dms_videomae2_learned_full	2	dms_videomae2_learned_sel_t15_w10_n979	engaged	test	451,840	451,834	1.000	101 / 101	25.10	25.10
abl_dms_videomae2_learned_full	2	dms_videomae2_learned_sel_t15_w10_n979	engaged	val	110,948	97,018	0.874	31 / 31	6.16	5.39

coverage — abl_dms_videomae2_ctx_full — 2026-09-16T05:01:08Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tiers A + B + C (anchor windows + the full val and test exposure; tier D = train exposure is not needed for operating points); layouts = the tier-A read-outs applied only (fit_readout.py --apply-only, no refit) to the full-exposure summaries results/baselines_dms/summ_full//.

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_videomae2_ctx_full	0	dms_videomae2_ctx_sel_t15_w10_n979	manual	test	512,934	512,766	1.000	158 / 158	28.50	28.49
abl_dms_videomae2_ctx_full	0	dms_videomae2_ctx_sel_t15_w10_n979	manual	val	70,330	69,423	0.987	46 / 46	3.91	3.86
abl_dms_videomae2_ctx_full	0	dms_videomae2_ctx_sel_t15_w10_n979	engaged	test	451,840	451,834	1.000	101 / 101	25.10	25.10
abl_dms_videomae2_ctx_full	0	dms_videomae2_ctx_sel_t15_w10_n979	engaged	val	110,948	97,018	0.874	31 / 31	6.16	5.39
abl_dms_videomae2_ctx_full	1	dms_videomae2_ctx_sel_t15_w10_n979	manual	test	512,934	512,766	1.000	158 / 158	28.50	28.49
abl_dms_videomae2_ctx_full	1	dms_videomae2_ctx_sel_t15_w10_n979	manual	val	70,330	69,423	0.987	46 / 46	3.91	3.86
abl_dms_videomae2_ctx_full	1	dms_videomae2_ctx_sel_t15_w10_n979	engaged	test	451,840	451,834	1.000	101 / 101	25.10	25.10
abl_dms_videomae2_ctx_full	1	dms_videomae2_ctx_sel_t15_w10_n979	engaged	val	110,948	97,018	0.874	31 / 31	6.16	5.39
abl_dms_videomae2_ctx_full	2	dms_videomae2_ctx_sel_t15_w10_n979	manual	test	512,934	512,766	1.000	158 / 158	28.50	28.49
abl_dms_videomae2_ctx_full	2	dms_videomae2_ctx_sel_t15_w10_n979	manual	val	70,330	69,423	0.987	46 / 46	3.91	3.86
abl_dms_videomae2_ctx_full	2	dms_videomae2_ctx_sel_t15_w10_n979	engaged	test	451,840	451,834	1.000	101 / 101	25.10	25.10
abl_dms_videomae2_ctx_full	2	dms_videomae2_ctx_sel_t15_w10_n979	engaged	val	110,948	97,018	0.874	31 / 31	6.16	5.39

coverage — abl_dms_l2cs_gaze_learned_full — 2026-09-16T05:14:34Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tiers A + B + C (anchor windows + the full val and test exposure; tier D = train exposure is not needed for operating points); layouts = the tier-A read-outs applied only (fit_readout.py --apply-only, no refit) to the full-exposure summaries results/baselines_dms/summ_full//.

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_l2cs_gaze_learned_full	0	dms_l2cs_gaze_learned_sel_t15_w10_n979	manual	test	512,934	512,091	0.998	158 / 158	28.50	28.45
abl_dms_l2cs_gaze_learned_full	0	dms_l2cs_gaze_learned_sel_t15_w10_n979	manual	val	70,330	67,387	0.958	46 / 46	3.91	3.74
abl_dms_l2cs_gaze_learned_full	0	dms_l2cs_gaze_learned_sel_t15_w10_n979	engaged	test	451,840	451,220	0.999	101 / 101	25.10	25.07
abl_dms_l2cs_gaze_learned_full	0	dms_l2cs_gaze_learned_sel_t15_w10_n979	engaged	val	110,948	96,699	0.872	31 / 31	6.16	5.37
abl_dms_l2cs_gaze_learned_full	1	dms_l2cs_gaze_learned_sel_t15_w10_n979	manual	test	512,934	512,091	0.998	158 / 158	28.50	28.45
abl_dms_l2cs_gaze_learned_full	1	dms_l2cs_gaze_learned_sel_t15_w10_n979	manual	val	70,330	67,387	0.958	46 / 46	3.91	3.74
abl_dms_l2cs_gaze_learned_full	1	dms_l2cs_gaze_learned_sel_t15_w10_n979	engaged	test	451,840	451,220	0.999	101 / 101	25.10	25.07
abl_dms_l2cs_gaze_learned_full	1	dms_l2cs_gaze_learned_sel_t15_w10_n979	engaged	val	110,948	96,699	0.872	31 / 31	6.16	5.37
abl_dms_l2cs_gaze_learned_full	2	dms_l2cs_gaze_learned_sel_t15_w10_n979	manual	test	512,934	512,091	0.998	158 / 158	28.50	28.45
abl_dms_l2cs_gaze_learned_full	2	dms_l2cs_gaze_learned_sel_t15_w10_n979	manual	val	70,330	67,387	0.958	46 / 46	3.91	3.74
abl_dms_l2cs_gaze_learned_full	2	dms_l2cs_gaze_learned_sel_t15_w10_n979	engaged	test	451,840	451,220	0.999	101 / 101	25.10	25.07
abl_dms_l2cs_gaze_learned_full	2	dms_l2cs_gaze_learned_sel_t15_w10_n979	engaged	val	110,948	96,699	0.872	31 / 31	6.16	5.37

coverage — abl_dms_l2cs_gaze_ctx_full — 2026-09-16T05:29:55Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tiers A + B + C (anchor windows + the full val and test exposure; tier D = train exposure is not needed for operating points); layouts = the tier-A read-outs applied only (fit_readout.py --apply-only, no refit) to the full-exposure summaries results/baselines_dms/summ_full//.

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_l2cs_gaze_ctx_full	0	dms_l2cs_gaze_ctx_sel_t15_w10_n979	manual	test	512,934	512,091	0.998	158 / 158	28.50	28.45
abl_dms_l2cs_gaze_ctx_full	0	dms_l2cs_gaze_ctx_sel_t15_w10_n979	manual	val	70,330	67,387	0.958	46 / 46	3.91	3.74
abl_dms_l2cs_gaze_ctx_full	0	dms_l2cs_gaze_ctx_sel_t15_w10_n979	engaged	test	451,840	451,220	0.999	101 / 101	25.10	25.07
abl_dms_l2cs_gaze_ctx_full	0	dms_l2cs_gaze_ctx_sel_t15_w10_n979	engaged	val	110,948	96,699	0.872	31 / 31	6.16	5.37
abl_dms_l2cs_gaze_ctx_full	1	dms_l2cs_gaze_ctx_sel_t15_w10_n979	manual	test	512,934	512,091	0.998	158 / 158	28.50	28.45
abl_dms_l2cs_gaze_ctx_full	1	dms_l2cs_gaze_ctx_sel_t15_w10_n979	manual	val	70,330	67,387	0.958	46 / 46	3.91	3.74
abl_dms_l2cs_gaze_ctx_full	1	dms_l2cs_gaze_ctx_sel_t15_w10_n979	engaged	test	451,840	451,220	0.999	101 / 101	25.10	25.07
abl_dms_l2cs_gaze_ctx_full	1	dms_l2cs_gaze_ctx_sel_t15_w10_n979	engaged	val	110,948	96,699	0.872	31 / 31	6.16	5.37
abl_dms_l2cs_gaze_ctx_full	2	dms_l2cs_gaze_ctx_sel_t15_w10_n979	manual	test	512,934	512,091	0.998	158 / 158	28.50	28.45
abl_dms_l2cs_gaze_ctx_full	2	dms_l2cs_gaze_ctx_sel_t15_w10_n979	manual	val	70,330	67,387	0.958	46 / 46	3.91	3.74
abl_dms_l2cs_gaze_ctx_full	2	dms_l2cs_gaze_ctx_sel_t15_w10_n979	engaged	test	451,840	451,220	0.999	101 / 101	25.10	25.07
abl_dms_l2cs_gaze_ctx_full	2	dms_l2cs_gaze_ctx_sel_t15_w10_n979	engaged	val	110,948	96,699	0.872	31 / 31	6.16	5.37

coverage — abl_dms_openface3_learned_full — 2026-09-16T05:50:21Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tiers A + B + C (anchor windows + the full val and test exposure; tier D = train exposure is not needed for operating points); layouts = the tier-A read-outs applied only (fit_readout.py --apply-only, no refit) to the full-exposure summaries results/baselines_dms/summ_full//.

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_openface3_learned_full	0	dms_openface3_learned_sel_t15_w10_n979	manual	test	512,934	512,091	0.998	158 / 158	28.50	28.45
abl_dms_openface3_learned_full	0	dms_openface3_learned_sel_t15_w10_n979	manual	val	70,330	67,387	0.958	46 / 46	3.91	3.74
abl_dms_openface3_learned_full	0	dms_openface3_learned_sel_t15_w10_n979	engaged	test	451,840	451,220	0.999	101 / 101	25.10	25.07
abl_dms_openface3_learned_full	0	dms_openface3_learned_sel_t15_w10_n979	engaged	val	110,948	96,699	0.872	31 / 31	6.16	5.37
abl_dms_openface3_learned_full	1	dms_openface3_learned_sel_t15_w10_n979	manual	test	512,934	512,091	0.998	158 / 158	28.50	28.45
abl_dms_openface3_learned_full	1	dms_openface3_learned_sel_t15_w10_n979	manual	val	70,330	67,387	0.958	46 / 46	3.91	3.74
abl_dms_openface3_learned_full	1	dms_openface3_learned_sel_t15_w10_n979	engaged	test	451,840	451,220	0.999	101 / 101	25.10	25.07
abl_dms_openface3_learned_full	1	dms_openface3_learned_sel_t15_w10_n979	engaged	val	110,948	96,699	0.872	31 / 31	6.16	5.37
abl_dms_openface3_learned_full	2	dms_openface3_learned_sel_t15_w10_n979	manual	test	512,934	512,091	0.998	158 / 158	28.50	28.45
abl_dms_openface3_learned_full	2	dms_openface3_learned_sel_t15_w10_n979	manual	val	70,330	67,387	0.958	46 / 46	3.91	3.74
abl_dms_openface3_learned_full	2	dms_openface3_learned_sel_t15_w10_n979	engaged	test	451,840	451,220	0.999	101 / 101	25.10	25.07
abl_dms_openface3_learned_full	2	dms_openface3_learned_sel_t15_w10_n979	engaged	val	110,948	96,699	0.872	31 / 31	6.16	5.37

coverage — abl_dms_openface3_ctx_full — 2026-09-16T06:03:41Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tiers A + B + C (anchor windows + the full val and test exposure; tier D = train exposure is not needed for operating points); layouts = the tier-A read-outs applied only (fit_readout.py --apply-only, no refit) to the full-exposure summaries results/baselines_dms/summ_full//.

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_openface3_ctx_full	0	dms_openface3_ctx_sel_t15_w10_n979	manual	test	512,934	512,091	0.998	158 / 158	28.50	28.45
abl_dms_openface3_ctx_full	0	dms_openface3_ctx_sel_t15_w10_n979	manual	val	70,330	67,387	0.958	46 / 46	3.91	3.74
abl_dms_openface3_ctx_full	0	dms_openface3_ctx_sel_t15_w10_n979	engaged	test	451,840	451,220	0.999	101 / 101	25.10	25.07
abl_dms_openface3_ctx_full	0	dms_openface3_ctx_sel_t15_w10_n979	engaged	val	110,948	96,699	0.872	31 / 31	6.16	5.37
abl_dms_openface3_ctx_full	1	dms_openface3_ctx_sel_t15_w10_n979	manual	test	512,934	512,091	0.998	158 / 158	28.50	28.45
abl_dms_openface3_ctx_full	1	dms_openface3_ctx_sel_t15_w10_n979	manual	val	70,330	67,387	0.958	46 / 46	3.91	3.74
abl_dms_openface3_ctx_full	1	dms_openface3_ctx_sel_t15_w10_n979	engaged	test	451,840	451,220	0.999	101 / 101	25.10	25.07
abl_dms_openface3_ctx_full	1	dms_openface3_ctx_sel_t15_w10_n979	engaged	val	110,948	96,699	0.872	31 / 31	6.16	5.37
abl_dms_openface3_ctx_full	2	dms_openface3_ctx_sel_t15_w10_n979	manual	test	512,934	512,091	0.998	158 / 158	28.50	28.45
abl_dms_openface3_ctx_full	2	dms_openface3_ctx_sel_t15_w10_n979	manual	val	70,330	67,387	0.958	46 / 46	3.91	3.74
abl_dms_openface3_ctx_full	2	dms_openface3_ctx_sel_t15_w10_n979	engaged	test	451,840	451,220	0.999	101 / 101	25.10	25.07
abl_dms_openface3_ctx_full	2	dms_openface3_ctx_sel_t15_w10_n979	engaged	val	110,948	96,699	0.872	31 / 31	6.16	5.37

coverage — abl_dms_ov_gaze_learned_full — 2026-09-16T06:18:03Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tiers A + B + C (anchor windows + the full val and test exposure; tier D = train exposure is not needed for operating points); layouts = the tier-A read-outs applied only (fit_readout.py --apply-only, no refit) to the full-exposure summaries results/baselines_dms/summ_full//.

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_ov_gaze_learned_full	0	dms_ov_gaze_learned_sel_t15_w10_n979	manual	test	512,934	512,200	0.999	158 / 158	28.50	28.46
abl_dms_ov_gaze_learned_full	0	dms_ov_gaze_learned_sel_t15_w10_n979	manual	val	70,330	65,373	0.930	46 / 46	3.91	3.63
abl_dms_ov_gaze_learned_full	0	dms_ov_gaze_learned_sel_t15_w10_n979	engaged	test	451,840	451,437	0.999	101 / 101	25.10	25.08
abl_dms_ov_gaze_learned_full	0	dms_ov_gaze_learned_sel_t15_w10_n979	engaged	val	110,948	96,792	0.872	31 / 31	6.16	5.38
abl_dms_ov_gaze_learned_full	1	dms_ov_gaze_learned_sel_t15_w10_n979	manual	test	512,934	512,200	0.999	158 / 158	28.50	28.46
abl_dms_ov_gaze_learned_full	1	dms_ov_gaze_learned_sel_t15_w10_n979	manual	val	70,330	65,373	0.930	46 / 46	3.91	3.63
abl_dms_ov_gaze_learned_full	1	dms_ov_gaze_learned_sel_t15_w10_n979	engaged	test	451,840	451,437	0.999	101 / 101	25.10	25.08
abl_dms_ov_gaze_learned_full	1	dms_ov_gaze_learned_sel_t15_w10_n979	engaged	val	110,948	96,792	0.872	31 / 31	6.16	5.38
abl_dms_ov_gaze_learned_full	2	dms_ov_gaze_learned_sel_t15_w10_n979	manual	test	512,934	512,200	0.999	158 / 158	28.50	28.46
abl_dms_ov_gaze_learned_full	2	dms_ov_gaze_learned_sel_t15_w10_n979	manual	val	70,330	65,373	0.930	46 / 46	3.91	3.63
abl_dms_ov_gaze_learned_full	2	dms_ov_gaze_learned_sel_t15_w10_n979	engaged	test	451,840	451,437	0.999	101 / 101	25.10	25.08
abl_dms_ov_gaze_learned_full	2	dms_ov_gaze_learned_sel_t15_w10_n979	engaged	val	110,948	96,792	0.872	31 / 31	6.16	5.38

coverage — abl_dms_ov_gaze_ctx_full — 2026-09-16T06:31:13Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tiers A + B + C (anchor windows + the full val and test exposure; tier D = train exposure is not needed for operating points); layouts = the tier-A read-outs applied only (fit_readout.py --apply-only, no refit) to the full-exposure summaries results/baselines_dms/summ_full//.

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_ov_gaze_ctx_full	0	dms_ov_gaze_ctx_sel_t15_w10_n979	manual	test	512,934	512,200	0.999	158 / 158	28.50	28.46
abl_dms_ov_gaze_ctx_full	0	dms_ov_gaze_ctx_sel_t15_w10_n979	manual	val	70,330	65,373	0.930	46 / 46	3.91	3.63
abl_dms_ov_gaze_ctx_full	0	dms_ov_gaze_ctx_sel_t15_w10_n979	engaged	test	451,840	451,437	0.999	101 / 101	25.10	25.08
abl_dms_ov_gaze_ctx_full	0	dms_ov_gaze_ctx_sel_t15_w10_n979	engaged	val	110,948	96,792	0.872	31 / 31	6.16	5.38
abl_dms_ov_gaze_ctx_full	1	dms_ov_gaze_ctx_sel_t15_w10_n979	manual	test	512,934	512,200	0.999	158 / 158	28.50	28.46
abl_dms_ov_gaze_ctx_full	1	dms_ov_gaze_ctx_sel_t15_w10_n979	manual	val	70,330	65,373	0.930	46 / 46	3.91	3.63
abl_dms_ov_gaze_ctx_full	1	dms_ov_gaze_ctx_sel_t15_w10_n979	engaged	test	451,840	451,437	0.999	101 / 101	25.10	25.08
abl_dms_ov_gaze_ctx_full	1	dms_ov_gaze_ctx_sel_t15_w10_n979	engaged	val	110,948	96,792	0.872	31 / 31	6.16	5.38
abl_dms_ov_gaze_ctx_full	2	dms_ov_gaze_ctx_sel_t15_w10_n979	manual	test	512,934	512,200	0.999	158 / 158	28.50	28.46
abl_dms_ov_gaze_ctx_full	2	dms_ov_gaze_ctx_sel_t15_w10_n979	manual	val	70,330	65,373	0.930	46 / 46	3.91	3.63
abl_dms_ov_gaze_ctx_full	2	dms_ov_gaze_ctx_sel_t15_w10_n979	engaged	test	451,840	451,437	0.999	101 / 101	25.10	25.08
abl_dms_ov_gaze_ctx_full	2	dms_ov_gaze_ctx_sel_t15_w10_n979	engaged	val	110,948	96,792	0.872	31 / 31	6.16	5.38

coverage — abl_dms_mp_face_ctx_full — 2026-09-16T06:51:18Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tiers A + B + C (anchor windows + the full val and test exposure; tier D = train exposure is not needed for operating points); layouts = the tier-A read-outs applied only (fit_readout.py --apply-only, no refit) to the full-exposure summaries results/baselines_dms/summ_full//.

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_mp_face_ctx_full	0	dms_mp_face_ctx_sel_t15_w10_n979	manual	test	512,934	502,131	0.979	157 / 158	28.50	27.90
abl_dms_mp_face_ctx_full	0	dms_mp_face_ctx_sel_t15_w10_n979	manual	val	70,330	62,783	0.893	46 / 46	3.91	3.49
abl_dms_mp_face_ctx_full	0	dms_mp_face_ctx_sel_t15_w10_n979	engaged	test	451,840	436,274	0.966	100 / 101	25.10	24.24
abl_dms_mp_face_ctx_full	0	dms_mp_face_ctx_sel_t15_w10_n979	engaged	val	110,948	96,170	0.867	31 / 31	6.16	5.34
abl_dms_mp_face_ctx_full	1	dms_mp_face_ctx_sel_t15_w10_n979	manual	test	512,934	502,131	0.979	157 / 158	28.50	27.90
abl_dms_mp_face_ctx_full	1	dms_mp_face_ctx_sel_t15_w10_n979	manual	val	70,330	62,783	0.893	46 / 46	3.91	3.49
abl_dms_mp_face_ctx_full	1	dms_mp_face_ctx_sel_t15_w10_n979	engaged	test	451,840	436,274	0.966	100 / 101	25.10	24.24
abl_dms_mp_face_ctx_full	1	dms_mp_face_ctx_sel_t15_w10_n979	engaged	val	110,948	96,170	0.867	31 / 31	6.16	5.34
abl_dms_mp_face_ctx_full	2	dms_mp_face_ctx_sel_t15_w10_n979	manual	test	512,934	502,131	0.979	157 / 158	28.50	27.90
abl_dms_mp_face_ctx_full	2	dms_mp_face_ctx_sel_t15_w10_n979	manual	val	70,330	62,783	0.893	46 / 46	3.91	3.49
abl_dms_mp_face_ctx_full	2	dms_mp_face_ctx_sel_t15_w10_n979	engaged	test	451,840	436,274	0.966	100 / 101	25.10	24.24
abl_dms_mp_face_ctx_full	2	dms_mp_face_ctx_sel_t15_w10_n979	engaged	val	110,948	96,170	0.867	31 / 31	6.16	5.34

coverage — abl_dms_ov_action_learned_full — 2026-09-16T07:26:59Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tiers A + B + C (anchor windows + the full val and test exposure; tier D = train exposure is not needed for operating points); layouts = the tier-A read-outs applied only (fit_readout.py --apply-only, no refit) to the full-exposure summaries results/baselines_dms/summ_full//.

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_ov_action_learned_full	0	dms_ov_action_learned_sel_t15_w10_n979	manual	test	512,934	512,766	1.000	158 / 158	28.50	28.49
abl_dms_ov_action_learned_full	0	dms_ov_action_learned_sel_t15_w10_n979	manual	val	70,330	69,423	0.987	46 / 46	3.91	3.86
abl_dms_ov_action_learned_full	0	dms_ov_action_learned_sel_t15_w10_n979	engaged	test	451,840	451,834	1.000	101 / 101	25.10	25.10
abl_dms_ov_action_learned_full	0	dms_ov_action_learned_sel_t15_w10_n979	engaged	val	110,948	97,018	0.874	31 / 31	6.16	5.39
abl_dms_ov_action_learned_full	1	dms_ov_action_learned_sel_t15_w10_n979	manual	test	512,934	512,766	1.000	158 / 158	28.50	28.49
abl_dms_ov_action_learned_full	1	dms_ov_action_learned_sel_t15_w10_n979	manual	val	70,330	69,423	0.987	46 / 46	3.91	3.86
abl_dms_ov_action_learned_full	1	dms_ov_action_learned_sel_t15_w10_n979	engaged	test	451,840	451,834	1.000	101 / 101	25.10	25.10
abl_dms_ov_action_learned_full	1	dms_ov_action_learned_sel_t15_w10_n979	engaged	val	110,948	97,018	0.874	31 / 31	6.16	5.39
abl_dms_ov_action_learned_full	2	dms_ov_action_learned_sel_t15_w10_n979	manual	test	512,934	512,766	1.000	158 / 158	28.50	28.49
abl_dms_ov_action_learned_full	2	dms_ov_action_learned_sel_t15_w10_n979	manual	val	70,330	69,423	0.987	46 / 46	3.91	3.86
abl_dms_ov_action_learned_full	2	dms_ov_action_learned_sel_t15_w10_n979	engaged	test	451,840	451,834	1.000	101 / 101	25.10	25.10
abl_dms_ov_action_learned_full	2	dms_ov_action_learned_sel_t15_w10_n979	engaged	val	110,948	97,018	0.874	31 / 31	6.16	5.39

coverage — abl_dms_ov_action_ctx_full — 2026-09-16T07:52:17Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tiers A + B + C (anchor windows + the full val and test exposure; tier D = train exposure is not needed for operating points); layouts = the tier-A read-outs applied only (fit_readout.py --apply-only, no refit) to the full-exposure summaries results/baselines_dms/summ_full//.

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_ov_action_ctx_full	0	dms_ov_action_ctx_sel_t15_w10_n979	manual	test	512,934	512,766	1.000	158 / 158	28.50	28.49
abl_dms_ov_action_ctx_full	0	dms_ov_action_ctx_sel_t15_w10_n979	manual	val	70,330	69,423	0.987	46 / 46	3.91	3.86
abl_dms_ov_action_ctx_full	0	dms_ov_action_ctx_sel_t15_w10_n979	engaged	test	451,840	451,834	1.000	101 / 101	25.10	25.10
abl_dms_ov_action_ctx_full	0	dms_ov_action_ctx_sel_t15_w10_n979	engaged	val	110,948	97,018	0.874	31 / 31	6.16	5.39
abl_dms_ov_action_ctx_full	1	dms_ov_action_ctx_sel_t15_w10_n979	manual	test	512,934	512,766	1.000	158 / 158	28.50	28.49
abl_dms_ov_action_ctx_full	1	dms_ov_action_ctx_sel_t15_w10_n979	manual	val	70,330	69,423	0.987	46 / 46	3.91	3.86
abl_dms_ov_action_ctx_full	1	dms_ov_action_ctx_sel_t15_w10_n979	engaged	test	451,840	451,834	1.000	101 / 101	25.10	25.10
abl_dms_ov_action_ctx_full	1	dms_ov_action_ctx_sel_t15_w10_n979	engaged	val	110,948	97,018	0.874	31 / 31	6.16	5.39
abl_dms_ov_action_ctx_full	2	dms_ov_action_ctx_sel_t15_w10_n979	manual	test	512,934	512,766	1.000	158 / 158	28.50	28.49
abl_dms_ov_action_ctx_full	2	dms_ov_action_ctx_sel_t15_w10_n979	manual	val	70,330	69,423	0.987	46 / 46	3.91	3.86
abl_dms_ov_action_ctx_full	2	dms_ov_action_ctx_sel_t15_w10_n979	engaged	test	451,840	451,834	1.000	101 / 101	25.10	25.10
abl_dms_ov_action_ctx_full	2	dms_ov_action_ctx_sel_t15_w10_n979	engaged	val	110,948	97,018	0.874	31 / 31	6.16	5.39

coverage — abl_dms_l2cs_gaze_chain — 2026-09-16T13:06:15Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tiers A + B + C + D (full grid: anchor windows + val / test / train exposure); chain-style arm = the legacy two-state hazard chain trained on the baseline's own history over the FULL train grid (labels t+0.2..t+5.0 s), hands channel shared from our causal pose pack; layouts on test / val (gold anchors + exposure).

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_l2cs_gaze_chain	0	dms_l2cs_gaze_chain_att3s	manual	test	512,934	509,593	0.993	158 / 158	28.50	28.31
abl_dms_l2cs_gaze_chain	0	dms_l2cs_gaze_chain_att3s	manual	val	70,330	62,358	0.887	46 / 46	3.91	3.46
abl_dms_l2cs_gaze_chain	0	dms_l2cs_gaze_chain_att3s	engaged	test	451,840	448,312	0.992	101 / 101	25.10	24.91
abl_dms_l2cs_gaze_chain	0	dms_l2cs_gaze_chain_att3s	engaged	val	110,948	95,861	0.864	31 / 31	6.16	5.33

coverage — abl_dms_ov_gaze_chain — 2026-09-16T13:32:27Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tiers A + B + C + D (full grid: anchor windows + val / test / train exposure); chain-style arm = the legacy two-state hazard chain trained on the baseline's own history over the FULL train grid (labels t+0.2..t+5.0 s), hands channel shared from our causal pose pack; layouts on test / val (gold anchors + exposure).

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_ov_gaze_chain	0	dms_ov_gaze_chain_att3s	manual	test	512,934	510,159	0.995	158 / 158	28.50	28.34
abl_dms_ov_gaze_chain	0	dms_ov_gaze_chain_att3s	manual	val	70,330	61,132	0.869	46 / 46	3.91	3.40
abl_dms_ov_gaze_chain	0	dms_ov_gaze_chain_att3s	engaged	test	451,840	448,294	0.992	101 / 101	25.10	24.91
abl_dms_ov_gaze_chain	0	dms_ov_gaze_chain_att3s	engaged	val	110,948	95,944	0.865	31 / 31	6.16	5.33

coverage — kin_subst/l2cs_gaze (layer-1 substitute driver-kinematics pack for U0_dms_l2cs_gaze_s<k>_kcv2_clip8_ln) — 2026-09-16T13:55:08Z

Missing policy (decisions.md 2026-09-14): a row without a substitute history (row_ok False: no cache / no frame at t / no valid step) is NaN on every channel and is neither trained on nor scored (the U0 trainer drops kin_row_ok False rows; paired_compare_joint scores the rows both layouts predicted, n printed in results/tables/dms_driver_onsets.md). Source: results/baselines_dms/kin_subst/l2cs_gaze/kinhist_causal_*.npz + provenance sidecars (baselines_dms/fit_u_base.py). Eligibility (plan §2): face-only baseline -> head_turn_onset only.

baseline	regime	split	rows	rows with history	share	valid steps given history	hours	hours with history	history source	anchor source	channels
l2cs_gaze	manual	train	632,463	631,865	0.999	0.978	35.14	35.10	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	l2cs_face_conf,l2cs_yaw,l2cs_pitch,l2cs_box_w,valid_flag,pad_5,pad_6,pad_7,pad_8,pad_9,pad_10
l2cs_gaze	manual	val	70,330	69,423	0.987	0.908	3.91	3.86	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	l2cs_face_conf,l2cs_yaw,l2cs_pitch,l2cs_box_w,valid_flag,pad_5,pad_6,pad_7,pad_8,pad_9,pad_10
l2cs_gaze	manual	test	512,934	512,804	1.000	0.993	28.50	28.49	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	l2cs_face_conf,l2cs_yaw,l2cs_pitch,l2cs_box_w,valid_flag,pad_5,pad_6,pad_7,pad_8,pad_9,pad_10
l2cs_gaze	engaged	train	824,859	808,766	0.980	0.984	45.83	44.93	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	l2cs_face_conf,l2cs_yaw,l2cs_pitch,l2cs_box_w,valid_flag,pad_5,pad_6,pad_7,pad_8,pad_9,pad_10
l2cs_gaze	engaged	val	110,948	97,018	0.874	0.989	6.16	5.39	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	l2cs_face_conf,l2cs_yaw,l2cs_pitch,l2cs_box_w,valid_flag,pad_5,pad_6,pad_7,pad_8,pad_9,pad_10
l2cs_gaze	engaged	test	451,840	451,834	1.000	0.992	25.10	25.10	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	l2cs_face_conf,l2cs_yaw,l2cs_pitch,l2cs_box_w,valid_flag,pad_5,pad_6,pad_7,pad_8,pad_9,pad_10

coverage — abl_dms_mp_face_chain — 2026-09-16T13:55:51Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tiers A + B + C + D (full grid: anchor windows + val / test / train exposure); chain-style arm = the legacy two-state hazard chain trained on the baseline's own history over the FULL train grid (labels t+0.2..t+5.0 s), hands channel shared from our causal pose pack; layouts on test / val (gold anchors + exposure).

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_mp_face_chain	0	dms_mp_face_chain_att3s	manual	test	512,934	470,596	0.917	157 / 158	28.50	26.14
abl_dms_mp_face_chain	0	dms_mp_face_chain_att3s	manual	val	70,330	51,667	0.735	46 / 46	3.91	2.87
abl_dms_mp_face_chain	0	dms_mp_face_chain_att3s	engaged	test	451,840	397,391	0.879	100 / 101	25.10	22.08
abl_dms_mp_face_chain	0	dms_mp_face_chain_att3s	engaged	val	110,948	90,246	0.813	31 / 31	6.16	5.01

coverage — kin_subst/mp_face (layer-1 substitute driver-kinematics pack for U0_dms_mp_face_s<k>_kcv2_clip8_ln) — 2026-09-16T13:55:53Z

Missing policy (decisions.md 2026-09-14): a row without a substitute history (row_ok False: no cache / no frame at t / no valid step) is NaN on every channel and is neither trained on nor scored (the U0 trainer drops kin_row_ok False rows; paired_compare_joint scores the rows both layouts predicted, n printed in results/tables/dms_driver_onsets.md). Source: results/baselines_dms/kin_subst/mp_face/kinhist_causal_*.npz + provenance sidecars (baselines_dms/fit_u_base.py). Eligibility (plan §2): face-only baseline -> head_turn_onset only.

baseline	regime	split	rows	rows with history	share	valid steps given history	hours	hours with history	history source	anchor source	channels
mp_face	manual	train	632,463	630,793	0.997	0.826	35.14	35.04	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	pca0,pca1,pca2,pca3,pca4,pca5,pca6,pca7,pca8,pca9,pca10
mp_face	manual	val	70,330	69,423	0.987	0.773	3.91	3.86	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	pca0,pca1,pca2,pca3,pca4,pca5,pca6,pca7,pca8,pca9,pca10
mp_face	manual	test	512,934	512,804	1.000	0.921	28.50	28.49	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	pca0,pca1,pca2,pca3,pca4,pca5,pca6,pca7,pca8,pca9,pca10
mp_face	engaged	train	824,859	792,397	0.961	0.848	45.83	44.02	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	pca0,pca1,pca2,pca3,pca4,pca5,pca6,pca7,pca8,pca9,pca10
mp_face	engaged	val	110,948	97,018	0.874	0.929	6.16	5.39	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	pca0,pca1,pca2,pca3,pca4,pca5,pca6,pca7,pca8,pca9,pca10
mp_face	engaged	test	451,840	451,834	1.000	0.881	25.10	25.10	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	pca0,pca1,pca2,pca3,pca4,pca5,pca6,pca7,pca8,pca9,pca10

coverage — abl_dms_openface3_chain — 2026-09-16T13:58:27Z

Missing policy (decisions.md 2026-09-14): a non-finite score is no alert with unchanged exposure (rows x 0.2 s); finite rows = w_mask & finite(column); decoded routes = routes with >= 1 finite row. Source: the pred_layout files listed (baselines_dms/eval_dms.py coverage_rows). Caches: extraction tiers A + B + C + D (full grid: anchor windows + val / test / train exposure); chain-style arm = the legacy two-state hazard chain trained on the baseline's own history over the FULL train grid (labels t+0.2..t+5.0 s), hands channel shared from our causal pose pack; layouts on test / val (gold anchors + exposure).

arm	seed	column	regime	split	rows	finite rows	share	decoded / routes	hours	finite hours
abl_dms_openface3_chain	0	dms_openface3_chain_att3s	manual	test	512,934	509,593	0.993	158 / 158	28.50	28.31
abl_dms_openface3_chain	0	dms_openface3_chain_att3s	manual	val	70,330	62,358	0.887	46 / 46	3.91	3.46
abl_dms_openface3_chain	0	dms_openface3_chain_att3s	engaged	test	451,840	448,312	0.992	101 / 101	25.10	24.91
abl_dms_openface3_chain	0	dms_openface3_chain_att3s	engaged	val	110,948	95,861	0.864	31 / 31	6.16	5.33

coverage — kin_subst/videomae2 (layer-1 substitute driver-kinematics pack for U0_dms_videomae2_s<k>_kcv2_clip8_ln) — 2026-09-16T14:58:44Z

Missing policy (decisions.md 2026-09-14): a row without a substitute history (row_ok False: no cache / no frame at t / no valid step) is NaN on every channel and is neither trained on nor scored (the U0 trainer drops kin_row_ok False rows; paired_compare_joint scores the rows both layouts predicted, n printed in results/tables/dms_driver_onsets.md). Source: results/baselines_dms/kin_subst/videomae2/kinhist_causal_*.npz + provenance sidecars (baselines_dms/fit_u_base.py). Eligibility (plan §2): all three driver events.

baseline	regime	split	rows	rows with history	share	valid steps given history	hours	hours with history	history source	anchor source	channels
videomae2	manual	train	632,463	631,745	0.999	0.481	35.14	35.10	summ_emb_anchor_5Hz_anchor_rows (odd 10 Hz steps invalid)	emb	pca0,pca1,pca2,pca3,pca4,pca5,pca6,pca7,pca8,pca9,pca10
videomae2	manual	val	70,330	69,423	0.987	0.482	3.91	3.86	summ_emb_anchor_5Hz_anchor_rows (odd 10 Hz steps invalid)	emb	pca0,pca1,pca2,pca3,pca4,pca5,pca6,pca7,pca8,pca9,pca10
videomae2	manual	test	512,934	512,748	1.000	0.485	28.50	28.49	summ_emb_anchor_5Hz_anchor_rows (odd 10 Hz steps invalid)	emb	pca0,pca1,pca2,pca3,pca4,pca5,pca6,pca7,pca8,pca9,pca10
videomae2	engaged	train	824,859	808,766	0.980	0.484	45.83	44.93	summ_emb_anchor_5Hz_anchor_rows (odd 10 Hz steps invalid)	emb	pca0,pca1,pca2,pca3,pca4,pca5,pca6,pca7,pca8,pca9,pca10
videomae2	engaged	val	110,948	97,018	0.874	0.493	6.16	5.39	summ_emb_anchor_5Hz_anchor_rows (odd 10 Hz steps invalid)	emb	pca0,pca1,pca2,pca3,pca4,pca5,pca6,pca7,pca8,pca9,pca10
videomae2	engaged	test	451,840	451,834	1.000	0.490	25.10	25.10	summ_emb_anchor_5Hz_anchor_rows (odd 10 Hz steps invalid)	emb	pca0,pca1,pca2,pca3,pca4,pca5,pca6,pca7,pca8,pca9,pca10

coverage — kin_subst/ov_action (layer-1 substitute driver-kinematics pack for U0_dms_ov_action_s<k>_kcv2_clip8_ln) — 2026-09-16T14:58:44Z

Missing policy (decisions.md 2026-09-14): a row without a substitute history (row_ok False: no cache / no frame at t / no valid step) is NaN on every channel and is neither trained on nor scored (the U0 trainer drops kin_row_ok False rows; paired_compare_joint scores the rows both layouts predicted, n printed in results/tables/dms_driver_onsets.md). Source: results/baselines_dms/kin_subst/ov_action/kinhist_causal_*.npz + provenance sidecars (baselines_dms/fit_u_base.py). Eligibility (plan §2): all three driver events.

baseline	regime	split	rows	rows with history	share	valid steps given history	hours	hours with history	history source	anchor source	channels
ov_action	manual	train	632,463	631,800	0.999	0.481	35.14	35.10	summ_emb_anchor_5Hz_anchor_rows (odd 10 Hz steps invalid)	emb	emb0,emb1,emb2,emb3,emb4,emb5,emb6,emb7,emb8,valid_flag,pad_10
ov_action	manual	val	70,330	69,423	0.987	0.482	3.91	3.86	summ_emb_anchor_5Hz_anchor_rows (odd 10 Hz steps invalid)	emb	emb0,emb1,emb2,emb3,emb4,emb5,emb6,emb7,emb8,valid_flag,pad_10
ov_action	manual	test	512,934	512,771	1.000	0.485	28.50	28.49	summ_emb_anchor_5Hz_anchor_rows (odd 10 Hz steps invalid)	emb	emb0,emb1,emb2,emb3,emb4,emb5,emb6,emb7,emb8,valid_flag,pad_10
ov_action	engaged	train	824,859	808,766	0.980	0.484	45.83	44.93	summ_emb_anchor_5Hz_anchor_rows (odd 10 Hz steps invalid)	emb	emb0,emb1,emb2,emb3,emb4,emb5,emb6,emb7,emb8,valid_flag,pad_10
ov_action	engaged	val	110,948	97,018	0.874	0.493	6.16	5.39	summ_emb_anchor_5Hz_anchor_rows (odd 10 Hz steps invalid)	emb	emb0,emb1,emb2,emb3,emb4,emb5,emb6,emb7,emb8,valid_flag,pad_10
ov_action	engaged	test	451,840	451,834	1.000	0.490	25.10	25.10	summ_emb_anchor_5Hz_anchor_rows (odd 10 Hz steps invalid)	emb	emb0,emb1,emb2,emb3,emb4,emb5,emb6,emb7,emb8,valid_flag,pad_10

coverage — kin_subst/ov_gaze (layer-1 substitute driver-kinematics pack for U0_dms_ov_gaze_s<k>_kcv2_clip8_ln) — 2026-09-16T16:11:46Z

Missing policy (decisions.md 2026-09-14): a row without a substitute history (row_ok False: no cache / no frame at t / no valid step) is NaN on every channel and is neither trained on nor scored (the U0 trainer drops kin_row_ok False rows; paired_compare_joint scores the rows both layouts predicted, n printed in results/tables/dms_driver_onsets.md). Source: results/baselines_dms/kin_subst/ov_gaze/kinhist_causal_*.npz + provenance sidecars (baselines_dms/fit_u_base.py). Eligibility (plan §2): face-only baseline -> head_turn_onset only.

baseline	regime	split	rows	rows with history	share	valid steps given history	hours	hours with history	history source	anchor source	channels
ov_gaze	manual	train	632,463	631,865	0.999	0.976	35.14	35.10	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	ov_face_conf,ov_head_yaw,ov_head_pitch,ov_head_roll,ov_gaze_x,ov_gaze_y,ov_gaze_z,ov_gaze_yaw,ov_gaze_pitch,ov_eye_closed_l,ov_eye_closed_r
ov_gaze	manual	val	70,330	69,423	0.987	0.891	3.91	3.86	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	ov_face_conf,ov_head_yaw,ov_head_pitch,ov_head_roll,ov_gaze_x,ov_gaze_y,ov_gaze_z,ov_gaze_yaw,ov_gaze_pitch,ov_eye_closed_l,ov_eye_closed_r
ov_gaze	manual	test	512,934	512,804	1.000	0.994	28.50	28.49	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	ov_face_conf,ov_head_yaw,ov_head_pitch,ov_head_roll,ov_gaze_x,ov_gaze_y,ov_gaze_z,ov_gaze_yaw,ov_gaze_pitch,ov_eye_closed_l,ov_eye_closed_r
ov_gaze	engaged	train	824,859	808,766	0.980	0.984	45.83	44.93	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	ov_face_conf,ov_head_yaw,ov_head_pitch,ov_head_roll,ov_gaze_x,ov_gaze_y,ov_gaze_z,ov_gaze_yaw,ov_gaze_pitch,ov_eye_closed_l,ov_eye_closed_r
ov_gaze	engaged	val	110,948	97,018	0.874	0.989	6.16	5.39	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	ov_face_conf,ov_head_yaw,ov_head_pitch,ov_head_roll,ov_gaze_x,ov_gaze_y,ov_gaze_z,ov_gaze_yaw,ov_gaze_pitch,ov_eye_closed_l,ov_eye_closed_r
ov_gaze	engaged	test	451,840	451,834	1.000	0.992	25.10	25.10	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	ov_face_conf,ov_head_yaw,ov_head_pitch,ov_head_roll,ov_gaze_x,ov_gaze_y,ov_gaze_z,ov_gaze_yaw,ov_gaze_pitch,ov_eye_closed_l,ov_eye_closed_r

coverage — kin_subst/openface3 (layer-1 substitute driver-kinematics pack for U0_dms_openface3_s<k>_kcv2_clip8_ln) — 2026-09-16T16:11:46Z

Missing policy (decisions.md 2026-09-14): a row without a substitute history (row_ok False: no cache / no frame at t / no valid step) is NaN on every channel and is neither trained on nor scored (the U0 trainer drops kin_row_ok False rows; paired_compare_joint scores the rows both layouts predicted, n printed in results/tables/dms_driver_onsets.md). Source: results/baselines_dms/kin_subst/openface3/kinhist_causal_*.npz + provenance sidecars (baselines_dms/fit_u_base.py). Eligibility (plan §2): face-only baseline -> head_turn_onset only.

baseline	regime	split	rows	rows with history	share	valid steps given history	hours	hours with history	history source	anchor source	channels
openface3	manual	train	632,463	631,865	0.999	0.978	35.14	35.10	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	pca0,pca1,pca2,pca3,pca4,pca5,pca6,pca7,pca8,pca9,pca10
openface3	manual	val	70,330	69,423	0.987	0.908	3.91	3.86	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	pca0,pca1,pca2,pca3,pca4,pca5,pca6,pca7,pca8,pca9,pca10
openface3	manual	test	512,934	512,804	1.000	0.993	28.50	28.49	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	pca0,pca1,pca2,pca3,pca4,pca5,pca6,pca7,pca8,pca9,pca10
openface3	engaged	train	824,859	808,766	0.980	0.984	45.83	44.93	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	pca0,pca1,pca2,pca3,pca4,pca5,pca6,pca7,pca8,pca9,pca10
openface3	engaged	val	110,948	97,018	0.874	0.989	6.16	5.39	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	pca0,pca1,pca2,pca3,pca4,pca5,pca6,pca7,pca8,pca9,pca10
openface3	engaged	test	451,840	451,834	1.000	0.992	25.10	25.10	cache_10Hz (dataset_adapter.route_hist / summarise.window_index)	cache_frames (per-frame channels of results/baselines_dms/cache//.npz; the summary anchor-source flag does not apply)	pca0,pca1,pca2,pca3,pca4,pca5,pca6,pca7,pca8,pca9,pca10