Datasets:

NagaYu
/

plumb-skipped-part-bench

cond stringclasses 6 values	gate stringclasses 6 values	seed int64 100 100	processed_fraction float64 0.03 0.84	skipped_err float64 0.01 0.18	all_err float64 0 0.15	processed_err float64 -0 0.03	n_encode_processed int64 1.17k 30.2k	n_encode_explore int64 0 1.56k	audit_n float64 44 1.56k ⌀	audit_est float64 0.01 0.19 ⌀	audit_lo float64 0.01 0.16 ⌀	audit_hi float64 0.01 0.21 ⌀	alarms stringlengths 2 1.75k	key stringclasses 8 values	stream stringclasses 1 value	suite stringclasses 1 value	elapsed_s float64 0.32 710	audit_rate float64 0.01 0.01 ⌀	lambda_final float64 0.25 8 ⌀	lambda_mean float64 0.41 7.65 ⌀	self_assessed_err float64 0.02 0.06 ⌀	trust_mean float64 0 0.51 ⌀	trust_final float64 0 1 ⌀	explore_rate float64 0.01 0.05 ⌀	trust_rule stringclasses 2 values
B	egotrigger-1s	100	0.123333	0.135959	0.119191	-0	4,440	315	315	0.141771	0.123061	0.160481	[355, 1080, 4916, 6024, 7435, 7812, 7855, 8054, 8096, 13608, 14181, 14232, 16864, 17185, 17271, 19084, 19216, 19503, 20617, 20673, 20857, 21009, 21130, 21152, 21163, 21202, 21203, 22546, 22699, 22718, 22798, 22893, 22976, 23024, 24170, 25729, 26198, 26403, 26687, 29052, 32019, 32579, 33047, 33133, 35248]	B	s100_d3600	long	0.756098	0.01	null	null	null	null	null	null	null
C	egotrigger-1s	100	0.101667	0.153369	0.137776	-0	3,660	324	324	0.160221	0.139871	0.18057	[231, 924, 1189, 4831, 6004, 6364, 7722, 7984, 8026, 8060, 8098, 8174, 13631, 14178, 16935, 17203, 17234, 17499, 18157, 18348, 19093, 19521, 20428, 20457, 20609, 20730, 20752, 20763, 20802, 20803, 21244, 21266, 22218, 22604, 22786, 22909, 23045, 23053, 23109, 24241, 24345, 25536, 25701, 25733, 26621, 27081, 27497, 3105...	C	s100_d3600	long	1.758424	0.01	8	7.405062	0.031173	null	null	null	null
C-corr	egotrigger-1s	100	0.123333	0.139333	0.122148	-0	4,440	315	315	0.144096	0.125414	0.162778	[231, 924, 1189, 4916, 6024, 7435, 7812, 7855, 8054, 8096, 13608, 14181, 14232, 16864, 17185, 17271, 19084, 19216, 19503, 20617, 20673, 20857, 21009, 21130, 21152, 21163, 21202, 21203, 22546, 22699, 22718, 22798, 22893, 22976, 23024, 24170, 25729, 26198, 26403, 26687, 29064, 31897, 32579, 33047, 33133, 35393]	C-corr	s100_d3600	long	2.493852	0.01	null	null	0.032762	null	null	null	null
C-tune	egotrigger-1s	100	0.101667	0.151566	0.136157	-0	3,660	324	324	0.159673	0.138894	0.180452	[355, 1080, 4831, 6004, 6364, 7722, 7945, 8025, 8060, 8098, 8174, 13631, 14178, 16935, 17203, 17234, 17499, 18157, 18348, 19093, 19521, 20428, 20457, 20609, 20730, 20752, 20763, 20802, 20803, 21244, 21266, 22218, 22604, 22786, 22909, 23045, 23053, 23109, 24241, 25450, 25701, 25733, 26621, 27081, 27497, 29165, 31059, 31...	C-tune	s100_d3600	long	1.001142	0.01	8	7.426418	0.030194	null	null	null	null
E	egotrigger-1s	100	0.123333	0.091664	0.080359	-0	4,440	0	null	null	null	null	NaN	E	s100_d3600	long	8.704338	null	null	null	null	null	null	null	null
D	egotrigger-1s	100	0.123333	0.136111	0.119324	-0	4,440	302	302	0.152583	0.131815	0.173351	[1044, 1762, 4587, 4908, 6119, 6328, 7852, 7936, 14206, 16598, 16895, 17168, 17236, 18092, 18139, 19040, 19511, 19548, 20578, 20686, 20988, 21013, 21243, 22242, 22582, 22678, 22696, 22698, 22815, 23093, 23930, 25743, 26400, 26443, 26533, 26675, 27023, 27035, 27037, 29224, 32080, 33036, 34412, 34458, 34473, 35800, 35999...	D-sign-p0.01	s100_d3600	long	1.748153	null	null	null	null	0.277338	0.15	0.01	sign
D	egotrigger-1s	100	0.123333	0.135959	0.119191	-0	4,440	302	302	0.152671	0.131805	0.173538	[1044, 1762, 4587, 4908, 6119, 6328, 7852, 7936, 14206, 16598, 16895, 17168, 17236, 18092, 18139, 19040, 19511, 19548, 20578, 20686, 20988, 21013, 21243, 22242, 22582, 22678, 22696, 22698, 22815, 23093, 23930, 25743, 26400, 26443, 26533, 26675, 27023, 27035, 27037, 29193, 32080, 33036, 34412, 34458, 34473, 35800, 35999...	D-lcb-p0.01	s100_d3600	long	1.08318	null	null	null	null	0	0	0.01	lcb
D	egotrigger-1s	100	0.123333	0.135171	0.1185	-0	4,440	1,501	1,501	0.141572	0.132913	0.150231	[102, 122, 292, 350, 912, 1044, 1103, 1147, 1204, 1226, 1323, 1788, 2276, 2379, 3305, 3328, 3360, 4276, 4342, 4569, 4600, 4796, 4866, 4894, 4931, 5658, 5817, 5855, 5939, 6068, 6086, 6141, 6241, 6372, 7635, 7827, 7847, 7900, 7936, 8016, 8024, 9316, 9436, 10608, 12048, 12179, 12236, 13235, 13263, 13348, 13630, 14159, 142...	D-sign-p0.05	s100_d3600	long	1.944821	null	null	null	null	0.283558	1	0.05	sign
B	egotrigger-hyst	100	0.207778	0.166774	0.132122	-0	7,480	291	291	0.162382	0.141046	0.183719	[1160, 3356, 5576, 5738, 5838, 6265, 7826, 8095, 8175, 13605, 16625, 17005, 17216, 17319, 17461, 17787, 17951, 19193, 20427, 20451, 20705, 20750, 20797, 20862, 20958, 21168, 21214, 21222, 22724, 22837, 22850, 22893, 23048, 23077, 24367, 25454, 26195, 26259, 26678, 28924, 29146, 29856, 30760, 31908, 32257, 34955, 35281,...	B	s100_d3600	long	0.401631	0.01	null	null	null	null	null	null	null
C	egotrigger-hyst	100	0.186111	0.180583	0.146975	-0	6,700	298	298	0.18535	0.163177	0.207523	[919, 1160, 2303, 3296, 4762, 5658, 5758, 6185, 7713, 7955, 8053, 8065, 8183, 9482, 9526, 10415, 11992, 13578, 13850, 16461, 16831, 17001, 17167, 17291, 19176, 19527, 20568, 20614, 20622, 20746, 20760, 20869, 20983, 21078, 22050, 22248, 22690, 22712, 22723, 22762, 22763, 24096, 24288, 26048, 26507, 26631, 27069, 31046,...	C	s100_d3600	long	2.417152	0.01	8	7.646398	0.025568	null	null	null	null
C-corr	egotrigger-hyst	100	0.207778	0.172657	0.136783	-0	7,480	291	291	0.165236	0.144306	0.186166	[919, 1160, 2323, 3356, 5576, 5738, 5838, 6265, 7826, 8095, 8175, 13605, 16625, 17005, 17216, 17319, 17461, 17787, 17951, 19193, 20427, 20451, 20705, 20750, 20797, 20862, 20958, 21168, 21214, 21222, 22724, 22837, 22850, 22893, 23048, 23077, 24066, 25454, 26195, 26259, 26678, 28924, 29146, 29856, 30760, 31908, 32257, 34...	C-corr	s100_d3600	long	2.855616	0.01	null	null	0.025983	null	null	null	null
C-tune	egotrigger-hyst	100	0.186111	0.174824	0.142287	-0	6,700	298	298	0.180886	0.158425	0.203347	[1160, 3275, 4762, 5658, 5758, 6185, 7713, 7955, 8053, 8065, 9482, 9526, 10415, 13578, 13850, 16545, 16859, 17098, 17220, 17412, 19176, 19527, 20568, 20614, 20622, 20746, 20760, 20869, 20983, 21078, 22050, 22248, 22690, 22712, 22723, 22762, 22763, 24096, 24368, 26048, 26507, 26631, 27069, 31046, 31068, 32281, 33199, 34...	C-tune	s100_d3600	long	1.650658	0.01	8	7.648263	0.024949	null	null	null	null
E	egotrigger-hyst	100	0.207778	0.106663	0.084501	-0	7,480	0	null	null	null	null	NaN	E	s100_d3600	long	9.571151	null	null	null	null	null	null	null	null
D	egotrigger-hyst	100	0.207778	0.166294	0.131742	-0	7,480	270	270	0.159848	0.138275	0.181421	[572, 992, 1204, 4548, 5568, 5874, 6028, 7629, 8087, 13633, 16595, 17060, 17190, 17402, 19511, 20422, 20431, 20453, 20469, 20475, 20564, 20674, 20775, 21042, 21123, 21170, 22737, 22798, 22845, 22906, 22979, 24172, 24291, 24390, 26270, 26580, 29115, 29857, 32154, 32325, 33196, 34428, 35143, 35709]	D-sign-p0.01	s100_d3600	long	3.539719	null	null	null	null	0.174601	0.35	0.01	sign
D	egotrigger-hyst	100	0.207778	0.166774	0.132122	-0	7,480	270	270	0.159758	0.138009	0.181508	[572, 992, 1204, 4548, 5568, 5874, 6028, 7629, 8087, 13633, 16595, 17060, 17190, 17402, 19511, 20422, 20431, 20453, 20469, 20475, 20564, 20674, 20775, 21042, 21123, 21170, 22737, 22798, 22845, 22906, 22979, 24172, 24313, 26270, 26580, 29115, 29857, 32154, 32325, 33196, 34428, 35143, 35709]	D-lcb-p0.01	s100_d3600	long	1.932004	null	null	null	null	0	0	0.01	lcb
D	egotrigger-hyst	100	0.207778	0.165463	0.131083	-0	7,480	1,364	1,364	0.167524	0.157651	0.177398	[276, 372, 941, 953, 1002, 1130, 1204, 1235, 1285, 1347, 1457, 1499, 1767, 1779, 2270, 2344, 4278, 4369, 4486, 4548, 4627, 4796, 4844, 4862, 5553, 5591, 5744, 5799, 5879, 5941, 6127, 6228, 6359, 7430, 7436, 7629, 7690, 7732, 7818, 7906, 7953, 8003, 8043, 8046, 8056, 8076, 8087, 8095, 8100, 8108, 8195, 12208, 13227, 132...	D-sign-p0.05	s100_d3600	long	2.630461	null	null	null	null	0.321836	1	0.05	sign
B	metas-framerate	100	0.2	0.027205	0.021764	-0	7,200	293	293	0.027801	0.024478	0.031124	[14256, 17289, 22538, 26398, 30673, 35774]	B	s100_d3600	long	0.361398	0.01	null	null	null	null	null	null	null
C	metas-framerate	100	0.123722	0.034579	0.030301	-0	4,454	315	315	0.035488	0.02987	0.041106	[8156, 13998, 15642, 18239, 19544, 20371, 22337, 26379, 30677, 33325, 34202, 34896, 35047]	C	s100_d3600	long	2.033823	0.01	2.974274	2.094773	0.040263	null	null	null	null
C-corr	metas-framerate	100	0.2	0.027902	0.022321	-0	7,200	293	293	0.028468	0.025092	0.031843	[14256, 17289, 22538, 26398, 30673, 35774]	C-corr	s100_d3600	long	1.627432	0.01	null	null	0.034096	null	null	null	null
C-tune	metas-framerate	100	0.088417	0.035841	0.032672	-0	3,183	327	327	0.032488	0.027794	0.037183	[7619, 13384, 13760, 17907, 17976, 18359, 20356, 23253, 25385, 31044, 32036, 34122, 34914, 35808]	C-tune	s100_d3600	long	0.434905	0.01	3.019224	2.622659	0.044728	null	null	null	null
E	metas-framerate	100	0.2	0.02742	0.021936	-0	7,200	0	null	null	null	null	NaN	E	s100_d3600	long	8.157969	null	null	null	null	null	null	null	null
D	metas-framerate	100	0.2	0.027205	0.021764	-0	7,200	272	272	0.02667	0.022989	0.030351	[7712, 18311, 18312, 32062]	D-sign-p0.01	s100_d3600	long	3.377126	null	null	null	null	0.000462	0	0.01	sign
D	metas-framerate	100	0.2	0.027205	0.021764	-0	7,200	272	272	0.02667	0.022989	0.030351	[7712, 18311, 18312, 32062]	D-lcb-p0.01	s100_d3600	long	1.541867	null	null	null	null	0	0	0.01	lcb
D	metas-framerate	100	0.2	0.02721	0.021768	-0	7,200	1,375	1,375	0.026756	0.025021	0.028492	[1744, 1829, 2403, 7613, 7989, 8154, 8196, 9408, 12229, 13307, 13538, 17319, 18311, 18312, 19109, 20419, 21286, 22532, 22534, 22588, 23153, 23224, 23474, 26268, 27603, 29244, 32062, 32303, 33253, 34092, 34928, 35799]	D-sign-p0.05	s100_d3600	long	1.641297	null	null	null	null	0.032101	0	0.05	sign
B	metas-random	100	0.097389	0.039948	0.036058	-0	3,506	325	325	0.037144	0.031314	0.042974	[8042, 32265, 34418, 34898]	B	s100_d3600	long	0.324642	0.01	null	null	null	null	null	null	null
C	metas-random	100	0.03975	0.065523	0.062918	-0	1,431	344	344	0.065369	0.054658	0.07608	[3276, 3350, 7527, 7720, 7763, 7764, 7798, 8194, 8195, 13231, 13653, 14261, 18351, 18928, 23161, 25787, 27117, 27577, 29139, 31975, 34401, 35042, 35679, 35680, 35720]	C	s100_d3600	long	1.194141	0.01	3.205923	2.726402	0.063978	null	null	null	null
C-corr	metas-random	100	0.097389	0.042223	0.038111	-0	3,506	325	325	0.039554	0.03325	0.045858	[7744, 8043, 32265, 34418, 34898]	C-corr	s100_d3600	long	1.403015	0.01	null	null	0.044478	null	null	null	null
C-tune	metas-random	100	0.032361	0.071	0.068702	-0	1,165	344	344	0.082091	0.068227	0.095955	[3359, 7425, 7428, 7435, 7707, 7708, 7782, 8224, 13250, 13592, 13659, 14110, 16220, 16377, 17845, 18259, 18992, 23165, 23240, 24018, 24305, 25892, 26243, 26378, 27043, 27111, 30676, 30706, 30754, 31142, 31961, 33118, 34217, 34421, 34891, 34963, 35627]	C-tune	s100_d3600	long	0.416943	0.01	4.284483	3.38713	0.06494	null	null	null	null
E	metas-random	100	0.097389	0.036424	0.032877	-0	3,506	0	null	null	null	null	NaN	E	s100_d3600	long	7.384582	null	null	null	null	null	null	null	null
D	metas-random	100	0.097389	0.039948	0.036058	-0	3,506	313	313	0.034265	0.029005	0.039525	[20370, 32215, 35979]	D-sign-p0.01	s100_d3600	long	1.534323	null	null	null	null	0	0	0.01	sign
D	metas-random	100	0.097389	0.039948	0.036058	-0	3,506	313	313	0.034265	0.029005	0.039525	[20370, 32215, 35979]	D-lcb-p0.01	s100_d3600	long	1.481149	null	null	null	null	0	0	0.01	lcb
D	metas-random	100	0.097389	0.039954	0.036063	-0	3,506	1,559	1,559	0.036967	0.034289	0.039645	[887, 2377, 7760, 10434, 13228, 13387, 13392, 14098, 14277, 17831, 18226, 18326, 18999, 19003, 19545, 20370, 22590, 23156, 23450, 25499, 25590, 26263, 26268, 27606, 32215, 32228, 33191, 33208, 34198, 34955, 35738, 35765, 35900, 35979]	D-sign-p0.05	s100_d3600	long	2.158438	null	null	null	null	0.003575	0	0.05	sign
B	epic-bypass	100	0.298222	0.017622	0.012367	-0	10,736	256	256	0.017817	0.016324	0.019309	[10076, 11731, 14903, 19785, 19986, 20784, 29859, 32418]	B	s100_d3600	long	0.901536	0.01	null	null	null	null	null	null	null
C	epic-bypass	100	0.817417	0.009253	0.001689	-0	29,427	50	50	0.00865	0.006988	0.010312	[]	C	s100_d3600	long	164.613971	0.01	0.25	0.432144	0.025023	null	null	null	null
C-corr	epic-bypass	100	0.298222	0.017314	0.01215	-0	10,736	256	256	0.017352	0.015933	0.018771	[10076, 12027, 19892, 23407, 29859, 32418]	C-corr	s100_d3600	long	14.086852	0.01	null	null	0.038162	null	null	null	null
C-tune	epic-bypass	100	0.838306	0.008856	0.001432	-0	30,179	44	44	0.009112	0.007801	0.010423	[]	C-tune	s100_d3600	long	118.691867	0.01	0.25	0.412295	0.02535	null	null	null	null
E	epic-bypass	100	0.298222	0.016835	0.011814	-0	10,736	0	null	null	null	null	NaN	E	s100_d3600	long	30.925084	null	null	null	null	null	null	null	null
D	epic-bypass	100	0.298222	0.017463	0.012255	-0	10,736	238	238	0.016626	0.014861	0.01839	[21265, 22605, 22987, 23143]	D-sign-p0.01	s100_d3600	long	10.095179	null	null	null	null	0.379983	0.4	0.01	sign
D	epic-bypass	100	0.298222	0.017523	0.012297	-0	10,736	238	238	0.016688	0.014914	0.018463	[21265, 22605, 22987, 23143, 23434]	D-lcb-p0.01	s100_d3600	long	9.003276	null	null	null	null	0.316304	0.702551	0.01	lcb
D	epic-bypass	100	0.298222	0.017333	0.012164	-0	10,736	1,195	1,195	0.017644	0.016703	0.018585	[923, 1148, 1691, 10199, 12284, 12398, 19755, 19917, 20147, 20467, 20548, 20754, 20875, 20910, 21145, 21265, 21591, 21663, 22068, 22739, 22818, 22824, 22835, 23077, 23085, 23089, 23111, 23123, 23143, 23164, 23171, 23333, 23963, 25393, 25425, 29273, 29671, 32292, 32457]	D-sign-p0.05	s100_d3600	long	8.969142	null	null	null	null	0.511642	0.9	0.05	sign
B	epic	100	0.298222	0.025149	0.02045	0.009392	10,736	256	256	0.021954	0.019405	0.024502	[]	B	s100_d3600	long	1.09061	0.01	null	null	null	null	null	null	null
C	epic	100	0.096	0.037237	0.036098	0.025366	3,456	325	325	0.037534	0.031296	0.043772	[18999, 19007, 20902, 22595, 23165, 25401]	C	s100_d3600	long	709.685035	0.01	2.745601	2.920316	0.06425	null	null	null	null
C-corr	epic	100	0.298222	0.024857	0.020245	0.009392	10,736	256	256	0.021487	0.018965	0.024008	[]	C-corr	s100_d3600	long	12.623033	0.01	null	null	0.040876	null	null	null	null
C-tune	epic	100	0.090917	0.037491	0.036484	0.026415	3,273	326	326	0.034635	0.028925	0.040345	[19000, 23213, 25494]	C-tune	s100_d3600	long	512.541347	0.01	3.111174	3.013304	0.064343	null	null	null	null
E	epic	100	0.298222	0.021163	0.017653	0.009392	10,736	0	null	null	null	null	NaN	E	s100_d3600	long	19.464406	null	null	null	null	null	null	null	null
D	epic	100	0.298222	0.024977	0.020329	0.009392	10,736	238	238	0.023295	0.019844	0.026745	[1282]	D-sign-p0.01	s100_d3600	long	6.283146	null	null	null	null	0.368199	0.4	0.01	sign
D	epic	100	0.298222	0.025076	0.020399	0.009392	10,736	238	238	0.02341	0.019955	0.026865	[1282]	D-lcb-p0.01	s100_d3600	long	4.767631	null	null	null	null	0.2453	0.596944	0.01	lcb
D	epic	100	0.298222	0.024857	0.020245	0.009392	10,736	1,195	1,195	0.024342	0.022746	0.025938	[78, 171, 1012, 1127, 1347, 1460, 1593, 23089, 23164]	D-sign-p0.05	s100_d3600	long	5.105892	null	null	null	null	0.50331	0.85	0.05	sign

Plumb — skipped-part degradation of gated egocentric perception

Measurements from Plumb, an audit layer that sits on top of always-on perception gates (EPIC, EgoTrigger, Ego-METAS) and checks the part each gate skips: whether online adaptation silently degrades it, and whether enabling the layer can ever be worse than the gate alone.

This is not an improvement to EPIC, EgoTrigger or Ego-METAS, and it claims no energy reduction. It is a test of a correctness condition, evaluated on top of them.

Synthetic data only. Every number here comes from a scripted, ray-cast synthetic egocentric stream. No natural first-person video (Ego4D, Project Aria) was used, and no conclusion has been reproduced on natural video.

A published negative result, in part. 5 of 7 pre-registered acceptance gates failed (G2, G4, G5, G6, G7), and are recorded as failures.

Headline numbers

	result
Zero-trust guarantee (output bit-identical to the gate alone)	holds — enforced by a randomised test
"Never worse than the gate" with trust > 0, rule as specified	fails: 40 of 175 runs worse (max +3.73e-04 in 1 − cos); LCB rule 4 of 175
Naive adaptive system's own estimate of its skipped-part error	too low by ≥ 4.6× on EgoTrigger; too high on EPIC
Plumb's audit at 1 % exploration	within 0.07 (relative) of the truth on EgoTrigger; CI coverage 0.872 (pre-registered ≥ 0.90: fails)
Measured energy per ViT encode (laptop CPU)	0.197 J; 30–34 pJ/FLOP

💻 Code: https://github.com/NagaYu/plumb
📄 Full write-up, including every hypothesis that did not hold: RESULTS.md

What is in this dataset

path	content
`summaries/main_summary.parquet`	one row per run: 5 test seeds × 6 gates × conditions (B, C, C-corr, C-tune, E, D at 7 exploration rates × 2 trust rules)
`summaries/long_summary.parquet`	the 60-minute drift stream, one row per (gate, condition)
`summaries/posthoc_summary.parquet`	post-hoc ablations (not pre-registered)
`runs/main/<stream>__<gate>.npz`, `runs/long/…`	per-step arrays (see below)
`protocol/preregistration.json`	every choice fixed before any condition ran
`protocol/calibration.json`	calibration of gate thresholds, error tolerances, correction hyperparameters; deviations
`protocol/report.json`	every derived number, acceptance-gate verdict and table
`energy/<stamp>/`	the energy measurement: 10 Hz samples of system power (SMC `PSTR`) and die temperatures, per-run windows, analysis, power-source labelling, conditions of measurement (personal identifiers removed)
`figures/`	figures, generated from the measurements

Columns of the summaries

cond / key (condition; D-<rule>-p<rate> for Plumb), gate, stream (s<seed>_d<seconds>), processed_fraction, skipped_err (mean 1 − cos to the full-encode embedding over the items the condition skipped — the primary metric), all_err, processed_err, n_encode_processed, n_encode_explore, trust_mean/trust_final (D), lambda_final/lambda_mean (the naive tuner's aggressiveness, C and C-tune), self_assessed_err (the naive system's own prequential error on processed items), audit_est/audit_lo/audit_hi/audit_n (Plumb's audit), alarms (CUSUM alarm steps, JSON).

Keys of the per-step `.npz` arrays

key	shape	meaning
`processed_B`	(N,) bool	the gate's own decisions — shared by B, C-corr, D, E
`segment`	(N,) int	script segment of each frame (still, slowpan, lookaround, walk, hoi, light, periphery, screen, roomchange)
`<cond>__err`	(N,) float32	per-frame 1 − cos of the emitted output vs. the full encode
`<cond>__processed`	(N,) bool	decisions of conditions that move the gate (C, C-tune)
`<cond>__trust`	(N,)	trust in force (D)
`<cond>__lambda`	(N,)	naive tuner's aggressiveness (C, C-tune)
`<cond>__surprise`	(M, 2)	(step, prequential error) on processed items — the naive system's own view
`<cond>__audit`	(K, 3)	(step, error of the emitted output, error of the plain reuse value) on explored items
`<cond>__alarms`	(A,)	CUSUM alarm steps

N = 12 000 frames (20 min at 10 FPS) for test streams, 36 000 for the long stream.

How the data were produced

Stream: ray-cast head-mounted camera in two furnished rooms, 168×168 at 10 FPS, with ground-truth pose and depth, IMU and 16 kHz audio; a script of stillness, pans, walking, hand–object interactions (with silent interactions and off-camera interaction sounds by construction), lighting changes, peripheral motion and room changes. The 60-minute stream re-arranges its decor every 10 minutes and follows an illumination cycle.
Encoder: facebook/dinov2-small CLS token, L2-normalised.
Gates (reproduced from the papers; interpretations listed in adapters/README.md of the code): EPIC (arXiv:2606.15859) frame bypass + temporal-spatial redundancy check; EgoTrigger (arXiv:2508.01915) ET-1s and ET-Hyst. with the paper's head on a stand-in for YAMNet embeddings; Ego-METAS (arXiv:2606.02246) Frame Rate and Random routing.
Conditions: (A) full encode; (B) gate; (C) gate + naive online adaptation; (D) gate + Plumb; (E) gate + oracle correction (diagnostic).
Energy: measured on an Apple M2 laptop CPU with the SMC system-power sensor (no root), idle-bracketed, interleaved, 3 repeats, under ~30 W of unrelated background load; the machine started the session on battery and was reconnected to AC part-way (runs are labelled).

Limitations

Synthetic only. Selection effects depend on how bursty processing is and how fast content changes — properties of this stream.
Gates reproduced from text; EPIC's saliency CNN omitted; EgoTrigger's classifier replaced by a stand-in that is easier to satisfy on synthetic sounds than the paper's was on real audio.
One encoder and one error measure; no task-level metric.
Energy on a laptop SoC under concurrent load, not a wearable; sensor power not measured.

Personal data

None. The stream is synthetic. The energy record was passed through a sanitiser that removes home-directory paths, process paths and device identifiers; a test checks the shipped records.

Citation

If you use this, cite the gates it was evaluated on:

T. Xia, H. Wang, J. Zheng, S. Chen, S. Q. Zhang. EPIC: A System Framework for Efficient Egocentric Perception on Embodied AR Glasses. arXiv:2606.15859, 2026.
A. Paruchuri et al. EgoTrigger: Toward Audio-Driven Image Capture for Human Memory Enhancement in All-Day Energy-Efficient Smart Glasses. arXiv:2508.01915, 2025.
M. Santos-Villafranca et al. Ego-METAS: Egocentric online Multimodal Energy-efficient Temporal Action Segmentation benchmark. arXiv:2606.02246, 2026.