Grasp-SFT route C — joint step 2000, flow head on training seeds
Checkpoint fontaine_grasp_sft_joint_corrected/step_002000 (the unseen-leg page's checkpoint, flow head served on the training band) · euler-10, execute-horizon 30, seeds 1000–1099 — the stage-B collection band the 313-demo SFT corpus was drawn from (64 kept / 36 collector-rejected among these 100 scenes) · leg 2 of the probe chain, finished 09:47Z 08-16, ~1.4 GPU-h (registered A §4: unseen 0–99, then train band 1000–1099) · headline: 42/100 ≈ the unseen sibling's 44/100 — no memorization signature; kept 29/64 vs rejected 13/36, CIs overlap (split section below)
42/100
successes (≤3 cm, held)
9 / 44
anchors: base / unseen sibling
75/100
moved the boat >0.5 cm
3.86 cm
mean progress toward disk
0
reset strikes
Against the anchors
Memorization split: kept vs collector-rejected scenes
Membership from the banked stage-B collect curve's kept_seeds: a kept training seed contributed a demo to the 313-demo SFT corpus; a rejected seed's scene was attempted but the scripted collector failed there, so the model never saw it (and those scenes skew harder — the rejection reason confounds the arms' gap). The decisive read is kept vs unseen: 29/64 (45%) on scenes the model trained on vs 44/100 on scenes it never saw — no memorization signature; the checkpoint generalizes rather than replays. The kept−rejected gap (+9 pp) sits well inside the overlapping Wilson CIs.
29/64
kept arm (45%, CI 34–57%)
13/36
rejected arm (36%, CI 22–52%)
+9 pp
kept − rejected rate gap (CIs overlap)
44/100
unseen sibling — the anchor the kept arm must beat to claim memorization
4.45 vs 2.82 cm
mean progress, kept vs rejected
Per-seed outcomes
Clips
fastest success (kept arm) — seed 1027, final 4.0 cm, success at tick 90median success (kept arm) — seed 1017, final 2.5 cm, success at tick 135nearest miss (kept arm) — seed 1008, final 1.5 cmfastest success (rejected arm) — seed 1092, final 4.0 cm, success at tick 106nearest miss (rejected arm) — seed 1073, final 6.0 cm
Per-seed table
seed
spawn cm
min cm
final cm
progress cm
success
1000
8.9
2.4
2.4
6.5
✓ @ 126
1001
11.6
3.4
3.4
8.2
✓ @ 138
1002
9.9
4.8
7.8
5.1
—
1003
10.4
1.9
1.9
8.5
✓ @ 133
1004
11.2
1.7
1.7
9.5
—
1005
7.6
5.7
5.9
1.8
—
1006
8.8
2.4
2.4
6.4
✓ @ 209
1007
8.1
1.5
1.5
6.5
✓ @ 117
1008
7.9
1.4
1.5
6.5
—
1009
10.2
10.1
10.1
0.1
—
1010
9.3
8.9
8.9
0.4
—
1011
9.0
1.7
2.2
7.3
—
1012
8.0
8.0
8.0
0.0
—
1013
11.7
10.9
12.4
0.7
—
1014
8.9
1.1
1.2
7.8
✓ @ 639
1015
8.8
7.8
8.5
0.9
—
1016
9.5
8.7
8.8
0.9
—
1017
9.0
2.5
2.5
6.5
✓ @ 135
1018
9.7
8.2
8.3
1.4
—
1019
9.8
9.3
9.4
0.4
—
1020
10.4
2.0
2.0
8.4
✓ @ 138
1021
11.1
1.4
1.4
9.7
✓ @ 136
1022
7.8
1.5
1.5
6.3
✓ @ 117
1023
7.4
6.8
7.3
0.7
—
1024
10.3
3.9
3.9
6.4
✓ @ 119
1025
9.4
4.0
4.0
5.4
✓ @ 127
1026
8.9
1.8
2.5
7.1
—
1027
9.3
4.0
4.0
5.3
✓ @ 90
1028
10.4
3.4
3.4
7.0
✓ @ 147
1029
8.7
5.8
9.8
2.8
—
1030
11.4
11.4
11.4
0.0
—
1031
7.7
6.4
7.5
1.3
—
1032
11.2
11.2
12.6
0.0
—
1033
8.4
8.4
9.1
0.0
—
1034
11.8
11.1
11.1
0.8
—
1035
9.8
3.5
3.5
6.3
✓ @ 650
1036
7.9
7.7
7.7
0.2
—
1037
10.1
2.3
2.3
7.8
✓ @ 111
1038
11.1
4.8
6.0
6.2
—
1039
10.2
2.1
2.1
8.0
✓ @ 138
1040
8.8
8.8
8.8
0.0
—
1041
10.6
10.6
10.6
0.0
—
1042
9.8
1.9
1.9
7.9
✓ @ 128
1043
11.1
4.0
4.0
7.1
✓ @ 146
1044
8.0
4.0
4.0
4.0
✓ @ 111
1045
10.9
9.9
13.2
1.0
—
1046
8.1
8.1
9.7
0.0
—
1047
9.7
3.6
3.6
6.2
✓ @ 287
1048
11.4
2.2
2.2
9.1
✓ @ 137
1049
9.1
2.2
2.5
6.9
—
1050
10.2
2.3
2.3
7.9
✓ @ 122
1051
11.1
9.8
9.8
1.3
—
1052
9.3
2.6
2.6
6.7
✓ @ 137
1053
10.8
10.8
10.8
0.0
—
1054
8.2
5.5
5.6
2.7
—
1055
11.6
9.5
10.4
2.1
—
1056
11.8
2.6
2.6
9.2
✓ @ 123
1057
9.3
3.2
3.2
6.2
✓ @ 208
1058
11.0
1.4
1.8
9.6
✓ @ 236
1059
10.1
2.5
2.5
7.5
✓ @ 128
1060
12.1
11.1
11.1
1.1
—
1061
9.3
9.0
9.0
0.2
—
1062
11.0
11.0
11.0
0.0
—
1063
9.2
4.8
10.0
4.4
—
1064
9.6
1.7
1.7
7.9
✓ @ 134
1065
11.7
11.7
11.7
0.0
—
1066
8.5
2.0
2.0
6.6
✓ @ 128
1067
8.5
3.0
3.0
5.4
✓ @ 125
1068
8.8
8.2
15.9
0.6
—
1069
8.9
3.9
3.9
4.9
✓ @ 152
1070
7.2
6.5
6.7
0.6
—
1071
7.6
7.0
7.0
0.6
—
1072
8.5
7.6
19.4
0.9
—
1073
7.3
5.8
6.0
1.6
—
1074
11.1
2.4
2.4
8.7
✓ @ 135
1075
7.9
7.8
8.4
0.1
—
1076
10.5
2.3
2.3
8.2
✓ @ 346
1077
12.1
11.7
11.7
0.4
—
1078
11.6
9.5
16.6
2.1
—
1079
8.6
3.7
3.7
4.9
✓ @ 120
1080
8.8
2.2
2.2
6.6
✓ @ 125
1081
7.4
7.3
7.7
0.0
—
1082
10.3
10.3
10.3
0.0
—
1083
7.9
6.6
6.6
1.3
—
1084
11.9
4.0
4.0
7.9
✓ @ 149
1085
10.2
9.7
9.7
0.5
—
1086
11.5
11.5
11.8
0.0
—
1087
11.2
10.9
10.9
0.4
—
1088
11.5
11.4
12.0
0.0
—
1089
8.1
6.8
6.8
1.3
—
1090
11.0
11.0
11.2
0.0
—
1091
11.0
3.6
3.6
7.4
✓ @ 102
1092
10.0
4.0
4.0
6.1
✓ @ 106
1093
9.6
9.4
9.4
0.2
—
1094
10.9
10.2
10.2
0.8
—
1095
10.2
1.5
3.3
8.7
✓ @ 836
1096
8.8
8.8
8.8
0.0
—
1097
9.2
6.9
7.3
2.3
—
1098
9.9
2.2
2.2
7.7
✓ @ 140
1099
9.3
3.9
3.9
5.4
✓ @ 111
Regenerate:
fontaine/scripts/grasp_sft_joint_unseen_report.py
from the banked flow_train.json only.
