Grasp-SFT v2 joint pdnorm — endpoint, flow head on unseen seeds
Checkpoint grasp_sft_v2_joint_1gpu_pdnorm/step_003000 (per-dataset flow norm, mixed v2 corpus; launched 2026-08-18 11:02Z, run complete 00:1xZ 08-19, ~12.9 GPU-h) · euler-10, execute-horizon 30, seeds 0–99, 30 s episodes; sim leg wears the sim demos' row (frozen serving-row rule) · frozen decision grid: ≤10 broken-class band / 11–19 ambiguous band / ≥20 exonerates the mix · panel anchors, wear-corrected class: 27.40 re-worn disc-1000 / 25.89 released pre-SFT (measured 08-18, at the null) / 25.15 repo-midpoint null / 8.37 state-copy (real bar) · verdict: 1/100 ≤ 10 — CONVICT: the mix is the prime suspect (read taken 03:17:39Z 08-19; panel endpoint 29.18 native / 27.44 truth-fit)
Panel anchor ladder
Wear-corrected class (wear audit 08-18) — the endpoint reads against this class, never the raw 58.14. The endpoint session re-runs pdnorm_panel_ladder_chart.py --endpoint <row> on GO, so the embedded sidecar pdnorm_panel_ladder.b64 carries the stamped magenta rung.
Estimator seam (pdnorm_endpoint_truthfit_rewear.py, output-side re-expression): endpoint native 29.18 → truth-fit 27.44 (seam +1.74); truth-fit ladder: disc-1000 27.40 / released 27.14 / repo-midpoint null 25.15. The native row stays the headline (a served rig wears recorded tables); the truth-fit row reads the ladder like for like.
1/100
successes (≤3 cm, held)
9 / 44 / 11
anchors: base / probe / disc1000 baseline
38/100
moved the boat >0.5 cm
0.84 cm
mean progress toward disk
0
reset strikes
Against the anchors
Paired read: pdnorm_endpoint3000 vs disc1000_demosonly
recorded non-gating read (rides alongside the frozen absolute bands) · pre-reg posts/2026-08-xx-prereg-grasp-sft-v2-joint-pdnorm.md · paired on the 100 shared unseen seeds · reset strikes 0/0. The demosonly baseline's 11/100 sits inside the pdnorm draft's own 11–19 ambiguous band, so the paired read rides alongside the frozen absolute bands — it is recorded, never gating.
-10
success-count delta, 1 vs 11 (CI95 [-16, -5], excludes 0)
p = 2.0e-03
McNemar exact, 0 vs 10 discordant seeds
-3.49 cm
mean progress delta (CI95 [-4.6808, -2.3506], excludes 0)
29%
per-seed progress win rate (0% ties)
Per-seed outcomes
Clips
fastest success — seed 29, final 3.9 cm, success at tick 247nearest miss — seed 30, final 5.6 cmmiss #2 — seed 61, final 6.5 cmmiss #3 — seed 32, final 6.8 cm
Per-seed table
seed
spawn cm
min cm
final cm
progress cm
success
0
10.5
10.5
25.5
0.0
—
1
7.3
6.3
23.5
1.0
—
2
10.2
10.0
11.0
0.2
—
3
10.6
9.7
13.4
0.9
—
4
10.3
10.3
12.2
0.0
—
5
8.6
8.6
15.3
0.0
—
6
10.1
9.6
10.1
0.5
—
7
7.8
7.6
19.0
0.1
—
8
7.1
7.0
20.1
0.1
—
9
11.0
4.1
8.3
6.9
—
10
11.6
11.2
13.9
0.4
—
11
9.4
7.3
7.3
2.1
—
12
7.3
7.1
21.6
0.1
—
13
8.6
8.6
12.0
0.0
—
14
10.6
9.5
9.9
1.1
—
15
8.3
8.3
10.4
0.0
—
16
9.7
6.4
10.5
3.3
—
17
11.4
8.4
8.5
3.0
—
18
8.3
7.8
13.8
0.4
—
19
7.4
7.2
19.3
0.2
—
20
9.4
9.3
13.4
0.1
—
21
9.4
8.4
12.5
1.0
—
22
10.6
10.6
15.9
0.0
—
23
9.0
9.0
11.5
0.0
—
24
9.7
9.7
20.7
0.0
—
25
11.6
11.6
14.2
0.0
—
26
10.5
10.4
10.4
0.1
—
27
10.5
10.5
27.7
0.0
—
28
8.5
7.6
8.0
0.9
—
29
9.5
3.9
3.9
5.5
✓ @ 247
30
9.6
5.2
5.6
4.4
—
31
12.0
11.4
19.0
0.6
—
32
9.0
6.7
6.8
2.4
—
33
9.0
7.7
7.7
1.3
—
34
8.0
7.1
7.3
0.9
—
35
9.4
8.7
8.7
0.8
—
36
9.8
7.9
8.3
1.8
—
37
9.0
8.2
8.2
0.8
—
38
10.4
10.4
13.8
0.0
—
39
9.2
7.5
12.8
1.7
—
40
8.9
8.7
8.7
0.2
—
41
9.3
9.3
10.4
0.0
—
42
10.1
10.1
12.3
0.0
—
43
11.6
7.7
10.6
3.8
—
44
10.5
10.3
10.4
0.2
—
45
9.3
9.2
12.3
0.1
—
46
11.9
9.3
9.4
2.6
—
47
8.7
8.7
11.7
0.0
—
48
8.8
8.8
14.2
0.0
—
49
8.8
8.1
8.4
0.7
—
50
8.5
8.4
13.6
0.1
—
51
11.1
8.9
17.2
2.2
—
52
9.2
9.0
11.4
0.1
—
53
8.6
7.1
10.1
1.5
—
54
9.2
9.2
13.7
0.0
—
55
8.5
8.5
24.6
0.0
—
56
11.8
11.5
11.6
0.3
—
57
9.8
9.7
11.3
0.0
—
58
7.5
7.0
25.7
0.6
—
59
8.0
7.7
13.4
0.3
—
60
11.4
9.5
13.1
1.9
—
61
8.0
6.2
6.5
1.8
—
62
8.4
8.4
10.2
0.0
—
63
7.2
7.1
10.3
0.1
—
64
10.2
10.2
10.5
0.0
—
65
9.4
6.6
7.4
2.9
—
66
9.5
8.1
8.1
1.4
—
67
10.9
10.9
21.1
0.0
—
68
7.8
7.4
12.5
0.4
—
69
9.3
9.0
9.2
0.3
—
70
8.3
8.2
9.0
0.0
—
71
9.1
9.1
15.4
0.0
—
72
12.1
12.1
21.4
0.0
—
73
10.2
10.1
14.2
0.1
—
74
10.1
10.1
12.9
0.0
—
75
10.8
10.0
10.9
0.9
—
76
7.9
7.6
22.5
0.3
—
77
9.6
9.6
13.7
0.0
—
78
7.4
7.4
27.9
0.1
—
79
12.0
12.0
15.0
0.0
—
80
11.8
7.2
7.2
4.6
—
81
10.6
9.1
13.0
1.4
—
82
9.1
8.8
8.8
0.3
—
83
11.3
11.3
18.6
0.0
—
84
8.0
8.0
12.5
0.0
—
85
9.6
9.6
14.2
0.1
—
86
10.9
10.8
24.5
0.1
—
87
8.2
8.0
17.6
0.1
—
88
8.4
8.4
8.8
0.0
—
89
8.2
7.8
10.3
0.5
—
90
8.4
7.9
7.9
0.5
—
91
7.4
7.1
9.8
0.3
—
92
11.3
10.5
10.5
0.9
—
93
10.3
8.5
8.5
1.8
—
94
8.0
7.8
15.9
0.2
—
95
7.6
7.1
8.9
0.5
—
96
8.7
7.4
9.7
1.3
—
97
9.9
4.6
11.8
5.3
—
98
8.9
7.8
10.3
1.2
—
99
9.1
9.1
20.6
0.0
—
Regenerate: fontaine/scripts/grasp_sft_joint_unseen_report.py
from the banked flow_unseen.json only.
