Grasp-SFT v1 — the 3-leg sim100 eval chain

One protocol, three legs, seeds 0–99 each (paired spawns, triple-keyed noise): flow + token heads of grasp_sft_v1_joint_8xa100 at step 500 (run2_step500, box download) and the step-3000 endpoint token head under the serving fix b779ba4 (the 0/100 endpoint token read was an inference-collator bug — per-item dataset quantiles instead of the merged training table). euler-10 / --serve-head ar greedy, execute-horizon 30, 30 s episodes · run 2026-08-17 06:0x–14:17:56Z on the local H100, ~6.2/12 GPU-h · verdict: the collapse is flow-specific and present from step 500 — the token head sits near its 500-step level at both ends (16→14) while flow never leaves the floor (4→5); the v1 normalization table mis-fit poisons the flow targets, not the shared trunk.

4/100

step500 flow

16/100

step500 token

14/100

endpoint token-fixed

5/100

endpoint flow (anchor)

0.69 cm

leg-3 median progress

54/100

leg-3 moved >0.5 cm

Against the anchors

Per-seed outcomes

Clips

leg 1 — step 500, flow head (euler-10)

fastest success — seed 8, final 3.9 cm, success at tick 84

median success — seed 40, final 3.2 cm, success at tick 504

nearest miss — seed 59, final 4.5 cm

leg 2 — step 500, token head (_arhead greedy)

fastest success — seed 12, final 4.0 cm, success at tick 82

median success — seed 42, final 3.9 cm, success at tick 110

nearest miss — seed 53, final 5.8 cm

leg 3 — step 3000, token head + serving fix (b779ba4)

fastest success — seed 55, final 3.9 cm, success at tick 144

median success — seed 22, final 3.5 cm, success at tick 224

nearest miss — seed 89, final 3.6 cm

Per-seed table (prog = spawn − best-point distance, cm)

seed	spawn cm	step500 flow			step500 token			endpoint token-fixed
seed	spawn cm	final	prog	succ	final	prog	succ	final	prog	succ
0	10.5	11.7	4.2	—	16.4	0.6	—	15.8	1.2	—
1	7.3	10.4	2.4	—	9.7	0.7	—	2.5	4.9	✓ @ 848
2	10.2	10.2	0.0	—	14.4	0.0	—	11.3	0.1	—
3	10.6	6.8	9.1	—	15.1	0.1	—	16.6	0.4	—
4	10.3	12.7	3.0	—	11.0	2.5	—	13.4	0.0	—
5	8.6	6.2	3.1	—	6.8	2.8	—	10.4	0.0	—
6	10.1	13.7	0.0	—	9.3	0.8	—	9.9	0.5	—
7	7.8	14.1	1.4	—	16.9	0.0	—	12.7	0.0	—
8	7.1	3.9	3.2	✓ @ 84	11.9	0.0	—	12.7	0.3	—
9	11.0	15.8	0.4	—	6.9	4.2	—	12.8	0.4	—
10	11.6	14.4	0.0	—	10.6	1.0	—	13.6	0.2	—
11	9.4	8.5	4.1	—	16.0	0.4	—	15.9	0.1	—
12	7.3	7.5	0.1	—	4.0	3.3	✓ @ 82	0.1	7.2	✓ @ 447
13	8.6	7.4	2.0	—	7.9	4.1	—	14.9	1.7	—
14	10.6	6.7	6.2	—	9.5	3.1	—	4.0	6.6	✓ @ 184
15	8.3	11.7	0.5	—	9.5	3.2	—	11.2	0.0	—
16	9.7	14.5	3.6	—	13.1	0.0	—	12.0	0.4	—
17	11.4	14.0	0.0	—	11.6	1.1	—	10.3	1.2	—
18	8.3	11.0	0.1	—	13.0	0.0	—	13.3	0.2	—
19	7.4	12.7	0.4	—	7.4	0.0	—	8.5	0.7	—
20	9.4	10.7	0.2	—	15.0	0.0	—	12.9	2.8	—
21	9.4	9.8	0.0	—	4.0	5.8	✓ @ 122	15.9	2.2	—
22	10.6	11.4	4.6	—	10.1	0.5	—	3.5	7.1	✓ @ 224
23	9.0	9.5	3.1	—	3.5	5.5	✓ @ 120	3.9	5.1	✓ @ 814
24	9.7	8.9	2.6	—	14.3	0.0	—	8.4	1.3	—
25	11.6	6.8	7.1	—	13.3	0.0	—	4.4	7.8	—
26	10.5	12.0	2.9	—	13.0	0.0	—	10.2	0.4	—
27	10.5	10.3	0.1	—	13.3	0.0	—	7.9	2.5	—
28	8.5	9.1	1.1	—	3.5	4.9	✓ @ 151	8.9	2.2	—
29	9.5	2.2	7.3	✓ @ 118	15.0	1.2	—	9.0	0.5	—
30	9.6	11.1	0.0	—	4.0	5.6	✓ @ 471	15.3	1.2	—
31	12.0	12.8	6.8	—	15.9	0.0	—	16.5	0.5	—
32	9.0	15.2	0.0	—	13.0	0.7	—	7.1	8.2	—
33	9.0	6.4	3.1	—	13.7	0.0	—	10.1	0.7	—
34	8.0	7.8	0.2	—	13.5	1.0	—	12.6	0.2	—
35	9.4	17.0	0.1	—	11.8	0.1	—	14.7	0.3	—
36	9.8	9.8	0.0	—	18.7	0.0	—	14.2	0.0	—
37	9.0	9.3	0.0	—	10.1	0.0	—	10.0	0.2	—
38	10.4	11.4	3.6	—	12.0	0.1	—	10.8	6.9	—
39	9.2	9.2	0.0	—	14.8	0.0	—	4.0	5.2	✓ @ 217
40	8.9	3.2	5.6	✓ @ 504	22.9	1.7	—	13.3	0.2	—
41	9.3	11.6	0.4	—	5.8	3.7	—	19.4	0.0	—
42	10.1	6.4	3.8	—	3.9	6.2	✓ @ 110	24.2	2.7	—
43	11.6	14.5	0.0	—	14.9	0.1	—	20.2	0.7	—
44	10.5	10.5	0.0	—	4.0	6.5	✓ @ 146	6.7	5.0	—
45	9.3	7.5	2.7	—	3.9	5.4	✓ @ 105	13.5	0.0	—
46	11.9	11.9	1.1	—	11.1	0.8	—	8.8	7.5	—
47	8.7	16.7	0.1	—	8.7	0.0	—	11.7	0.2	—
48	8.8	10.2	1.0	—	8.6	0.7	—	6.9	2.3	—
49	8.8	9.2	3.0	—	10.0	0.0	—	7.2	1.8	—
50	8.5	12.0	1.7	—	6.8	1.8	—	10.7	1.2	—
51	11.1	11.4	0.1	—	9.3	1.8	—	3.7	7.4	✓ @ 245
52	9.2	7.2	3.0	—	12.2	0.4	—	9.5	0.0	—
53	8.6	8.8	0.0	—	5.8	3.0	—	6.0	2.6	—
54	9.2	4.0	5.5	✓ @ 519	10.3	0.3	—	10.3	0.0	—
55	8.5	12.3	1.5	—	11.0	1.5	—	3.9	4.5	✓ @ 144
56	11.8	11.1	5.5	—	11.7	0.5	—	12.6	0.2	—
57	9.8	14.6	2.5	—	26.7	1.5	—	7.0	5.2	—
58	7.5	7.5	1.9	—	2.9	4.6	✓ @ 136	4.1	3.4	—
59	8.0	4.5	3.6	—	9.4	0.0	—	10.6	2.0	—
60	11.4	11.8	0.0	—	10.5	0.9	—	3.9	7.5	✓ @ 703
61	8.0	8.6	0.0	—	8.2	0.0	—	4.7	3.8	—
62	8.4	14.2	1.4	—	12.0	0.0	—	14.1	0.0	—
63	7.2	10.7	0.0	—	3.9	3.3	✓ @ 94	11.3	0.0	—
64	10.2	9.5	1.9	—	9.0	1.3	—	18.4	0.2	—
65	9.4	12.6	0.4	—	13.2	0.0	—	13.1	0.0	—
66	9.5	15.6	1.5	—	10.6	0.0	—	12.3	0.0	—
67	10.9	11.0	0.4	—	8.6	2.4	—	18.2	1.1	—
68	7.8	8.7	2.8	—	7.5	0.3	—	9.7	0.0	—
69	9.3	9.2	0.1	—	16.3	0.0	—	15.4	0.0	—
70	8.3	9.1	2.6	—	13.2	0.0	—	10.4	0.0	—
71	9.1	9.2	0.1	—	9.1	0.0	—	15.9	0.7	—
72	12.1	5.0	7.2	—	6.2	7.6	—	4.0	8.1	✓ @ 217
73	10.2	6.1	6.5	—	5.8	4.4	—	12.0	0.0	—
74	10.1	10.8	1.2	—	3.2	6.8	✓ @ 98	8.9	3.8	—
75	10.8	8.0	8.0	—	16.1	0.1	—	4.9	7.0	—
76	7.9	10.0	1.0	—	3.4	4.6	✓ @ 103	11.1	0.0	—
77	9.6	11.0	2.5	—	11.3	2.6	—	4.0	5.6	✓ @ 173
78	7.4	6.8	0.6	—	8.4	0.0	—	13.3	1.0	—
79	12.0	24.2	0.1	—	8.9	3.1	—	8.5	3.6	—
80	11.8	13.6	0.0	—	18.2	0.1	—	13.8	0.3	—
81	10.6	11.7	0.0	—	10.6	0.0	—	16.4	0.2	—
82	9.1	16.5	0.0	—	10.9	0.0	—	4.0	5.1	✓ @ 157
83	11.3	10.0	5.5	—	8.5	4.2	—	15.0	0.3	—
84	8.0	6.1	2.4	—	3.6	4.5	✓ @ 102	13.9	0.0	—
85	9.6	10.8	0.0	—	14.0	0.0	—	17.2	0.0	—
86	10.9	7.4	4.5	—	11.4	2.1	—	11.3	0.1	—
87	8.2	11.6	0.0	—	2.9	5.2	✓ @ 93	10.2	1.0	—
88	8.4	4.9	4.2	—	8.4	0.0	—	8.2	0.6	—
89	8.2	11.8	1.2	—	11.0	0.1	—	3.6	4.6	—
90	8.4	8.7	0.9	—	8.1	0.4	—	16.0	0.3	—
91	7.4	8.5	1.5	—	2.7	4.7	✓ @ 411	3.1	4.3	✓ @ 173
92	11.3	24.3	0.4	—	6.3	5.5	—	14.0	0.1	—
93	10.3	8.6	2.2	—	7.9	3.2	—	17.8	0.2	—
94	8.0	8.0	0.4	—	12.7	0.9	—	5.8	6.8	—
95	7.6	7.6	0.0	—	3.9	3.6	✓ @ 82	3.0	4.6	✓ @ 305
96	8.7	8.5	2.2	—	14.9	0.0	—	7.9	1.2	—
97	9.9	7.4	4.4	—	6.1	4.1	—	10.3	0.0	—
98	8.9	13.0	0.0	—	6.0	3.8	—	8.7	2.1	—
99	9.1	9.7	0.8	—	8.6	0.8	—	14.4	0.3	—

Regenerate: fontaine/scripts/sft_v1_chain_report.py from the banked leg JSONs only. Frozen summary: analysis__sft_v1_chain.json.