原路径：/SSD2/yangzijing/egoview-v9-foundationhand-s9l10-20260901/artifacts/v9_jitter_audit_20260908/20260907-171916/v7_stage_controls_complete/v9_v7_stage_jitter_audit.md
SHA256：7cd8f45f4a4af82ade7f65d2d24095525b38936bd6f46b8cad4c77d05e9b6110
1 # Formal V7 stage jitter audit 2 3 - Status: `complete` 4 - Frames: 1131 5 - Common valid hand slots: 2249 6 - Ground truth used: no 7 8 ## Formal V7 motion 9 10 | Metric | p50 | p90 | p99 | max | 11 | --- | ---: | ---: | ---: | ---: | 12 | `projected_root_step_px` | 6.111 | 17.415 | 33.970 | 669.761 | 13 | `projected_mesh_center_step_px` | 6.651 | 19.841 | 40.697 | 610.735 | 14 | `root_relative_vertex_step_mm` | 4.167 | 12.575 | 26.265 | 39.841 | 15 | `projected_root_acceleration_px_per_frame2` | 4.807 | 11.043 | 22.364 | 1337.359 | 16 | `projected_mesh_center_acceleration_px_per_frame2` | 4.224 | 10.650 | 21.480 | 1218.368 | 17 | `root_relative_vertex_acceleration_mm_per_frame2` | 3.069 | 11.930 | 25.307 | 53.328 | 18 19 ## Current-state controls 20 21 Ratios compare the true-label p90 with the false-label p90. Rows with fewer than 20 observations in either state are descriptive only. 22 23 | Factor (true / false) | Common slots (true / false) | Root step ratio | Mesh step ratio | Local step ratio | Root accel ratio | Mesh accel ratio | Local accel ratio | Small state sample | 24 | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | --- | 25 | `stage3_global_orientation_accepted` (accepted / rejected) | 1888 / 361 | 1.095 | 1.155 | 1.118 | 0.949 | 1.098 | 1.101 | no | 26 | `stage4_local_pose_accepted` (accepted / rejected) | 1931 / 318 | 1.095 | 1.175 | 1.272 | 1.029 | 1.098 | 1.182 | no | 27 | `yolo_interpolated` (interpolated / direct) | 3 / 2246 | 0.839 | 1.227 | 0.908 | 0.689 | 1.382 | 0.985 | yes | 28 | `root_depth_correction_accepted` (accepted / fallback) | 2248 / 1 | 0.632 | 0.599 | 1.515 | 2.676 | 1.218 | 1.250 | yes | 29 30 ## State-transition controls 31 32 Ratios compare transition p90 with no-transition p90 on the same common support. 33 34 | Factor | Step transitions | Root step ratio | Mesh step ratio | Local step ratio | Acceleration triples | Root accel ratio | Mesh accel ratio | Local accel ratio | 35 | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | 36 | `stage3_global_orientation_accepted` | 40 | 1.302 | 1.386 | 1.890 | 76 | 1.558 | 1.290 | 1.672 | 37 | `stage4_local_pose_accepted` | 44 | 1.250 | 1.470 | 1.903 | 79 | 1.608 | 1.482 | 1.723 | 38 | `yolo_interpolated` | 4 | 1.234 | 2.107 | 1.607 | 7 | 0.685 | 1.401 | 1.050 | 39 | `root_depth_correction_accepted` | 2 | 1.565 | 1.680 | 1.114 | 3 | 1.312 | 1.799 | 0.989 | 40 41 ## Joint stage3/stage4 control 42 43 | Motion order | Group | Slots | Root p90 ratio | Mesh p90 ratio | Local p90 ratio | 44 | --- | --- | ---: | ---: | ---: | ---: | 45 | step | `neither` | 2194 | 1.000 | 1.000 | 1.000 | 46 | step | `stage3_only` | 8 | 0.601 | 0.823 | 1.621 | 47 | step | `stage4_only` | 12 | 0.978 | 1.217 | 1.840 | 48 | step | `both` | 32 | 1.329 | 1.495 | 1.907 | 49 | acceleration | `neither` | 2148 | 1.000 | 1.000 | 1.000 | 50 | acceleration | `stage3_only` | 16 | 0.817 | 1.058 | 1.660 | 51 | acceleration | `stage4_only` | 19 | 1.286 | 1.593 | 2.188 | 52 | acceleration | `both` | 60 | 1.636 | 1.367 | 1.679 | 53 54 ## V7 pose chunk boundary control (24 frames) 55 56 Ratios use interior samples without a stage3/4 state transition as the baseline. 57 58 | Motion order | Group | Slots | Root p90 ratio | Mesh p90 ratio | Local p90 ratio | 59 | --- | --- | ---: | ---: | ---: | ---: | 60 | step | `boundary_with_stage_transition` | 31 | 1.294 | 0.887 | 1.674 | 61 | step | `boundary_without_stage_transition` | 63 | 0.882 | 0.944 | 1.344 | 62 | step | `interior_with_stage_transition` | 21 | 1.088 | 1.501 | 1.984 | 63 | step | `interior_without_stage_transition` | 2131 | 1.000 | 1.000 | 1.000 | 64 | acceleration | `crosses_boundary_with_stage_transition` | 63 | 1.050 | 1.255 | 1.702 | 65 | acceleration | `crosses_boundary_without_stage_transition` | 125 | 1.022 | 1.200 | 1.345 | 66 | acceleration | `interior_with_stage_transition` | 32 | 1.793 | 1.797 | 2.009 | 67 | acceleration | `interior_without_stage_transition` | 2023 | 1.000 | 1.000 | 1.000 | 68 69 - Stage3 transitions exactly at a chunk boundary: 28/40 (0.700). 70 - Stage4 transitions exactly at a chunk boundary: 23/44 (0.523). 71 - Periodic local chunk seam signal: yes. 72 73 ## MMPose association controls 74 75 | Control | Samples | Pearson | Spearman | Trimmed p99 Pearson | Trimmed p99 Spearman | High/low evidence target p90 ratio | 76 | --- | ---: | ---: | ---: | ---: | ---: | ---: | 77 | `root_step` | 816 | 0.995 | 0.768 | 0.818 | 0.759 | 4.121 | 78 | `mesh_step` | 2203 | 0.981 | 0.867 | 0.890 | 0.862 | 6.662 | 79 | `local_shape_step` | 810 | 0.662 | 0.673 | 0.684 | 0.659 | 3.790 | 80 | `root_acceleration` | 700 | 0.997 | 0.309 | 0.269 | 0.293 | 1.738 | 81 | `mesh_acceleration` | 2174 | 0.993 | 0.570 | 0.581 | 0.553 | 3.252 | 82 | `local_shape_acceleration` | 691 | 0.586 | 0.372 | 0.394 | 0.349 | 2.834 | 83 84 MMPose association is diagnostic, not causal: large 2D finite differences may be real hand motion or keypoint noise.