idle

last completed: 2026-06-10 — yam_overview.svg v2 — replaced row 3
volume_construction panel with a 4-panel horizontal strip of matched-viewpoint
3D renders (same elev=22, azim=-55, same axis limits, same stride-5
subsample of 17 576 display voxels):
  (a) volume points — gray opacity-only, no colour (raw structure)
  (b) PCA of pre-MLP features (80-d concat [F_scene | F_wrist | γ(z)])
  (c) PCA of post-MLP features (32-d after the 3-layer fusion MLP)
  (d) argmax trajectory (T=32 viridis dots + lines + red ☆ EE start)
      against the volume points at very low alpha as background
Dashed red border on panel (d) signals "this is the action output".

Row 4 (FEATURE VOLUME · fusion mechanism) simplified — dropped the z-slice PCA
panels since (b) and (c) now show the same pre→post effect in 3D. The row's
right side now just has a pointer back to row 3 panels b/c saying "same MLP
applied to all voxels · PCA colour change pre→post = what the MLP learned".

Row 5 (SCORING) also simplified — dropped the standalone prob_volume_3d image
since the argmax trajectory is now panel (d) above. Right side of row 5 has an
"argmax_t · per t" callout that points back to row 3 panel (d) in red.

Canvas now 1500 × 2090 (was 2160). Slightly tighter.

deliverables:
- /data/cameron/para/paper/figs/svg/yam_overview.svg                          (3.0 MB)
- /data/cameron/para/paper/figs/generated/yam_overview.png                    (3000×4180, 2× DPR)
- /data/cameron/agents_stuff/agents/figure_maker/outbox_yam_overview_v2.png   (preview)
- /data/cameron/penpot/build_yam_overview_diagram.py                          (SVG builder, updated)
- /data/cameron/para/paper/figs/data/yam_overview/extract_panels.py           (added 4 matched-viewpoint renders)
- /data/cameron/para/paper/figs/data/yam_overview/vol_points.png              (new)
- /data/cameron/para/paper/figs/data/yam_overview/vol_pre_pca.png             (new)
- /data/cameron/para/paper/figs/data/yam_overview/vol_post_pca.png            (new)
- /data/cameron/para/paper/figs/data/yam_overview/vol_argmax_traj.png         (new)

v2 rough edges flagged:
- pre-MLP PCA panel (b) shows mostly purple/magenta uniformity in the central
  cuboid — the height_emb component dominates the top PCs. Could fit PCA on
  just the (scene_feat | wrist_feat) 64 dims to surface image-feature structure
  instead of height structure. Let me know if you want that variant.
- post-MLP PCA panel (c) shows a clear yellow blob where the cup is — visible
  difference vs (b) which is the main story. Stride 5 is dense enough that the
  yellow blob reads clearly.
- panel (d) argmax trajectory column is thin/vertical because the model's
  predicted reach is mostly a downward Z motion at this frame; this is real
  behaviour but reads as a small streak in the 3D panel. Could pick a frame
  with more XY motion if you want a chunkier trajectory.

prior:
- 2026-06-10 yam_overview.svg v1 — initial 8-row figure with volume_construction
  panel that was too cluttered (replaced this turn)
- 2026-06-10 model_overview_pixelnerf.svg v1 — original mockup predecessor