Turn the body from its support.
Test internal twist second.

The reference reads as a coordinated body turn toward the non-holding side. Sweep30 adds chest–pelvis torsion, but misses much of the body travel, lower-belly movement and opposite-side sweep.

Analysis only. Original Source29 and Sweep30 A/B are unchanged. The non-holding-foot pivot is the next hypothesis to test; this single front view does not establish a unique hidden pivot or weight distribution.

All three videos share playback time. Source: 30fps; native A/B: 60fps. One frame step advances native time by 1/60s. Overlays: cyan = gray head envelope, pink = nose, green = broom binding, purple = left cream edge, blue = sampled exposed sole. These are image features, not a reconstructed skeleton.

The main mismatch is spatial

The face-presentation cue already moves 113% as much as the reference, while the head-envelope midpoint and broom binding move only 50%. Increasing head or chest yaw alone does not address the missing body transport.

What the comparison did not test

1. A fixed central pivot. Both versions rotate around the same central point. Neither selects the non-holding foot as a support reference or expresses contact-relative body displacement.
2. A chosen 12° internal difference. Chest and pelvis headings were authored, not measured. Quiet lower-body presentation was assumed before proving that common supported motion was insufficient.
3. The same upper motion in A and B. Head, shoulders, arms and broom are equal by construction. That comparison cannot show whether changing body transport would improve the action.
4. Two changes bundled together. A/B toggles both relative yaw and the gray-leg anchoring fade. The existing C pilot isolates relative yaw more closely, but still uses the same pivot.
5. A stroke that misses its destination. The inherited local arm movement stays too small to reach the opposite foot. A better torso field alone cannot fix the shared hand/tool path.
6. Engineering checks were narrower than performance. Smoothness, stationary paw vertices and reopening prove that the rig follows its chosen rules. They do not prove that those rules reproduce the source.

Supported turn is plausible. Exclusive foot loading is unknown.

The non-holding foot's sampled outside edge and sole remain fixed in all 120 source frames, with stable results across three mask cutoffs. The holding sole is also fixed in 108 unoccluded samples; the other 12 are excluded. No step or lift is observed.

Both contacts staying planted does not identify which foot bears the load. A pivot-based rotation and a central rotation plus translation can also produce identical geometry. The useful next test is a contact-relative body mechanism, not a claim to have recovered an invisible joint.

First compare a central common-body case with a non-holding-contact case using the same local arm/head scores. Let the shoulder and tool move with the changed body frame. Only then test a small chest correction. Keep leg accommodation equal when isolating torso twist.

Read the full evidence report · Read the next experiment design · Original Sweep30 rig comparison and editable native

Every source frame, with observable feature ownership

Full source sequence — 120 frames

Source frames 0–29Source frames 30–59Source frames 60–89Source frames 90–119

Enlarged feet and trunk — 120 frames

Feet/trunk frames 0–29Feet/trunk frames 30–59Feet/trunk frames 60–89Feet/trunk frames 90–119

Source and A/B feature key sheets

source · 14 annotated keysA-common · 14 annotated keysB-articulated · 14 annotated keys

Seven full-resolution source overlays

Feature ownership · source f000Feature ownership · source f044Feature ownership · source f052Feature ownership · source f060Feature ownership · source f068Feature ownership · source f078Feature ownership · source f096