# Sweep30: a supported chest/pelvis rig extension

Sweep30 adds separately editable chest and pelvis headings to a new copy of
Sweep25. A broad torso field and gray-leg accommodation connect the turning
chest to planted paws. The seven-pose native pilot and complete continuous
frame review pass with limitations. The portable comparison includes the
unchanged source, five native videos and an editable native download.
This is a candidate, not owner adoption or a full motion-release certificate.

## Result and controlled comparison

The four-second action has one preparation, inward work gesture and relaxed
return. Its event roles follow Sweep29's independent front original. Angles
and in-betweens are hand-authored, not estimated joints or video-coordinate
fitting. The previous arm shape, material length and shoulder/elbow action are
retained deliberately to isolate torso responsibility.

A and B share the exact chest, shoulder, arm, hand, head and broom tasks.
A rotates the body/thigh branch as one rigid frame. B turns the chest above a
quieter pelvis and fades gray lower attachment motion near the fixed paws.
The tail follows the lower frame. The comparison also isolates relative turn
and leg anchoring by disabling each at the work-end pose.

Across all seven keys and four views, B reduces lower-body rotation without
changing the arm task. The belly pigment follows the broad torso twist. The
front and oblique views show the changed relationship between the chest and
supported legs; top view primarily shows head and tail directions because
the large head hides the torso. The sixteen isolated shaded views show no
observed new abrupt waist pinching, fold, crack or exposed attachment gap in
this pilot. Existing simplified gray-leg/paw overlap and arm-root embedding
remain visible limitations. Bright clay highlights limit subtle shading
judgment; oblique/side surfaces were also inspected at their native640px.

The native is intentionally more restrained than the generated source's full
broom excursion. This experiment establishes useful separate rig controls;
it does not claim that lower-body articulation alone reproduces every part of
the source performance. The generated tail silhouette is not adopted as new
anatomy or pigment authority.

## Native verification

Original Sweep25 SHA256:
`1cfc32aa8a05d2d26d8c3602839e8312e855f29dd4f77405b7fefc1a8044dfc6`.
New native `pilot01/sweep30-torso.blend`,220981730bytes, SHA256:
`81f0a7904bc861989eb894c2be5aeca347b9aeec3ca5447870a24cd30eb7d793`.
The complete native is byte-identical on Mac and Windows.

Blender5.2.0LTS, buildfbe6228777e7, evaluates the persistent Geometry Nodes and
ordinary animation/driver controls. Source hashes and observed results are
bound in `pilot01/pilot.json`. The actual pre-extension native lacks the four
required independent-frame controls; that is the recorded RED capability
failure. Build, reopen or disk failures are never counted as RED.

| Check | Observed pilot result |
| --- | --- |
| Actual native time samples | 57 |
| Joined body field versus independent calculation | maximum3.0241e-7 model units |
| Displayed paw drift | 0 |
| Chest-bound shoulder frame error | maximum2.1479e-7 |
| Lower-frame tail error | maximum3.0211e-7 |
| Rigid head assembly error | maximum3.3069e-7 |
| Fixed-grip complete prop error | maximum7.5703e-7 |
| Blade floor/lift error | maximum9.1315e-8 |
| A/B upper-task difference at seven keys | 0 |
| Rest, endpoint, reopen error | 0 |
| Actual control changes and exact resets | 7 before saving,4 after reopening |
| Topology, material slots and rest attributes | exact at rest and work pose |
| Color pilot / isolated shaded views | 64 / 16, all reviewed |

Five pure-logic tests verify zero/bypass behavior, planted lower sections,
the rigid shoulder shelf, section-radius preservation for pure yaw, smooth
field joins and temporal endpoint continuity. An analytic sample of14175
points across the authored score has minimum Jacobian determinant0.6332857
and maximum singular stretch1.389826. These are diagnostics of the specified
field, not an all-surface appearance or full-control-domain certificate.

## Continuous capture and resource recovery

The capture uses the actual live native arm/tool and torso fields on exact
evaluated rest snapshots. Whole-scene comparisons against the full native
protect head members, materials and attachment ownership. It does not use
interpolated poses or a second approximate offline animation implementation.

Pilot FIFO20261004-062929-fac09a85 completedrc0. Initial capture
20261004-063926-74763845 stopped after139complete times when shared Windows
free space reached the5GB floor. All695completed frame images were retained.
The native and render settings were not changed to address that environment
stop. The continuation verifies those hashes and resumes at index139.

To recover capacity,740old decoded reference PNG/JPEG files totaling355628011
bytes were copied to Mac and every hash verified. Only those exact Windows
duplicates were removed. All four provider originals, native files, source
inputs, reports and other sessions' data were preserved. The cleanup receipt
is `cleanup-decoded-result.json`; initial free space for the continuation was
5.549GB. Continuation20261004-065601-c702853b completed with return code0.

All240times were captured at60fps in five series: A front and B front,
oblique, side and top. Twenty comparisons with the complete native scene
have maximum error2.4080e-7 model units; blade floor/lift error over the
capture is at most1.2294e-7. Five videos fully decode to240frames each.
All25transferred videos/sheets match their recorded SHA256 values.

The main agent inspected all1200view frames in20chronological sheets,
including the index139resume boundary. The single sweep and return has
no observed new contour jump, waist crease, open attachment gap or extra
tool. The front A/B difference is deliberately modest: B keeps the lower
gray silhouette more centered over the unchanged paws. Side and oblique
show smooth accommodation of the lower body under the turning chest.
Top view is useful for heading, but the head occludes most torso surfaces.
The unchanged long initial/final holds make the active section short.
See `sequence-review.json` for the bound frame-review record.

## Browser verification

The comparison plays the reference and both native videos together. Actual
UI checks cover1x/0.5x/0.25x, all four B views, linked/unlinked reference,
key-pose seek, axes, end-stop, previous frame and loop wrap. A near-end
play/pause promise race was fixed in the page only; the corrected session
reported no console warnings/errors. The native and captured media did not
change for this UI repair.

## Qualification and scope

`qualification-ledger.json` tracks all15historical architecture issues and
all11gate groups. All remain open at full-release scope. The receipts above
support a bounded, useful sweep-rig experiment; they do not close jump
phases, airborne foot shape, general asymmetric/alternating support,
all-control collision/strain ranges, force balance or target runtime export.

No independent reviewer or external model was used. Geometry checks and
main-agent visual findings are separate. Existing adopted motion registries,
Sweep25 and the original reference videos remain unchanged. Read
RIG_GUIDE.md for the persistent controls and what requires revalidation.