# Sweep38 cleaning animation: verification and scope

The requested cleaning action is implemented in a separate native file: reach, pull inward three times with floor-supported bristles and airborne returns,
then settle to the adopted hold. Duration8s,480frames at60fps, closed endpoint481.
New Hailuo3r03 informs pull/recovery timing and floor-directed attention. All158
source frames plus1x/0.5x playback were reviewed. It is agent-selected, not yet
owner-approved. Its generated anatomy, foot stance and tool proportions are excluded.
Sweep37 is preserved as the provisionally adopted native sweep1.0.0.

Native: `sweep38-cleaning.blend`,221243895bytes.
SHA256:`1506208bafef6e9a4b97712071d4af8d7190e5c2fc9ae512e301639b80b73370`.
Adopted37 SHA256:`9dfbaed464f1ea440d0443d96724ef9b9ec2d19a54ed6171d90e6ee18a4a558f`.
Originalr03 SHA256:`5304bbedaf7c540efe6734f4c142bee7be19f5633731e1c95d0c39e8b5520e15`.

## Motion and construction

Ten custom-property curves use C1 shape-preserving Hermite interpolation with a
shared tangent per knot, represented directly by editable native Bezier handles.
The independent score and native curves agree at961times per channel. The loaded
pulls are0.82–2.08,3.02–4.28 and5.28–6.55seconds, closely corresponding to the
qualitative source brackets. The score moves Stroke1→0.08 while floor-supported,
then lifts the tips0.075units for each return. Working posture stays above0.85
through the active pull/return interval. Balance gives the original free arm18degrees maximum shoulder swing
and6degrees raise; local arm construction is unchanged. Source duration is6.583s;
settling6.55..8s is authored, not generated source footage.

The original Body core is placed rigidly. Each retained180 upper thigh and198
lower elliptical/quadratic profile is reconstructed from its own moved hip and
fixed currentG ankle before the original bounded composition. The old height warp
is absent. Original arm material length and sections and the F2 broom are retained.

## Evidence

- Independent coupled profile domain:35190part evaluations,
  1035time samples and17Body-contribution levels. Segment length ratio
  0.967723–1.029732.
  Positive ellipse/segment and the original centerline acceleration bound pass.
- Actual native:13key-event,2control and2reopened probes.
  Detailed cap tests examine all cap vertices with independent containment rays.
  Source manifold/self-intersection checks cover hold, first work and recovery.
- Full action:480output times plus236half-frame probes.
  Maximum deviation from the independent constructive geometry model:
  3.351264e-06native units. This measures preservation
  of the original shape construction, not pixel-fitting error against the video.
  Fixed-paw maximum errors L0,R0.
  Broom floor/lift support error<=1.15483999e-07native units.
- Every full-action probe checks actual cap/paw triangle non-intersection and
  enclosed-side witnesses. Total reported crossings:0.
  Existing medial arm embedding remains; maximum increased composition depth
  0.000254858518is below the inherited0.0007remesh tolerance.
- Visual review:2880whole-character viewframes,144orbit images,
  39pilot viewframes,24body-only views and9control views.
  All six480-frame movies fully decode at60fps. Browser results are recorded
  separately in PUBLICATION.md after local and deployed playback checks.
- Source/rest parity, control edit/reset, scripts-disabled reopen and endpoint are
  exact. No external libraries or unpacked image dependencies are needed to play.

The render snapshots preserve exact evaluated geometry, face/material assignments,
shader attributes and evaluated material driver values. A total of
459distinct complete render states occur. Pixels are reused
only when that complete digest matches, while the actual native is evaluated and
checked at every output time. There is no mesh interpolation between rendered poses.

## Limits

The native channel curves are C1; that does not prove vertex-correspondence
continuity of the changing composed surface topology. Actual full-sequence views
are reviewed in addition to the mathematical and geometric predicates.

The inherited Blender object-driver positive near-zero snap is accounted for
explicitly. 5input instances are recorded. Ideal raw residuals
remain in the evidence; checks against evaluated driver inputs stay strict:
position<=1.77953848e-07,
orientation<=9.64043263e-08.
No general source-shape or contact tolerance is relaxed.

This is a held-broom, planted-feet candidate. Global physical balance, arbitrary
independent controls, stepping, tool pickup/release, bone/runtime transport and
full15issue/11gate qualification remain open. Main-agent implementation and review
under the active instruction; no independent sign-off or owner adoption is claimed.
See evidence/qualification-ledger.json for inherited open issues and scoped evidence.

Earlier source/native originals and adopted releases are preserved. Verified
Windows working replicas retired for capacity remain backed up on the Mac with
hash receipts; restore those inputs before rebuilding their historical scripts.