# Shoulder movement v0.10 — sampled deltoid correction

10 September 2026. [Open the prototype](index.html). The full-body explorer remains [musculoskeletal v0.6](../musculoskeletal-v06/index.html). This regional `Blender/shoulder-motion-v10` is separate from the original retained `implementation-v10` app.

## What is now in the browser

The linked shoulder retains 83 structures, including both cuffs, three deltoid portions per side, teres major, shoulder bones/labra and fixed thoracic context. The anterior and middle deltoid portions on both sides now use a sampled correction through the raise/rotation controls. An on/off switch compares the correction with the original linear deformation. Other muscles retain the earlier rig. The four measured stray humerus triangles are removed from two small geometry overrides; all other source position buffers and source tissue assignments are preserved. Earlier versions and the full v0.5 master remain unchanged.

Selection, labels, wheel/button zoom, two-axis rotation, source-pose reset, hiding, origin/insertion receiving-bone references and the existing two-stage separation are retained. On the corrected muscles, picking, labels, bounds and separation use their current posed geometry. Tendon faces share their muscle's display transform and retain their material identities. Exact attachment footprints remain disabled; source-position outlines are not footprints.

## Blender correction field

`build_deltoid_motion_grid_v10.py` starts from the repaired v0.9 topology candidate. It applies preserve-volume skinning to four anterior/middle deltoid structures. Where the resulting mesh loses volume, it adds a small volume/contact correction while fixing all tendon-associated vertices and provisional attachment seeds against **additional corrective displacement**. This does not mean those vertices stay fixed in world space: they still follow the posed rig. The earlier 98% activation switch was removed so the corrective follows a continuous volume deficit rather than switching abruptly during motion.

The grid contains 15 poses: raise progression 0/25/50/75/100% crossed with rotation adjustment -25/0/+25 degrees. Eight independently solved intermediate poses, at raise 12.5/37.5/62.5/87.5% and adjustment ±12.5 degrees, were withheld from the grid to check interpolation. This gives 92 muscle-pose checks across both sides. The numeric controls remain authored development parameters, not physiological ranges or a measured shoulder rhythm.

The binary field stores world-coordinate deltas added after the original linear deformation. The browser interpolates those deltas and updates only eight corrected draw parts when a pose or the comparison switch changes. It recomputes their normals and bounds. Camera movement reuses these buffers. Other parts continue using GPU bone transforms. The saved `shoulder-motion-base-v10.blend` contains the repaired rig with targeted preserve-volume settings; the extra corrective field is external and is **not automatically applied inside that Blender scene**.

## What the evidence proves

`motion-grid-review.json` records all grid and withheld solves. The tested corrected source-volume ratios range from 0.99833 to 1.04976; some poses still expand by nearly 5%. No additional more-than-1-mm inside-humerus vertices were introduced relative to the starting deformation in these four muscles, and interpolation introduced no additional tested inside vertices relative to the independently solved poses. The test uses vertex samples and three-ray majority parity against each repaired closed humerus. It does not check all triangle intersections, scapula/clavicle contacts, neighbouring muscles or physiological tissue pressure. Existing source/rig overlap remains; there is no collision-free or anatomical-acceptance claim.

At the withheld poses, the maximum interpolated-position difference from the independently solved corrective was 0.551 mm. This is an engineering approximation error, not an anatomical tolerance. The browser vertex mapping was also checked against both source coordinates and source rig weights; maximum weight difference was zero. All source position buffers in the humerus overrides are unchanged; exactly two measured Bone triangles per side are omitted.

`qa-v10.json` checks nine browser poses, including non-grid control values, against the exported correction data across all eight corrected draw parts. Maximum buffer-coordinate difference was below 0.00027 mm. This checks export/application consistency, separately from the larger withheld-pose approximation error above. The correction switch reproduces the earlier linear result when off. GPU picking on a posed left deltoid, posed separation framing, neutral reset, camera-buffer reuse and 390-pixel layout passed. No page or HTTP errors. Desktop/isolated/mobile screenshots were inspected. Warm corrective updates in that local run took approximately 1.3–3.5 ms; this is processing time for these parts, not a frame-rate guarantee across devices or the whole body.

## Still unfinished

The change is visually modest and the model remains coarse. Shoulder movement is authored, not biomechanically calibrated. The remaining cuff/deltoid/teres shapes need their own deformation review; the current correction covers only four muscle structures. Missing cuff tendon boundaries, verified fibre directions, anatomical attachment areas, wider contacts and muscle wrapping remain unresolved. Source volume preservation does not prove correct activation, force, shape or pennation.

The full musculoskeletal goal, abdominal/lat corrections, other joints and cross-joint muscles, lesson/practice parity and the ten supplied final visual references remain in scope. None is declared complete by these regional checks. Continue wider contacts/attachments and the next regional rig before broad visual polishing. Skills Locker accounts/details remain future integration for Mitch; no purchases or paid services were introduced. The anatomy repository remains local without a configured remote.

## Reproduce and preserve

Run `tools/blender/build_deltoid_motion_grid_v10.py`, then `export_shoulder_motion_browser_v10.py` through the installed local Blender helper. Run `tools/build_shoulder_motion_browser_v10.py` and `tools/qa/qa-shoulder-motion-v10.cjs`; set `FIT_ANATOMY_QA_OUTPUT` for QA artifacts. Dependencies include the v0.5 catalogue/geometry/study data, v0.6 layout/joint presets, v0.8 rig/weights and v0.9 repaired Blender base. Release hashes cover nested geometry overrides as well as the binary field, scripts' outputs and screenshots.

[Source and model attribution](ATTRIBUTION.md) · [Prior deformation evidence](../../Research/2026-09-10/SHOULDER_DEFORMATION_EVIDENCE.md) · [Full completion register](../../Library/MUSCULOSKELETAL_COMPLETION_REGISTER.md)

## 10 September — resumed contact investigation checkpoint

The user authorised continuing until 5% remaining, superseding the historical 10% limit. The v0.10 wider audit found 168 added scapular vertex-contact occurrences across five poses and both sides. Its correction is now OFF by default, retained for comparison. Full-body v0.6 and regional v0.10 remain the available viewers.

The v0.11 candidate is NOT promoted and has no viewer yet. The first attempt added 128 scapular occurrences; its rejected report is preserved. The second build adds parity-based backtracking towards original rig positions, retaining all tendon/seed positions. Latest completed build: added contacts across 23 poses = {'HUMERUS': 0, 'SCAPULA': 0, 'CLAVICLE': 0}; interpolation comparison = {'HUMERUS': 0, 'SCAPULA': 37, 'CLAVICLE': 0}. Read raw reports for scope. Existing baseline overlap, a 1 mm threshold and vertex-only sampling limit these tests; they are not anatomical approval.

Next: review this candidate, independently audit its exported grid against original baseline at intermediate poses (the build interpolation counter compares against the corrected solve), then export, integrate and visually/browser-test only if acceptable. Export and wider-audit scripts are in tools/blender. Do not use prospective success prose in staging record_shoulder_v11.py without actually passing all checks. The Blender base does not apply the external correction grid automatically.

All source originals, connective materials, established learner features, origin/insertion pipeline, unresolved abdominal/lat work and the whole visual goal remain in scope. No paid services or resets. Mitch handles future Skills Locker integration; no remote repository is configured.

