# Coordinated body/foot control: bounded mechanical success, artistic review open

The coordinated sequence passes the existing monitor's compiled mechanical checks on all 24 freshly rendered frames. The scripted framewise comparator, given exactly the same expanded pose set and controls, passes 17/24. Its seven failures concern added correction speed. This establishes a narrow benefit of sequence-aware selection in this development case; it is not a claim of novel IK, general semantic editing, or superiority over an expert animator or a commercial system.

| Fresh-render measurement | Framewise | Coordinated | Declared limit |
|---|---:|---:|---:|
| Mechanical frame passes | 17/24 | 24/24 | All compiled gates |
| Provisional support observations within tolerance | 32/32 | 32/32 | 20 mm |
| Worst support-clearance change | 0.006 mm | 15.001 mm | 20 mm |
| Worst added correction speed | 6.084 m/s | 0.432 m/s | 0.5 m/s |
| Added sampled external occlusion | 4.62 percentage points | 4.62 percentage points | 5 percentage points |
| Maximum lower-leg length change | 0.146% | 0.013% | 5% |
| Head/hair/backpack translation residual | 0.0172 mm | 0.0172 mm | 1 mm |
| Fresh sole replay discrepancy | 0 | 0 | 0.1 mm |

The 32 foot/frame observations are temporally correlated, not independent trials.

The coordinated solution trades up to 15 mm of allowed clearance variation for bounded correction motion. It does not dominate the comparator on every measurement. The framewise method preserves the sampled clearance more exactly. Both use the original armature and existing foot IK, with a one-metre camera-right displacement. The coordinated solution also uses the permitted 0.5 m toward-camera depth offset during frames 83–103 and body-height offsets of roughly 10–15 cm. Camera and external object transforms remain unchanged.

## What is substantively different

The prior transfer candidate adjusted root height only. This implementation evaluates the deform-weighted foot regions in GEO-suit, drives left and right existing foot controls independently, measures explicitly approved terrain rather than treating grass as support, checks head and suit external occlusion separately, and solves body/foot correction trajectories across the sequence. Constraints are replayed in fresh Blender render processes and evaluated with the existing monitor-core module.

The scene adapter explicitly identifies the rig, feet, head, suit, and three support surfaces. It is not a learned or fully automatic semantic interpretation. The prior dependency-discovery result is reused for assembly membership; it is not newly claimed as exhaustive.

The coarse search evaluates 2,040 poses. It finds spatially feasible poses at every frame, but no admissible path through the later segment under the motion bound. Its 2.5 cm height spacing itself exceeds the allowed 2.083 cm per-frame correction, and support/swing transitions introduce additional discontinuities.

The recorded post-hoc refinement adds continuous proposals using standard SciPy SLSQP. It jointly solves body height and the two foot correction trajectories. Support targets can vary within a ±1.5 cm corridor inside the original 2 cm clearance limit. Swing targets stay tied to the body offset. All acceptance thresholds stay unchanged. Each proposal is re-evaluated through the actual rig, terrain and visibility; it is not accepted on numerical optimization success alone. Both framewise and sequence selectors receive the same expanded candidate set. `REFINEMENT.md` and the original `selected-coarse.json` make this development step explicit.

## Comparison with the earlier rendered attempt

The previous root-height-only candidate was freshly rescored under this round's foot-mesh, provisional phase and visibility definitions. It passes 15/32 support observations, has 152.59 mm worst support error, and violates the visibility budget in 11 frames (up to 23.08 percentage points additional sampled occlusion). The new coordinated result passes 32/32 and the visibility limit throughout.

These rescored figures differ from the earlier 28/48 bone-clearance result because the measurement and phase definitions changed. Do not compare those denominators directly. The old images and measurements remain untouched; `legacy-replay.json` links the same old images by hash.

## Scope of acceptance and visible review

The source has an authored action discontinuity at frame 83. It remains in every render; temporal checks explicitly restart there. The edit is not claimed to remove that source jump or produce a continuous walking-gait benchmark. Added correction speed excludes the source animation's own motion and is recomputed from fresh rendered-run geometry.

Support phases are provisional: each foot is considered supported when its source clearance is within 2 cm of its segment minimum. Some original clearances are negative, including roughly 6 cm penetration in the sampled source geometry. The experiment preserves those authored relationships within tolerance; it does not prove physically valid ground contact. Maximum additional penetration in the coordinated result is 9.32 mm. The sole proxy averages the lowest 20% of 140 evaluated foot-weighted vertices per foot; this is not complete shoe-surface collision detection.

Visibility uses external geometry rays to 79 head vertices and 65 suit vertices, not a segmentation mask or pixel-level readability model. Transparency, materials, all render-only dependencies, joint limits and full-body collisions are not exhaustively checked. Occlusion may remain significant when it was already present in the source. A 4.62-point increase is within the declared budget, not proof that no readability was lost.

Assistant inspection covered all 24 coordinated PNGs. The visible hair stays attached and facial features remain coherent at preview resolution. The lower body remains partly hidden by foreground terrain, so the pictures alone cannot verify individual foot contacts in every frame. This is a mechanical pass with artistic acceptance open, not an assertion of production-ready animation. All 48 full monitor assessments remain blocked pending artist review.

## Files and verification

- `PROTOCOL.md` is hashed before the coarse search; `REFINEMENT.md` separately records the continuous extension.
- `reports/candidates.json`, `selected-coarse.json`, `continuous-proposals.json`, `continuous-evaluated.json`, `candidates-refined.json` and `selected.json` retain solver evidence, including unsuccessful numerical proposals.
- `reports/render-framewise.json` and `render-coordinated.json` record 48 new PNG hashes, fresh measurements, preserved camera/environment transforms and replay errors.
- `reports/checked-*.json`, `results.json`, and `contracts.json` record final calculations and the independent existing contract evaluator. The complete contract includes the unanswered artist-acceptance check.
- `reports/prior-evidence.json` preserves hashes for the 96 prior transfer renders and its protocol/results. The source `.blend` remains unsaved.
- The three-way video explicitly reuses the exact prior source movie and displays newly rendered framewise and coordinated movies, all 24 frames at 24 fps. Preview settings remain 640×360, Eevee16, subdivision1, embedded scripts disabled. Source shader-override execution and original-look fidelity remain unverified.
- Resource-supervised Blender runs use four threads, a 13 GiB process-tree RSS guard and disk/RAM reserves. The coarse search took about five minutes; the two renders about 87 seconds each. Logs and resource receipts are retained.

The next claim test should freeze this completed method and apply it to an unedited interval or another asset with independently labeled support and visibility requirements. Re-tuning on this same clip does not establish generalization. The framewise comparison here isolates sequence selection and is not a substitute for a competent artist/tool baseline.

Settlers — Blender Studio / Blender Foundation, CC BY 4.0. Original source provenance remains in the existing director_asset_audition_v0_1 manifest; the exact source entrypoint hash is recorded in the experiment reports.