# V7 — complete generated atlas parts

V7 replaces V6's whole-character polygon slices and edge-color patches with
complete independently generated parts. New upper/lower green cloth is one
continuous material, and the left outer coat is continuous from shoulder to
hem. Fourteen new assets plus reviewed inherited head/accessory/tail/background
assets form 30 layers, 32 unique textures and 33 scene image slots.

The LLM authored prompts, inspected the outputs, rejected incorrect parts,
chose anatomical anchors and adapted event-timed deformation. Code extracts
already separate alpha components, verifies exact source pixels/hashes,
constructs contour/edge meshes and executes saved controls. No render-loss
fitting, differentiable rendering, source-video texture cutouts or color-padding
layers are used. The V5 tail layer and pelvis animation remain exactly equal.

## Play without models or original video

Serve the included `model/` directory:

```sh
cd model
python -m http.server 8765 --bind 127.0.0.1
```

Open http://127.0.0.1:8765/. The player includes bones, mesh/feature edges,
brush weights, solo layers, exploded parts and visibility inspection presets.
All runtime assets are local. No image generator or original frames are read.

## Code verification / export from cached assets

From the source root, Python 3.10+, Node.js and Chromium/Google Chrome:

```sh
python -m venv .venv
.venv/bin/pip install -r experiments/atlas_v7/requirements.txt
.venv/bin/python -m unittest discover -s tests -p 'test_*v7.py'
.venv/bin/python experiments/atlas_v7/build.py
.venv/bin/python experiments/atlas_v7/controls.py
.venv/bin/python experiments/atlas_v7/export.py
.venv/bin/python experiments/atlas_v7/render.py
```

The archive includes the frozen atlas registry and inherited provenance
manifest. Build/export validate those cached assets. Running `prepare.py`
instead re-extracts parts and collects the original reference hashes; it needs
the original workspace and reference frames. `audit.py` additionally verifies
the original protected-file snapshot; its geometry/coverage report is included
in `evidence/`. `report.py` needs the saved historical renders and comparison
videos in the original workspace. These are not required to rebuild or play
the final model. Image generation is optional and will not recreate identical
pixels; all nine exact prompts and original generated outputs are retained.

## Files and review boundaries

- `prompts/`, `reviews.json`, `art_reviews.json`: exact generation requests, rejected outputs and
  the reason each selected part was accepted.
  Art-review receipts bind acceptance to the raw image hash and individual
  part. Extraction cannot approve new images by refreshing these receipts.
- `prepare.py`, `assets.py`, `assets.json`: atlas extraction and strict source
  registry. Atlas alpha extraction is allowed; cutting a whole character into
  visible fragments is not. Native component extraction assigns already
  isolated alpha islands without creating colors or extending opacity.
- `placements.json`, `build.py`: source/target anatomical anchor pairs; uniform
  scale plus rotation, draw order and body-specific skinning.
- `controls.py`, `authored_controls.json`: source-event timing and saved local
  controls. V6 optical-flow evidence informs the agent; no automated flow warp.
- `connections.json`, `attachments.py`, `audit.py`: eight attachment groups,
  three authored points each, inspected every half frame; triangle orientation
  every quarter frame; Python/JS geometry comparison for every layer.
- `render.py`: verifies the browser loaded the exact scene snapshot before
  saving final frames and foreground-removal views.
- `export.py`, `bundle.py`, `report.*`: standalone player and evidence page.
  Report packaging compares the complete exported scene, each texture and
  runtime file with the audited source; a stale export is rejected.

Completeness means enough real material for the target animation's occlusions
and attachment ranges. Legs continue below the video viewport; unseen feet
are not reconstructed. Coverage probes are not an exhaustive proof of no
gaps, and zero triangle flips is not a perceptual similarity score. Garment
folds/proportions, main-body trajectory and planar head yaw remain approximate.

No skill was installed; old_stuff and earlier versions are read-only.