{
  "date": "2026-09-28",
  "hypothesis": "Fine semantic parts allow separate motion; aligned local generation should reduce appearance drift compared to coarse packed atlas.",
  "budget": "First one head atlas call; expand only after visual inspection. Ordinary CPU image tools and frozen LPIPS.",
  "factors": [
    "old coarse source layers",
    "fine source partitions with identical inherited deformation",
    "fine generated atlas with identical deformation",
    "agent-authored independent fine-part controls"
  ],
  "fixed": [
    "source frame 0",
    "renderer",
    "tail and body trajectories",
    "evaluation masks",
    "metrics"
  ],
  "selection": "validation hair-region LPIPS primary; reject if character LPIPS worsens >2% versus v1; inspect frame0 texture drift separately",
  "scope": "First head/hood/ears/hair because these currently share one deforming sheet. Body remains control.",
  "test_policy": "Previous test frames/results have been exposed in v1. Do not claim a fresh holdout; use dev/val only in this follow-up.",
  "source_scene": "experiments/scenes/11_midpoint_poses.json",
  "source_scene_sha256": "9916cc8263a95e0b1e237fb96522f5eabc9fb944469e8ed2c860069de6a62d66"
}