Claude Skill

verification

Cross-DCC verification toolset - deterministic capture review views, in-band image statistics, scene-vs-spec validation, and reference-vs-render comparison sheets. Pure Python, read-only, DCC-agnostic. Use to close the quality loop without shipping pixels to a model: flag white/n

LLM Mart · 0 points · 0 views 2 listing impressions 0 install-command copies

#verification

Download dcc-mcp-dcc-mcp-core-python_dcc_mcp_core_skills_verification-8d8c969.zip · 9 KB

Install

skills CLI npx skills add https://github.com/dcc-mcp/dcc-mcp-core/tree/main/python/dcc_mcp_core/skills/verification
Claude Code claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install dcc-mcp-dcc-mcp-core@llmmart
Git git clone https://github.com/dcc-mcp/dcc-mcp-core.git

The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole dcc-mcp/dcc-mcp-core collection as a plugin from our marketplace. Git is the plain clone.

Skill manifest

Verification

A cross-DCC, read-only verification toolset (affinity: any) that closes the capture-and-look quality loop with deterministic, machine-checkable signals. Inspired by img2threejs's "deterministic-first, model-last" review design, it computes evidence in-band — statistics, hashes, and verdicts — and never ships pixels to a model unless a caller explicitly asks for image/both payloads.

The four tools map to the four legs of the deterministic capture contract:

  • capture_review_views — the fixed front / side / top / three-quarter view plan, viewport-only, fails loudly on a missing, wrong-size, or blank frame.
  • image_stats — mean luma, stddev, histogram, uniformity, and bad-frame flags (white playblast, near-black render, missing display transform).
  • validate_scene_vs_spec — hierarchy names, material bindings, per-mesh UV coverage, part existence, non-manifold, and Euler checks, before export.
  • make_comparison_sheet — reference + renders side-by-side into one image; packages evidence and computes no scores.

Pixel-level metrics (pHash / dHash / aHash, silhouette IoU, SSIM, CIEDE2000, Sobel edge maps) live in dcc_mcp_core.verification as deterministic library functions, not skill tools, so callers gate on them directly.

Safety Contract

Every tool is read-only and side-effect free except make_comparison_sheet, which writes exactly one output image at the requested path. Tools never mutate the input captures and never launch a DCC. PPM/PGM inputs decode through the standard library; other formats (PNG/JPEG/EXR) decode through ffmpeg on PATH (or vx ffmpeg) and fail loudly when the decoder is unavailable.

Files (dcc-mcp-core)
  • scripts
    • capture_review_views.py 3.3 KB
      Loading…
    • image_stats.py 1.3 KB
      Loading…
    • make_comparison_sheet.py 2.9 KB
      Loading…
    • validate_scene_vs_spec.py 1.9 KB
      Loading…
    • _common.py 8.1 KB
      Loading…
  • SKILL.md 2.8 KB
    ---
    name: verification
    description: >-
      Cross-DCC verification toolset - deterministic capture review views, in-band
      image statistics, scene-vs-spec validation, and reference-vs-render comparison
      sheets. Pure Python, read-only, DCC-agnostic. Use to close the quality loop
      without shipping pixels to a model: flag white/near-black/gamma-broken frames
      from stats alone and catch missing-UV/missing-part exports before export.
      Not for scoring or aesthetic review - this toolset packages evidence and
      computes no scores.
    license: MIT
    compatibility: Uses ffmpeg/ffprobe on PATH (or via vx) only for non-PPM image decode and comparison-sheet compositing.
    metadata:
      dcc-mcp:
        dcc: python
        recall-context:
          app_type: python
          domain: verification
          workflow_stage: verification
          task_category: verify
        version: "0.1.0"
        layer: infrastructure
        search-hint: "verification, capture, image stats, luma, histogram, comparison sheet, scene spec, uv coverage, blank frame, review views"
        tags: "verification, capture, image-stats, comparison-sheet, scene-spec, uv, dcc-agnostic"
        tools: tools.yaml
    ---
    
    # Verification
    
    A cross-DCC, read-only verification toolset (`affinity: any`) that closes the
    capture-and-look quality loop with deterministic, machine-checkable signals.
    Inspired by img2threejs's "deterministic-first, model-last" review design, it
    computes evidence in-band — statistics, hashes, and verdicts — and never ships
    pixels to a model unless a caller explicitly asks for `image`/`both` payloads.
    
    The four tools map to the four legs of the deterministic capture contract:
    
    - `capture_review_views` — the fixed front / side / top / three-quarter view
      plan, viewport-only, fails loudly on a missing, wrong-size, or blank frame.
    - `image_stats` — mean luma, stddev, histogram, uniformity, and bad-frame flags
      (white playblast, near-black render, missing display transform).
    - `validate_scene_vs_spec` — hierarchy names, material bindings, per-mesh UV
      coverage, part existence, non-manifold, and Euler checks, before export.
    - `make_comparison_sheet` — reference + renders side-by-side into one image;
      packages evidence and computes no scores.
    
    Pixel-level metrics (pHash / dHash / aHash, silhouette IoU, SSIM, CIEDE2000,
    Sobel edge maps) live in `dcc_mcp_core.verification` as deterministic library
    functions, not skill tools, so callers gate on them directly.
    
    ## Safety Contract
    
    Every tool is read-only and side-effect free except `make_comparison_sheet`,
    which writes exactly one output image at the requested path. Tools never mutate
    the input captures and never launch a DCC. PPM/PGM inputs decode through the
    standard library; other formats (PNG/JPEG/EXR) decode through `ffmpeg` on
    `PATH` (or `vx ffmpeg`) and fail loudly when the decoder is unavailable.
    
  • tools.yaml 5 KB
    Loading…

Comments (0)

Sign in to join the conversation.

No comments yet.

Reviews (0)

No reviews yet.