Verification
SkillMediaCross-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.
Use Verification in Claude, ChatGPT or Ahel Desktop
Free. Sign in, add Verification and connect your AI. About a minute.
Also: Claude Code · Cursor · Codex
Then ask your AI: use the Verification skill
Details
Instructions available. Your AI can read the instructions. Execution depends on the setup they require.
Account requirements not reviewed. Check the skill instructions before use; Ahel provides instructions and does not run this skill.
No other account needed.
Add Ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.
What this skill tells your AI
The instructions your AI receives, as published by dcc-mcp/dcc-mcp-core in python/dcc_mcp_core/skills/verification/SKILL.md and read by Ahel’s review.
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.
Signals
- GitHub stars
- 51
- Forks
- 4
- Last commit
- Oct 2026
Advanced
- Item type
- skill
- Key
verification-dcc-mcp- Source
- github.com/dcc-mcp/dcc-mcp-core
github.com/dcc-mcp/dcc-mcp-core