Why Artifact Forge exists

I wanted an AI fabrication system that could explain what it built, prove that the requested features exist, and admit when it could not build them.

Most AI-to-3D tools generate meshes and hope for the best. Artifact Forge takes the opposite path: a product is a declaration — archetype, parameters, constraints, features — and every build ends with a validation report and an honesty report. Not “exported successfully”, but proof that the requested features were actually built, and an explicit admission of what was not.

Artifact Forge is created by an independent systems architect, creative technologist, musician and maker. It connects software architecture, digital fabrication and product design into one open-core system.

How it was built

The goal was simple to state and hard to do: let someone with zero 3D/CAD modeling skills produce working household parts and small devices — a cable clip that actually holds the bundle, a hose adapter that actually seals, an enclosure whose lid actually snaps.

The engine was developed in close collaboration with AI models — Claude Fable 5, Claude Opus 4.8 and GPT 5.5 — over two major iterations, and the iterations tell the real story:

• Iteration one leaned on the models for everything: describe a part, let the AI produce the geometry. It worked just often enough to be dangerous — a mesh can look right and still be wrong in every way that matters (walls too thin, screws unreachable, a “hook” that is actually a closed ring).
• Iteration two inverted the architecture. The knowledge moved out of the model and into a declarative parametric grammar: typed archetypes with parameters, named regions and contracts, plus a registry of validators that measure every claim on real geometry. LLM involvement dropped to roughly 20% — the model now only translates human intent into a YAML declaration and proposes edits; it never draws geometry, and everything works with the AI switched off.

A few decisions did most of the work, in plain terms:

• Geometry is checked before it exists. Every part is first built as an exact 2D profile with named measurements — no meshes, no CAD kernel. Wall thickness or a mouth opening is read off the profile like a number on a drawing, so most mistakes are caught in milliseconds, before any heavy geometry runs.
• A part cannot lie about its features. A feature counts as “built” only after its validators pass on the measured result — and the data schema literally cannot express “built” for something unsupported. The honesty report says what was requested, what was proven, and what the engine could not do.
• Protected regions. Every archetype declares zones that decoration must never touch — screw seats, flex hinges, contact surfaces. You can perforate, texture or “biomorph” a part, and the engine guarantees the functional zones come out untouched.
• Edits are rebuilds, not surgery. “Make it stronger” or “print it without supports” becomes a typed patch to the declaration; the part is rebuilt from source and the engine verifies that everything you asked to preserve survived — checked, not promised.
• A typo is a loud error, never a silent skip. Every name in a declaration — archetype, feature, check, joint — must exist in a registry at load time. Nothing degrades quietly.

The result is a parametric model of everyday objects: describe the thing you need in a few lines, get geometry with a validation report attached — no CAD skills required.

The core engine is Apache-2.0. Models you generate from the open-core archetypes are yours — print them, sell the prints, remix the YAML. Domain packs plug into the same fail-fast registries under the same honesty rules; the path for your own starts at Packs.

Let's build together

Artifact Forge is open by design, and it grows fastest with more hands on it. If you make things, break things, write validators, or just have an idea for a part the engine should know how to build — I'd love to hear from you. Early contributors shape where this goes, and there's a standing invitation to become a maintainer as the project matures.

Reach out, say hello, propose a collaboration: pinelover2024@gmail.com