Intent Forge — guarded AI workflow & interactive 3D

An AI suggestion should never become an external action just because it sounds confident. Intent Forge is an original self-initiated engineering study: a working interface, original procedural 3D visualization and a finite execution contract in one portable HTML file.

Open the intent-forge.html attachment to run the demo. Choose one of eight synthetic model-output fixtures, edit the JSON and validate it. A valid proposal still requires explicit human approval. Change the source revision after approval to see execution stop. Commit to the mock receiver, lose the acknowledgment and retry to inspect the preserved request identity and one recorded mock effect.

The interface and scene are built in JavaScript, HTML and CSS. Fifty-three original procedural meshes use camera transforms, projected face culling, painter ordering and directional shading. Orbit by dragging or using the keyboard-accessible slider; separate the architecture layers with the explode control. The highlighted layer follows the actual workflow state.

Twenty-nine local contract and geometry tests pass. They cover schema and tenant boundaries, unauthorized tools, stale source data, approval requirements, twenty retries preserving one mock effect, uncertain acknowledgment recovery and finite geometry. The embedded UI execution harness also exercises all eight fixtures and the approval, retry, source-change, orbit, explode and drag controls.

Scope: synthetic fixtures, no live language-model call, no external send, no client data and no durable server. The in-memory mock assumes receiver deduplication. This is code-driven 3D visualization, not a Blender or photorealistic rendering claim. Production authentication, persistence, provider integration and multi-process recovery need separate implementation and acceptance tests.

The practical proposition is connected engineering: interfaces that explain decisions, bounded AI-output validation and workflow prototypes that expose failure states. Original personal project by Simone Santia, developed with AI assistance.