# thespianjas `thespianjas` is the global Jeffrey digital-twin subsystem for the AC monorepo. It is not owned by one podcast or campaign. Podcast, Reel, lecture, piece, and live-performance pipelines consume versioned assets from here. ## Architecture 1. **Identity** — canonical studio references resolve from `papers/jeffrey-platter` through `identity.json`. 2. **Reconstruction** — provider adapters turn those references into a textured, relightable GLB. Meshy 6 multi-image is the first adapter; Tripo H3.1 and SAM 3D Body are named benchmark candidates. 3. **Studio** — `metal/` is the primary native macOS renderer. Its SceneKit scene graph is explicitly backed by Metal; key, fill, rim, ground, camera, skeletal playback, and later facial controls remain ours to direct. The browser studio is only a quick compatibility viewer. 4. **Performance** — podcast audio drives energy, breathing, head emphasis, and eventually facial blendshapes. A video lipsync provider may polish exported shots, but it is downstream of the canonical relightable twin. 5. **Compositing** — `marketing/essay-reels` adds word-timed captions and social safe-area layout to a rendered thespianjas performance. The canonical slug is globally unique: `thespianjas`. Generated binaries are large and versioned locally under `assets/versions/vNNN/`; manifests and recipes are tracked in git. ## First build ```fish node thespianjas/bin/generate.mjs --version v001 --provider meshy cd thespianjas/metal && swift run ThespianJas ../assets/versions/v001/idle.glb ``` Meshy accepts one to four views and returns a textured PBR GLB. The initial studio platter has a strong face view and two seated body views; a dedicated front/side/back neutral-pose capture will materially improve the next rig. The generated `rigged.glb` and `idle.glb` contain the humanoid skeleton. They do not promise facial ARKit blendshapes; viseme/blendshape authoring is a separate performance stage. ## Provider notes (July 2026) - **Meshy 6 multi-image:** 1–4 views, PBR textures, A/T pose options, humanoid auto-rig, GLB/FBX/USDZ/OBJ. Best first complete asset lane. - **Tripo H3.1:** useful single-image benchmark for likeness/geometry. - **SAM 3D Body:** inexpensive human-specific reconstruction benchmark. - **Kling / Sync lipsync:** video finishing layers, not substitutes for the relightable GLB. They accept rendered video or a portrait plus audio. Identity outputs are publishable only in Jeffrey-operated AC surfaces. Never silently substitute another person's photographs or train from private imagery.