SOLID LIGHT: CO-CREATE

Two minds. Two perspectives. One impossible creation.

Inspiration

Visual-hull reconstruction and shadow puzzles make a beautiful promise: two flat perspectives can define something spatial. We explore a cooperative twist. A person owns the front silhouette; LUMA owns the perpendicular side silhouette. Neither can quietly correct the other's work.

What it does

Draw a ribbon of cyan light. LUMA proposes amber side profiles and tests the resulting voxel structure. A small traveler crosses only when a walkable path exists in that actual solid. In the second puzzle, the human outline stops short. LUMA discovers that no side-view change can supply missing length, highlights the absent columns, and asks the player to extend their drawing. A new human revision triggers a real replan.

How we built it

TypeScript, Meta IWSDK 1.1.0-rc.2, its ECS and shared spatial pointer input, Vite, Three.js re-exported by IWSDK, and the official IWER emulator. A worker intersects two 32×32 masks into 32³ occupancy, evaluates bounded candidate side projections, and runs BFS over standable surfaces with floor thickness, footprint, step and clearance checks. Exposed voxels render in one instanced mesh. Revision checks prevent old computation replacing current input. Assets are procedural and sounds synthesized.

How LUMA actually thinks

LUMA is a deterministic local search planner, not an LLM and not a language model. It observes the current cyan mask, proposes a bounded set of amber side profiles (default budget: 24 candidates / 800 ms), validates walkability, clearance, shelter and light cost, ranks scores, and either commits the best valid solid or requests a human edit for what it cannot change. It cannot see intent and it cannot invent missing length — that is why Incomplete Span is a genuine human repair. The in-game How LUMA thinks panel shows the last decision and why. Every critical puzzle runs without paid inference or a cloud service. Optional OpenAI suggestions, off by default, never replace this local planner.

Challenges and testing

64 unit tests and 15 browser tests cover the planner, walkability checks, layout and input. The demo was captured in the official Quest 3 emulator (IWER) with hand input; on-headset testing is next, so we make no hardware frame-rate claim. A late fix moved LUMA's depth panel so no panels overlap the play area.

A longer co-creation journey

Four connected chapters build on each other: First Bridge, Incomplete Span, Light Garden, and Last Spark. The Garden adds a real canopy constraint. Last Spark adds a strict light budget that can require the human to trim their projection. Four workshop presets vary span, creature footprint and budget. Sturdy/Efficient preferences change tested side profiles without bypassing validation.

Animated loop teaching, immediate strokes, actionable gap feedback, an enlarged depth readout, a thick emissive volume, and recoverable editing support the core puzzle. Optional OpenAI suggestions remain separate from the authoritative local planner. A bounded real provider connectivity test passed; a public inference proxy is not deployed.

Best First Five Minutes

Minutes 0–1: learn the closed loop on The First Bridge and watch a validated crossing. Minutes 1–3: Incomplete Span — LUMA’s solid visibly falls short, you extend the cyan outline with your hands, and a celebratory crossing lands. Minutes 3–5: inspect How LUMA thinks, and in a headset lean to see real depth (desktop orbit is the fallback). Progress is saved for Continue vs New game.