Actual engine stills. Metal path tracing for Bistro and Flight Helmet; baked probes for Harbor.
01 / Metal path tracing
Bistro
The public Bistro exterior in sunlight, with multi-bounce lighting. An offline 1024 × 768 render at 512 samples per pixel, a 12-bounce limit, and textures capped at 512 px.
World-space probes light the ship and cargo meshes in a fixed sailing pose. The ocean keeps its existing sky, sun, and reflection lighting.
Baked probesMesh diffuse GIFixed pose
Earlier recording: sailing through MCP. A scripted pilot tests four steering forks, chooses an action, and takes reachable cargo aboard. Developer-mode automation · 2 crates collected · not an autonomous LLM player.
03 / Metal path tracing
Materials, up close.
The CC0 Flight Helmet: 94,722 triangles, six materials, and authored color, normal, and roughness maps. This offline Metal path-traced still shows its materials under multi-bounce lighting.
DeepSeek Flash reads a 30 m radar and sends game intentions. The earlier native replay matches a real model run, including every action and final world hash.
Real model callsPlayer gatewayHash-verified replay
Six frozen microtasks, three fresh trials each. Two repairs and four gameplay features.
63
Native behavioral assertions
Positive, negative and boundary cases, plus types, determinism, fork, replay and reload.
15/18
Clean agent completions
Three provider disconnects happened after valid patches. Their artifacts and interrupted status stay in the data.
Small gameplay tasks, through the engine API.
Collection radius · parcel value scoring · collection victory · dash cooldown · once-only milestones · armored damage and defeat.
DeepSeek Flash inspects, writes and applies scripts using native developer MCP tools. The grader and reference source stay outside the candidate project.
One experimental task set and one model. Read the raw calls, final patches, elapsed time, token usage and cost upper bounds in the full report.
Apple M5 · Metal · 300 measured frames after 60 fully loaded warm-up frames.
High-detail mesh stress
Render and wait for GPU idle — lower is better
Mean wall time per completed frame
64 helmets 6.06M triangles
10.52 ms
256 helmets 24.25M triangles
28.81 ms
1,024 helmets 97.00M triangles
82.43 ms
5 October 2026 · forward renderer · 1920 × 1080 · authored PBR maps, directional shadows and postprocessing enabled. Static rendering without simulation, presentation or PNG encoding. Geometry counts precede culling; this is not a full-game FPS claim or a PT measurement. Method, p95 and raw results.
The short version
A small technology map.
Simulation
Host
Rust · Bevy ECS
Rules
TypeScript · QuickJS-ng · oxc
Physics
Rapier · enhanced determinism
Presentation
Renderer
wgpu · WGSL
Targets
Metal · Vulkan · WebGPU
Editor
React · dockview · Monaco
Interfaces
People
Web editor · CLI
Agents
CLI · MCP
Tools
HTTP · WebSocket · JSON Schema
Start with a running sample
Open a sample. Learn the API.
Build the host, open a sample, and learn the scripting and command surfaces.