# Windows native720 demo

The existing local backend and frontend are preserved. This Space is static; its hosted page cannot use a visitor's localhost as a public GPU server.

On the prepared demo PC, first check the running service rather than launch another GPU model:

```powershell
Invoke-RestMethod http://127.0.0.1:8002/health
```

Verify model Overworld/Waypoint-1.5-1B, model_loaded true. If running, use the existing frontend at http://127.0.0.1:3000. If startup is needed, in terminal 1:

```powershell
Set-Location 'C:\Users\filma\Documents\Morph\Claude\backend'
powershell -ExecutionPolicy Bypass -File .\experiments\Start-LiveMemoryBackend.ps1
```

This existing launcher uses the prepared Codex Plus virtual environment and local model cache. It refuses an occupied 8002 port. Do not use -Restart on an active world. For another PC, adapt the Python path parameter and populate the model cache using the pinned revisions in MODEL_MANIFEST.json. The launcher uses offline mode; it cannot download missing weights.

In terminal 2:

```powershell
Set-Location 'C:\Users\filma\Documents\Morph\Claude\frontend'
$env:ENGINE_TARGET='http://127.0.0.1:8002'
$env:VITE_WORLD_API_BASE_URL='http://127.0.0.1:8002'
npm ci
npx vite --port 3000 --strictPort --host 127.0.0.1
```

Skip npm ci on the already running prepared machine. Keep the regular Brave window foregrounded, choose Create Anything, generate a short first-person scene description, inspect the image, then Start Exploring. Confirm LIVE and native720. Use WASD and pointer-lock when available, otherwise drag-look. Show pause, resume and reset. Keep movement gentle; do not promise drift-free long exploration.

Fallback: choose a previously generated saved image in the same local application. Reuse its theme and metadata; do not label a saved image as fresh generation. If the GPU service fails, show the Space's explicitly recorded evidence and the separate deterministic Rescue Mission. Never silently use mock video.

Rescue Mission and AI Showdown are separate from GPU exploration. Show the score, stars and A* comparison after a mission. Show existing agent evaluation data without claiming new training results. Physical keyboard hold validation should be performed by the presenter before judging; it is not newly verified by this packaging task.

Troubleshooting: a localhost server is private to that PC. Static Spaces do not host the inference backend. Verify GPU service health, port conflicts, LIVE label and scene preview. A low canvas FPS may be browser scheduling even with healthy GPU throughput. Shared-lock image enhancement can pause generation. Restore a clean saved scene when necessary and disclose the transition.