PacmanLocalJev: the unchanged 322M Laya checkpoint plays Pac-Man on Windows with CUDA

# raboija/PacmanLocalJev

**LocalLaya · fast typed decisions vs. Pac-Man**

A native Windows demo of the real **Laya multilingual 322M** model making typed decisions locally, on NVIDIA CUDA or CPU. The game runs beside live model probabilities, executed moves, inference timing and planner intervention counts. No API key or cloud inference.



[Download v0.1.0](https://github.com/raboija/PacmanLocalJev/releases/tag/v0.1.0) · [How it works](docs/ARCHITECTURE.md) · [Raw results](reports/RESULTS.md) · [General API](docs/API.md)

## What this is

Every game move calls the unchanged, general-purpose Laya checkpoint. It returns a direction `choice` and two `noul` boolean estimates. The same model handles arbitrary text questions through a local API; no Pac-Man classifier is trained or substituted.

The gameplay is **planner-guided**: a deterministic search describes possible routes and identifies its preferred move. Laya classifies those descriptions. An execution filter and progress-recovery rule can override the model's proposal, with every override visible in the UI. The model receives structured game information, not pixels.

The work in this repository is the Windows/NVIDIA integration, bounded CUDA graph execution, original game and route adapter, native UI, transparent telemetry and reproducible evaluation. Laya's model and training are upstream work. `PacmanLocalJev` is the project name; this is not an official Jev product, a Jev implementation, or a new general-intelligence model.

## Run on Windows

1. Install **Python 3.11** from [python.org](https://www.python.org/downloads/windows/) and [Microsoft Edge WebView2 Runtime](https://developer.microsoft.com/en-us/microsoft-edge/webview2/).
2. Download and extract the **Windows source ZIP** from [Releases](https://github