wdobry / laya-playground

# Laya playground
A website, three games, a benchmark and an agent skill for Laya, the open-source decision model: typed questions in, real probabilities out, one forward pass, no generated text. It runs on your own machine.
The live site is at **brainfunctioncollapse.com/laya**. It has no model behind it and replays recorded runs. Clone this repository and the same pages run against the real model.
Laya is created by Nandakishor M of Convai Innovations. This repository is not his and does not contain the model: it installs his package and downloads his open weights.
## Run it
 [code] 
Then open . The first start downloads 2.3 GB of open weights and takes about 90 seconds to load them. Apple silicon, NVIDIA or plain CPU.
You get the whole site, live:
- `/` the landing page, where the model plays Flappy, a lane runner and Tetris itself, about 30 decisions a second
- `/playground` the editor: write some text and a few typed questions, see every answer with its probabilities, compare all three checkpoints
- `/about` why this exists
The web server itself is the Python standard library and binds to loopback, so nothing is reachable from your network. The only dependency is `laya`.
### Without the model
Any static file server shows the recorded version, exactly as the public site does:
 [code] 
Then open . The games replay real recorded decisions and say so; the playground answers its presets from recordings.
## The agent skill
One file teaches a coding agent to add Laya to a project properly: the API, questions that work, thresholds, calibration, and the traps found building this site.
 [code] 
Claude Code reads `.claude/skills` natively. Any agent that accepts a markdown instruction file can use the file as it is.
## The benchmark
Laya and TypeSafe AI's hosted Jev, on the same 500 labelled examples with the same questions and no tuning. Jev is more accurate out of the box. Laya matches it on simple questions, answers several times faster from a laptop, is free, and is yours to fine-tune. The numbers on the site are rendered from `static/data/versus.json`, never typed by hand.
To reproduce it:
 [code] 
Two of the source datasets restrict redistribution, so the sampled texts and the raw API responses are not in this repository. `eval/provenance.json` records where every example comes from. The harness only measures: nothing from Jev is ever fed into Laya.
## What is in here
| Path | What it is |
| --- | --- |
| `server.py`, `poc.py` | the local model server and the proof of concept it grew from |
| `index.html`, `about.html`, `playground.html` | the three pages |
| `static/` | styles, scripts, and the recorded data the public site replays |
| `static/demos/` | the three games: Flappy, a lane runner and Tetris. Each one describes its situation in a sentence and asks one typed question |
| `skills/laya-integration/SKILL.md` | the agent skill |
| `eval/` | the benchmark: dataset builder, tasks, runner |
| `tools/` | recorders for the replays, and small si