laya-web: Laya running entirely in the browser as an 8-bit ONNX model

# laya-web

[Laya](https://huggingface.co/convaiinnovations/laya) — a System One decision model —
running entirely in the browser, as an 8-bit quantized copy of the English checkpoint.

**Live at [laya-web.pages.dev](https://laya-web.pages.dev)** · weights on
[Hugging Face](https://huggingface.co/nvkudva/laya-web-q8)



Laya is not a chat model. It reads a **state** — any text or JSON — scores the **options
you enumerate**, and returns one calibrated probability distribution per question in a
single forward pass. No sampling, no tokens out.

## Features

- **Three question types.** `noul` (true/false), `choice` (pick one named option),
  `score` (expectation over ordered levels).
- **Calibrated probabilities**, not raw softmax: temperatures fitted per question type
  and option count, plus an entropy-based confidence on every answer.
- **Nothing leaves the tab.** onnxruntime-web over WebAssembly — no API key, no backend,
  and no request carrying your state anywhere.
- **524 MB of weights, cached across reloads** in the Cache API. The page shows how much
  is cached and lets you delete it.
- **Verified against the PyTorch reference.** `/parity.html` re-runs both gates in the
  browser: byte-identical token ids, then end-to-end probabilities within 0.02.
- **Five worked presets** — email triage, moderation, routing, review scoring, escalation
  — each one state answered by three questions at once.

## How it works

Every question becomes one sequence, with a `[MASK]` marker in front of each option:

```
[CLS] <type> question: <instructions> [SEP] [MASK] opt0 [MASK] opt1 … [SEP] <state> [SEP]
```

1. **Tokenize** with transformers.js, straight from `tokenizer.json` — no model needed.
2. **Encode** the sequence with the quantized ModernBERT-large encoder (512 tokens).
3. *