Before You Send: live tone, formality and fight-risk readings as you type

# Before You Send

**A message tone checker that runs while you type.** Paste or write a message you're nervous about (a text to an ex, an email to your boss, a Slack reply) and see how it will land, before you hit send.

You get three live readings and a language badge:

| Reading | What it shows | Comes from |
|---|---|---|
| **Tone** | friendly, neutral, passive-aggressive, angry or anxious, with a probability for each | Laya `choice` question |
| **Formality** | a 1 to 5 slider | Laya `score` question |
| **Fight risk** | a thermometer, plus a "Maybe sleep on it? 😬" banner above 60% | Laya `noul` question |

The "Copy & Send" button turns green when the risk is low.

The judging is done by [**Laya**](https://github.com/NandhaKishorM/laya), an open-source *decision model*. It is not a chat model: it never writes text, it picks from options you give it, in about 40 to 110 ms on a laptop. That speed is what makes checking on every pause in typing practical.

> **Status:** works end to end with real Laya. It is a prototype: it misses subtle passive-aggression and its formality slider barely moves (see [Findings](#findings-what-works-and-what-doesnt)).

---

## How it fits together



The browser waits 150 ms after you stop typing, then sends your text to `POST /analyze`. The FastAPI server passes it to Laya (the dashed area) and returns three typed answers, which the page turns into the pill, slider and thermometer. Everything outside the dashed box is plumbing.

---

## The classification layer: how Laya works

This is the part that makes the project different from "ask an LLM to classify it".

### It classifies. It doesn't write.



| | Regular LLM | Laya |
|---|---|---|
| How it answers | Writes tokens one by one, then you parse the text | Reads the input once and s