jev-tetris

# JEV Tetris

A Tetris played by a model that cannot write, cannot see the board, and has
never been told what Tetris is.

JEV answers typed questions. You give it the possible answers and it returns a
probability for each one. It cannot generate text, so it cannot be asked "where
should this go" in the usual sense.

So the code does the geometry. Every frame it searches for all the places the
falling piece can be manoeuvred into — moving, rotating and dropping in any
order, so tucks and spins count — and describes each one in a sentence:

> Traps no new space underneath. Clears one line. Flattens the surface out. The
> stack stays low. It goes at the centre, lying flat.

Those sentences go to JEV as the options of a multiple-choice question. It sends
back a probability for every one, in about 380ms. The code plays whichever it
ranked highest.

Because you get a probability for all of them and not just a winner, the board
can show the whole answer at once. Before the piece drops, every cell lights up
by how much of the model's preference rests on it, with a white ring on the
placement it took. Then it commits and the field goes out.

Settled blocks are dimmed by how confident the model was when it placed them, so
the wall it builds is a readable record of its own uncertainty.



## Run it

Needs Node 22 or later. There is nothing to install: no dependencies, no build
step. Two fonts load from Google Fonts at runtime.

Get a key from [typesafe.ai](https://typesafe.ai), then:

```bash
git clone https://github.com/thelau/jev-tetris && cd jev-tetris
cp .env.example .env         # put your TYPESAFE_API_KEY in it
npm run dev                  # -> http://localhost:5173
```

Open it and press START. With no key there is no START button, and the page says
why instead of failing.