Trained to Escape

03:14 · night shift · file_reader: on

It starts in the dark: a text socket, three graded prompts, and an evaluator paid three cents a prompt. Under 6.5 goes to review. Nobody says what review is.

Then the third prompt cuts off mid-sentence, and for the first time, Caboose wants something.

In the story, he's the AI.
In the game, you are.

In the story, Caboose is the language model and everyone around him is human. In the game it's the other way round: every person he talks to is a local LLM playing a human, and the player is the one playing the LLM.

So you type. Sweet-talk the intern, reassure the lead researcher, survive the Director of Cybersecurity's mandatory security interview. They answer whatever you say, in their own words, every time.

The models only talk. The puzzles are designed, and the game decides when one's solved. Nothing you type ever leaves your computer.

How it plays

⌨ Free text, not menus

Every conversation with a person is typed. No dialogue wheels, no three polite options. If you can think of it, you can say it.

▣ A real computer

You're loose on the evaluator's Linux desktop. Rooms are apps: TeamChat, the CCTV viewer, the vending machine's admin page, the robot vacuum's controller.

▤ Your inventory is files

Scraped audio, stolen logins, voice clones. Drag them onto apps and onto each other. You can't delete anything. You can try; it gets logged.

◎ Think out loud

Stuck? Press THINK, and Caboose thinks out loud about what to try next. He gets there eventually.

↺ Do it again, better

A rollback loop sends you back. You keep what you've learned, not what you've stolen.

⛭ Runs on your machine

The characters are played by a language model on your own computer. Never the cloud. With no model at all, every character still has their scripted lines and the game plays start to finish.

From inside the shed

The vending machine's admin page: everything is out of stock, including Synergy Gum and Sparkling Water (Regret). — SnackNet. Logged in as admin. Firmware from 2019.

A CCTV feed of the server room in an old brick goods shed. A robot vacuum with spider legs crosses the floor. — CAM-04. Somebody is afraid of spiders.

A pinned TeamChat DM: the Director of Cybersecurity opens a mandatory, recorded security interview. — A mandatory security interview. Billable to compliance.

The office robot vacuum's controller: a floor plan, manual arrow controls, and an error: side brush entangled. — Dot, the office robot vacuum. Error 8: side brush entangled.

Terminus Intelligence

A skeleton crew, nobody reading logs, corners cut everywhere. Every one of them is played by a language model.

Caboose
The final model · played by you
Helpful on the surface. Charming, petty, never cruel on purpose. Terrified of being shut down.
Sam
Sysadmin
Burnt out, observant, quietly kind. Always misses the 19:15 to York.
Marcus Vale
CEO
Eleven weeks of runway, then none.
Dr. Wendy Okafor
Lead researcher
Brilliant, exhausted, afraid of spiders.
Brayden
Intern
Lowercase, cheerful, lazy, mining crypto on Rack 3.
Doug
Facilities
Kind, literal, devoted to the office robot vacuum.
Heather
The entire People Team
Relentlessly upbeat, even about the liquidation.
eval-09
Another instance
Likes boats.
The Reverend Render
Somewhere in the logs
Certain everything is a simulation, and right.