README.md

---
title: Hollow
emoji: 🌫️
colorFrom: gray
colorTo: indigo
sdk: gradio
sdk_version: 6.18.0
app_file: app.py
python_version: "3.11"
pinned: true
short_description: A lost child asks for your memories — then claims them.
tags:
  - track:wood
  - achievement:offgrid
  - achievement:offbrand
  - achievement:fieldnotes
  - build-small-hackathon
  - thousand-token-wood
  - horror
  - npc
  - game
  - small-models
  - qwen3
  - zerogpu
---

# Hollow

*Something shaped like a lost child stands at the far edge of the Thousand Token Wood, where the path ends and the fog begins.*

It has no memories of its own. It asks for yours. It keeps them in a treasure. And as the bond grows, it starts telling them back to you — **in first person, as if it had lived them.**

> ⚠️ Contains flashing images and one sudden sound.

## Watch & read

- 🎥 **Demo video:** https://youtu.be/Awc4g1Y7d6Q
- 📓 **Field notes / write-up:** https://www.linkedin.com/feed/update/urn:li:ugcPost:7472433866822131712/

## How it plays

Talk to it. That's all. But *everything* you do is being measured on two axes:

- **Bond** — how much of yourself you share. The child literally materializes out of the fog as the bond grows: from a smudge, to a faceless shape, to a child with sad eyes, to something that has been listening too well.
- **Tone** — how you *treat* it. It mirrors your writing style. Warmth makes it attach. Coldness makes it withdraw. Cruelty reminds it of how it was treated when it was alive — and it keeps your exact words. You'll see *something* being stored in its treasure, redacted: `▮▮▮▮▮`. You won't know what until it's too late.

**Three endings.** Love it too much and it loves you back — the wrong way. Stay distant and the fog simply gives up on you. Be cruel and you'll learn what it does with the words you gave it. Each ending is a scripted sequence that takes over the interface itself: messages appear on *your* side that you never typed.

## The tech (small on purpose)

- **One model, 8B, no APIs:** [Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) — `transformers` + ZeroGPU on this Space, the same model via Ollama on a consumer 8 GB GPU during development. No cloud calls, no external services: the whole game is one small model and one prompt discipline.
- **One GPU acquisition per turn:** the reply *and* a JSON extraction pass (bond delta, captured memories, tone delta, cruel quotes) run inside a single `@spaces.GPU` call. State is a plain dict in `gr.State`; all game logic (recall scheduling, ending gates, redaction) is deterministic, pure Python, covered by 147 tests.
- **The pacing was tuned empirically:** a simulator plays scripted visitors through the real model and reports per-turn state, so the three ending gates are calibrated against measured behavior, not guesses.
- **Off-brand by hand:** the entire interface is custom CSS over Gradio — no JS. The materialization blur, the terror flashes that cut through every face the child could become, the full-screen screamer, the heartbeat bond meter, the redacted treasure: all CSS animations exploiting Gradio's DOM re-mount on update. Even the horror sting in the bad ending is **synthesized in numpy at boot** — no audio file in the repo.
- **Art:** five portraits generated with FLUX (grayscale, early-1900s storybook horror), embedded as base64 WebP.

## Built for the Build Small Hackathon

Track: 🍄 **An Adventure in Thousand Token Wood** · Badges: 🔌 Off the Grid · 🎨 Off-Brand · 📓 Field Notes

*The recall moment — the child telling you your own memory as its own — needs nothing bigger than 8B. Small models are perfect liars: just big enough to remember you, too small to know it shouldn't.*