git404hub

what is duck-harness fr?

tufalabs/duck-harness — explained in plain English

Analysis updated 2026-05-18

58HTMLAudience · researcherComplexity · 4/5Setup · moderate

tl;dr

The winning solution to ARC-AGI-3's first milestone: a tool-using AI puzzle solver, shared with its full benchmark run and an interactive viewer.

vibe map

mindmap
  root((The Duck))
    What it does
      ARC-AGI-3 puzzle solver
      Tool using approach
      Won Milestone 1
    Parts
      Inference runner
      TAAF framework
      Example run and viewer
      Kaggle notebook
    Viewing
      Base Python deps only
      No GPU needed
    Running
      Needs GPU
      vLLM and PyTorch
      Local server or OpenRouter

Code map

Detail Auto

An interactive map of this repo's files and how they connect — its source is parsed live in your browser. Click Visualize to build it.

filefunction / class

what do people make with this?

VIBE 1

Browse an interactive viewer of a complete ARC-AGI-3 competition run without any GPU setup.

VIBE 2

Study a winning tool-using approach to an AI reasoning benchmark competition.

VIBE 3

Reproduce the full solver run on Kaggle hardware using the included notebook.

VIBE 4

Build on the TAAF framework to run your own ARC-AGI-3 experiments.

what's the stack?

PythonvLLMPyTorchuvOpenRouter

how it stacks up fr

tufalabs/duck-harnessliangdabiao/weekend-city-triprevfactory/font-harness
Stars585858
LanguageHTMLHTMLHTML
Setup difficultymoderatemoderateeasy
Complexity4/53/53/5
Audienceresearchergeneraldesigner

Figures from each repo's GitHub metadata at analysis time.

how do i run it?

Difficulty · moderate time til it works · 30min

Viewing an example run needs only base Python deps in seconds, but actually running the solver needs a GPU and multi-gigabyte dependencies like vLLM and PyTorch.

The README does not state a license.

in plain english

The Duck is a research project that shares the winning solution to the first milestone of ARC-AGI-3, a competition that tests whether AI systems can reason and adapt to solve new visual puzzle games they were never specifically trained on. This repository contains a full, self-contained writeup of that solution: the code for the solver itself, one complete run of it against the benchmark, and an interactive viewer for looking through that run in detail. The solver, called The Duck, plays these puzzle games by using tools rather than by direct end-to-end prediction, and it runs on top of a framework the team built called TAAF, short for the Tufa ARC-AGI Framework. The README points to a technical writeup on Kaggle, a blog post comparing the approach to others, and both a podcast episode and a YouTube video where the team explains the solution further. The repository is organized into a few clear parts: the actual inference runner with its prompts, run results, scoring code, and viewer, the TAAF framework it runs on, which provides the benchmark and game interaction logic, a complete example run covering 25 official games with 20 attempts each, and a Kaggle notebook that can reproduce a full run from scratch. If you just want to look through the example run without doing anything computationally heavy, you only need basic Python tooling. The instructions have you sync a small set of base dependencies with a package manager called uv, then start a local viewer that opens the bundled example run in a browser, or point it at a different run folder if you have one. Actually running the solver yourself is a bigger step: it needs a GPU and a much larger set of dependencies, including vLLM and PyTorch, and it plays games either through a locally hosted vLLM server or through OpenRouter, an API that provides access to different AI models. The included Kaggle notebook reproduces a full competition run on Kaggle's own hardware, which is the easiest path to actually training or evaluating the full solution.

prompts (copy fr)

prompt 1
Help me set up uv and view the example ARC-AGI-3 run in this repository.
prompt 2
Explain in simple terms what ARC-AGI-3 tests and why a tool-using solver approach helps.
prompt 3
Walk me through what I would need to actually run The Duck's solver myself, including the GPU requirements.
prompt 4
Explain what the TAAF framework provides versus what The Duck solver itself does.
prompt 5
Summarize how the Kaggle notebook reproduces a full competition run.

Frequently asked questions

what is duck-harness fr?

The winning solution to ARC-AGI-3's first milestone: a tool-using AI puzzle solver, shared with its full benchmark run and an interactive viewer.

What language is duck-harness written in?

Mainly HTML. The stack also includes Python, vLLM, PyTorch.

What license does duck-harness use?

The README does not state a license.

How hard is duck-harness to set up?

Setup difficulty is rated moderate, with roughly 30min to a first successful run.

Who is duck-harness for?

Mainly researcher.

peek the repo → explain another one

This repo across BitVibe Labs

double-check against the repo, no cap.