ARC-AGI-3 Kaggle Harness

ARC-AGI-3 Kaggle Harness is my experiments applying world models, LLM-based approaches, neuro-symbolic AI, solvers, and deterministic models to the ARC Prize 2026 ARC-AGI-3 Kaggle competition. It contains local experiment code, scripts to generate the Jupyter notebooks published to Kaggle for public-game runs on GPU, and the submission pipelines for both public and private (hidden) game runs.

The experiments went through a number of iterations, from v1 through v3, and are by no means complete — this is still an ongoing, evolving set of experiments. The scorecard tracking local and Kaggle run results lives at kaggle-v3/SCORECARD.md.

These experiments were conducted with help and healthy debate with OpenAI Codex on the directions, hypotheses, and implementation along the way.

This repository was extracted from Ouroboros so the submission experiments, model packaging, notebook generation, and result artifacts can evolve independently from the Ouroboros agent.

Layout

• kaggle/: legacy deterministic and Qwen local/Kaggle harness (v1).
• kaggle-v2/: the ouro2 exploration, induction, and director harness (v2).
• kaggle-v3/: the attributed Duck/TAAF hybrid harness with guarded Kaggle publishing and submission pipelines (v3).

Each generation has its own README, tests, configuration, and setup scripts. Generated notebooks, model data, run results, logs, virtual environments, vendored dependencies, and Kaggle credentials are intentionally ignored.

Attributions and influences

This is an independent research and experimentation repository. It combines original Ouroboros code with adapted open-source code, public benchmark infrastructure, model/runtime dependencies, and ideas from public research.

Adapted code and benchmark infrastructure

• Duck / TAAF: kaggle-v3/src/inference/ and kaggle-v3/src/taaf/ are adapted from Tufa Labs' public MIT-licensed Duck ARC-AGI-3 inference harness, including the Tufa ARC-AGI Framework (TAAF). The v3 experiment lanes and Ouroboros additions are separately developed; kaggle-v3/THIRD_PARTY_NOTICES.md records the audited source commit, dataset provenance, and contributor credit.
• ARC Prize Foundation: the benchmark and environment tooling come from ARC-AGI-3 and the ARC-AGI-3-Agents framework. The v1 and v2 setup workflows download that framework, while v3 uses the arc-agi and arcengine packages.

Models and runtimes

• Qwen: local and Kaggle experiments use Qwen3.5/Qwen3.6 models from the Qwen team. Model weights and model-card terms remain subject to the applicable Qwen and checkpoint licenses.
• vLLM, Ollama, and MLX: the inference adapters support the vLLM, Ollama, and MLX ecosystems. These are runtime dependencies, not source code copied into this repository.
• Kaggle: public validation and hidden-game submission use the Kaggle platform and its CLI; Kaggle's own terms govern remote kernels, datasets, credentials, and competition submissions.

Research influences

• Retained reasoning and compaction: the duck-memory and duck-reasoning experiments were informed by OpenAI's ARC-AGI-3 harness analysis. The local implementation is an independent Qwen/vLLM experiment and does not claim to reproduce OpenAI's model or score.
• Adaptive, self-auditing harnesses: the duck-poetiq lane is inspired by Poetiq's public ARC research, especially iterative feedback, self-auditing, selective computation, and model-agnostic orchestration. No Poetiq source code is vendored here.
• World models and verification: the v2 ouro2 lane develops explicit rule induction, replay/backtesting, planning, and reality-over-model verification for unfamiliar games. These ideas are documented in kaggle-v2/HOW-IT-WORKS.md and are this repository's own implementation experiments rather than a copied external solver.

See kaggle-v3/THIRD_PARTY_NOTICES.md and kaggle-v3/LICENSE for the detailed redistribution notice.

Quick start

git clone https://github.com/secondorderai/arc-agi-3-harness.git
cd arc-agi-3-harness

cd kaggle-v3
python3.12 -m venv .venv
.venv/bin/pip install -e '.[dev]'
.venv/bin/python -m pytest

See the generation-specific README files for local gameplay, notebook generation, validation gates, and Kaggle submission commands.

Never commit Kaggle credentials. Use the local ignored .kaggle/access_token file or the Kaggle CLI's supported environment-based authentication.

Author

Crafted by Henry at SecondOrder AI. Made in Australia.

License

This repository is licensed under the MIT License. Adapted third-party components and dependencies may carry additional notices or license terms; see kaggle-v3/THIRD_PARTY_NOTICES.md for the Duck/TAAF attribution and redistribution details.