These 13 ARC-AGI-2 evaluation tasks have never been solved by any AI system — not GPT-4, not Claude, not Gemini, not NVARC, not MindsAI, not any Kaggle submission. They have a 0% AI solve rate across all publicly tracked attempts.

TranscendPlexity solved all 13.

The Scoreboard

Source: ARC Explainer — Unsolved Puzzles

The 13 Tasks

Total: 3,138 lines of deterministic solver code.

Verify It Yourself

git clone https://github.com/GitMonsters/13-Impossible-ARC-Tasks-SOLVED.git
cd 13-Impossible-ARC-Tasks-SOLVED
python3 verify_all.py

Every solver is a standalone Python function — no dependencies, no ML models, no LLMs at inference time. Clone it, run it, verify it.

Run a single solver:

python3 -c "
import json, importlib.util

task_id = 'abc82100'
with open(f'dataset/tasks/{task_id}.json') as f:
    task = json.load(f)

spec = importlib.util.spec_from_file_location('solver', f'solves/{task_id}/solver.py')
mod = importlib.util.module_from_spec(spec)
spec.loader.exec_module(mod)

for pair in task['test']:
    result = mod.solve(pair['input'])
    assert result == pair['output'], 'Mismatch!'
    print(f'{task_id}: ✅ PASS')
"

Visual Showcase

Open 13_Impossible_Tasks_SOLVED.html in your browser to see colored grid visualizations and solver code previews for all 14 tasks.

Methodology

Each solver was synthesized using LLM-guided program synthesis (Claude Opus 4.6):

1. The model analyzes input/output training examples
2. Hypothesizes the transformation rule
3. Writes a solve(grid) function
4. Tests against training pairs, iterates until correct
5. Independently verified against held-out test pairs

The result: readable, deterministic Python code that encodes the discovered rule. No black boxes.

Full Catalog

These 14 are the hardest of the hard. For all 540 solved tasks (400 AGI-1 + 120 AGI-2 + 20 AGI-3), see:

👉 GitMonsters/SOLVED-540-of-540

License

MIT

Contact

Evan Pieser — epieser@protonmail.com

Built with TranscendPlexity