Verantyx

82.6% on ARC-AGI-2

Hand-crafted solvers + Claude Sonnet 4.5 program synthesis

Every solution is a verifiable Python program.

826/1000 tasks solved — LLM writes code, system verifies

Built by kofdai × OpenClaw — Human Logic + AI Implementation

Quick Install
git clone https://github.com/Ag3497120/verantyx-v6 && cd verantyx-v6 && pip install -r requirements.txt
ARC-AGI-2
0.0%
826/1000 tasks
HLE
0.0%
Humanity's Last Exam
⚡
Hand-crafted solvers + Claude Sonnet 4.5. Every solution is a verifiable program.

How the Cross Engine Solves

INPUT
✦
Phase 1: Cross DSL
Neighborhood rules — 57% of solutions
◆
Phase 2: Standalone Primitives
Flip, rotate, crop, scale
▣
Phase 3: Stamp / Pattern
Object detection → pattern fill
⟐
Phase 4: Composite Chains
2-3 step compositions
⟳
Phase 5: Iterative Cross
Residual learning without gradients
⧉
Phase 6: Puzzle Reasoning
Declarative spatial predicates
⌬
Phase 7: ProgramTree
CEGIS-based synthesis
OUTPUT ✓
Solution verified on all training pairs

How It Works

Hand-crafted solvers + LLM program synthesis — every solution is a verifiable program

▦

Input Grid

⚙

Cross Engine

🧠

LLM Synthesis

✓

Verification

▣

Output

Stage 1: Cross Engine (24.4%)

30+ hand-crafted solvers using cross-structure analysis, object movement, panel decomposition, and iterative residual learning. Pure symbolic — no LLMs, no neural networks.

Stage 2: LLM Program Synthesis (+58.2%)

Claude Sonnet 4.5 writes Python transform(grid) functions for each unsolved task. 5-6 parallel agents process batches of 50 tasks via OpenClaw. The LLM never outputs answers — it writes code.

Verifiable by Design

Every generated program is deterministically verified against all training examples. Only pixel-perfect transforms survive. No hallucination, no guessing — just provably correct code.

7-Phase Symbolic Pipeline

1
十

Cross DSL

Neighborhood Rules
57% of solutions
2
⟲

Standalone Primitives

flip, rotate, crop
Pure transformations
3
▣

Stamp

Pattern Fill
Template matching
4
⊕

Composite Chains

Multi-step pipelines
Sequential operations
5
∞

Iterative Cross

Residual learning
Recursive refinement
6
∴

Puzzle Reasoning Language

High-level abstractions
Symbolic logic
7
⊢

ProgramTree Synthesis

CEGIS
Counter-example guided
⚡
Each phase is a nested cross structure — symbolic reasoning all the way down

Evolution

From 11.3% to 82.6% — hand-crafted plateau at 24%, then Claude Sonnet 4.5 synthesis

0%20%40%60%80%100%11.3%v1915.5%v3218.9%v4721.6%v5723.4%v7224.4%v8282.6%+Synth

v19–v82: hand-crafted solvers · +Synth: Claude Sonnet 4.5 program synthesis (+58.2%)

Explore Verantyx

View Source Code

Explore the full codebase — 304 files, 100K lines

→

Models & Datasets

HuggingFace Spaces — live ARC-AGI-2 solver demo

→

Follow @kofdai

Latest updates and research progress

→

Verantyx Logic

iOS app — modal logic puzzles only

※ Modal logic problems only. Not the full ARC solver.

→

Support This Research

Verantyx is built by a student in Kyoto with no GPU cluster — just a MacBook and Claude as a development partner. The engine costs nothing to run, but building it requires API credits that add up fast.

☕

Supporter

$5/mo

Sponsors badge, early release notes, README shoutout

POPULAR
🔬

Researcher

$20/mo

Inference logs for all 1,000 tasks, failure analysis, private Discord

🏗️

Architect

$50/mo

Experimental branches, DSL drafts, monthly roadmap, direct Q&A

💜 Sponsor on GitHub⭐ Star on GitHub

You're not just funding a project — you're proving that a single researcher with the right tools can compete with billion-dollar labs on the hardest AI benchmarks.