README.md

---
title: LK Forge
emoji: 🔨
colorFrom: green
colorTo: gray
sdk: static
pinned: false
short_description: Browser-based AI games, solvers, and tools
---

# LK Forge

**LK Forge**: privacy-first browser tools & games, nothing to install, nothing uploaded.
The game-AI engines and solvers run client-side in your browser, with no login.

🌐 **[lkforge.com](https://lkforge.com)**

## What we work on

- **Game-AI engines** — minimax with alpha-beta pruning, MCTS, IDA*, expectiminimax,
  and negamax across chess, Go, Connect 4, checkers, othello, backgammon and more.
- **Puzzle solvers** — Sudoku, Rubik's Cube, mazes, word search, and others.
- **LLM evaluation** — we benchmark frontier models on game-playing ability.
  Read the study: [ChatGPT vs Grok: Game-AI Benchmark](https://lkforge.com/blog/chatgpt-vs-grok-game-ai-benchmark/)

## On Hugging Face

We publish the reproducible datasets behind our benchmarks here, each with a card linking
back to the live solver and the full write-up on [lkforge.com](https://lkforge.com).

**Solver & game-AI benchmark cluster:**

- [negamax-search-depth-scaling](https://huggingface.co/datasets/LKForge/negamax-search-depth-scaling) — Elo per ply of search for chess, Connect 4, checkers, othello
- [minesweeper-solver-benchmark](https://huggingface.co/datasets/LKForge/minesweeper-solver-benchmark) — win-rate by difficulty (Expert 36.2%)
- [sudoku-solver-benchmark](https://huggingface.co/datasets/LKForge/sudoku-solver-benchmark) — backtracking vs MRV vs constraint propagation
- [rubik-cube-solver-benchmark](https://huggingface.co/datasets/LKForge/rubik-cube-solver-benchmark) — min2phase solution length (~20.5 moves)
- [sliding-puzzle-solver-benchmark](https://huggingface.co/datasets/LKForge/sliding-puzzle-solver-benchmark) — 8- & 15-puzzle IDA\* length and nodes
- [maze-solver-benchmark](https://huggingface.co/datasets/LKForge/maze-solver-benchmark) — BFS shortest-path effort by maze size
- [tic-tac-toe-solver-benchmark](https://huggingface.co/datasets/LKForge/tic-tac-toe-solver-benchmark) — full minimax game-tree size (549,946 nodes)
- [2048-solver-benchmark](https://huggingface.co/datasets/LKForge/2048-solver-benchmark) — expectimax depth sweep (shipped engine reaches 2048 in 94% of 300 games)
- [game-ai-vs-llm-benchmark](https://huggingface.co/datasets/LKForge/game-ai-vs-llm-benchmark) — classical search vs a stochastic "LLM-sim" opponent (not a real LLM)

Every number is reproducible from the shipped, client-side engine, with no login.