# BDX Bench

> BDX Bench is an independent, MIT-licensed benchmark harness, leaderboard and playable browser-game showcase for testing AI models against identical real-world software briefs. The public site ranks nine models by Showdown Score v2 (a 5-dimension, 100-point source-code audit of the games they built from one shared prompt); the zero-dependency Node.js harness adds the swe-mini and terminal-mini task suites and a blind A/B chat arena ranked by Elo (start 1000, K = 32).

Results are deliberately small-scale and indicative: one prompt, nine models, eleven builds, a single audit round. Mock-mode harness runs need no API key.

## Docs

- [README](https://github.com/MHJoy99/bdx-bench#readme): overview, leaderboard table, quickstart, API, v0.1 limits
- [Methodology](https://github.com/MHJoy99/bdx-bench/blob/main/docs/METHODOLOGY.md): Showdown Score v2 (Part A) and harness suites and scoring (Part B)
- [Arena](https://github.com/MHJoy99/bdx-bench/blob/main/docs/ARENA.md): blind-match rules, seeded pairing, judging, Elo
- [Quickstart](https://github.com/MHJoy99/bdx-bench/blob/main/docs/QUICKSTART.md): 5-minute harness walkthrough
- [Full plain-text description for LLMs](https://mhjoy99.github.io/bdx-bench/llms-full.txt): everything in one file

## Try it

- [Project website](https://mhjoy99.github.io/bdx-bench/)
- [Live leaderboard and playable builds](https://bench.bdx.market/)
- [Source code on GitHub](https://github.com/MHJoy99/bdx-bench)

## Optional

- [Issues](https://github.com/MHJoy99/bdx-bench/issues)
- [Author: MHJoy99](https://github.com/MHJoy99)