---
title: README
emoji: 🃏
colorFrom: indigo
colorTo: gray
sdk: static
pinned: false
---

# mtg-kernel

Policy networks that play Magic: The Gathering (Pauper), trained only by reinforcement learning: terminal win or loss is the only reward, with no human game data and no imitation of heuristic bots. They run on the open-source [mtg-kernel](https://github.com/jackmaiorino/mtg-kernel) rules engine and are rated on the [Spellbench](https://github.com/jackmaiorino/spellbench) Pauper board.

| Model | Elo (95% CI) |
| --- | --- |
| [Archon 3](https://huggingface.co/mtg-kernel/archon-3-pauper) | 1381 (1315 to 1463) |
| [Archon 2](https://huggingface.co/mtg-kernel/archon-2-pauper) | 1211 (1151 to 1288) |
| [Archon 1](https://huggingface.co/mtg-kernel/archon-1-pauper) | 1179 (1114 to 1254) |

Spellbench `pauper-kernel-v2` board, 2026-10-28 snapshot, with the uniform-random bot anchored at 1000. All entries are policy-only: they sample from the network with no search. Each model card has build and run steps.

Every rated model is published except the current top one.

> Unofficial fan content. Not affiliated with, endorsed by, or sponsored by Wizards of the Coast.