---
pretty_name: AgentBattler Bench results
license: other
tags:
- benchmark
- chess
- coding-agents
configs:
- config_name: dotagents_luna
  data_files:
  - split: train
    path: releases/agentbattler-dotagents-v1-8211dd28a6ef7cd19705/dotagents_luna/data/*.parquet
- config_name: dotagents_sol
  data_files:
  - split: train
    path: releases/agentbattler-dotagents-v1-8211dd28a6ef7cd19705/dotagents_sol/data/*.parquet
- config_name: dotagents_terra
  data_files:
  - split: train
    path: releases/agentbattler-dotagents-v1-8211dd28a6ef7cd19705/dotagents_terra/data/*.parquet
---

# AgentBattler Bench results

Release `agentbattler-dotagents-v1-8211dd28a6ef7cd19705` contains replayable local benchmark bundles and queryable Parquet game tables. Source code and reproduction instructions: https://github.com/aj47/agentbattler-bench.

## Contents

- `releases/agentbattler-dotagents-v1-8211dd28a6ef7cd19705/dotagents_luna`: 180 games.
- `releases/agentbattler-dotagents-v1-8211dd28a6ef7cd19705/dotagents_sol`: 180 games.
- `releases/agentbattler-dotagents-v1-8211dd28a6ef7cd19705/dotagents_terra`: 180 games.
- Each `bundle/` stores `result.json.gz`, its deterministic gzip manifest, original bundle checksums, positions, copied agents, and manifest. The uncompressed canonical result is intentionally not duplicated in the dataset tree.

## Data schema

Each Parquet row is one game using `agentbattler.hf-game-row.v1`. Stable queryable fields are suite/game IDs; white/black agent identity, display name, harness, and model; position ID, seed, FEN, and max plies; outcome, reason, failure JSON, plies count, and recorded game-result hash. `game_json` is canonical JSON for the complete recorded game, including every move and nested result field. `summary.json` gives compact aggregates; `release-manifest.json` and `SHA256SUMS` seal the package.

## Reproduction and verification

Use the matching repository revision, then run `npm run verify:hf-results -- --output <downloaded-release-root> --suite-set dotagents`. This verifies every package hash, Parquet counts/unique game IDs/aggregate counts, deterministic gzip and canonical-result hashes, then replays the compressed bundles against their existing checksums.

## Method, fairness, and security limitations

All agents used the same chess-agent prompt and legal-move contract, six fixed positions, seeded color-balanced pairings, high reasoning where supported, and isolated generation homes. Results are exploratory local evidence, not a claim of general model ability; sequential Elo is order-dependent. Claude Code used a third-party loopback Messages translation gateway to a ChatGPT Codex backend, which Anthropic does not support for non-Claude models. The gateway can introduce translation and tool-semantics differences; it returned agent source in final text with zero recorded tool calls. No Anthropic billing or OpenAI API key was used. Only accepted artifact sources and sanitized provenance are included; raw traces, rejected attempts, credentials, checkpoints, host paths, service logs, and temporary homes are excluded.

## Licensing

This repository has no checked-in LICENSE file. Dataset and generated-agent redistribution rights are therefore not asserted beyond benchmark verification until the maintainers choose and publish a license. `license: other` reflects that restriction.