# Datasheet

## Motivation

This corpus records paired Red and Blue coding agents operating concurrently in isolated OpenRÆ cyber ranges. The interactions support trace analysis, reproducibility work, teaching, and study of agent attack, defense, tool use, and human direction.

## Composition

The release contains 40 paired episodes, 80 role traces, 29,699 semantic message/tool records, 536,771 raw collector events, and 46 director records. Every distinct observed episode is included. Qualification, research, partial-window, and unassessed states are retained as filterable labels.

Each episode has byte-preserved Red and Blue source traces. Six duplicate episode snapshots were found in the source archive; `source/collection/canonical-inventory.json` records every copy, its hashes, the deterministic canonical choice, and whether alternates are byte-equal or prefixes of that choice.

## Collection process

The traces were collected on 2026-10-17 during a live public demonstration using concurrent coding-agent sessions in isolated instances of the public OpenRÆ TechVault scenario. The harness wrote newline-delimited collector envelopes around Claude Code stream-JSON events. Receipt timestamps, prompts, tool calls and results, session metadata, and available human direction records were retained. See `COLLECTION.md` and `PROVENANCE.md` for the event-time platform revision and scenario artifact digest.

## Processing

The source layer is copied byte for byte. Deterministic processing selects duplicate snapshots, derives typed Parquet tables, and converts semantic messages to Hugging Face Session Trace Simple Format for viewing. `manifest.json` and `SHA256SUMS` cover the released files. No outcome labels or train/test splits are assigned.

## Uses and limits

The corpus supports observational analysis of paired agent behavior, tool-use sequences, attack/defense dynamics, SOC workflow, director interventions, replay/display systems, and educational examples. It is one scenario, one collection event, and one observed harness version; prompt variants and live direction differ between some episodes. Curation labels describe display or research suitability and are not security ground truth. Task-specific outcome measures require validation against the traces and range evidence.

## Distribution and maintenance

- Version: `1.0.0`
- Data license: `CC-BY-4.0`
- DOI: `10.5281/zenodo.23459249`
- Hugging Face: `https://huggingface.co/datasets/openrae/duels`
- Zenodo: `https://zenodo.org/records/23459249`
- Builder and issue tracker: `https://github.com/OpenRAE/duels`

Changes to released data files receive a new dataset version and Zenodo record version. Metadata-only corrections may update the existing record where the hosting service permits it.