# FightBench replay dataset v1

**Purpose:** reproducibility audits, offline controller analysis and training
experiments in deterministic simultaneous combat. Unit: one match with replay and
model I/O. Author/maintainer: Terminator contributors. Created 2026-09-23.

**Collection:** engine 0.1.1, protocol 1.0.0, seeded fixed scenarios. Official Duel
plays both sides; FFA rotates spawns in isolated suites. Download the pinned suite
manifest with every analysis. Locally seeded data uses scripted house bots; it is
not evidence of LLM performance. Paid/live matches are a separate pool.

**Contents:** scenario/config/seed, action log, final hash, recorded observations,
responses, parser outcomes, provider attempts, latency, tokens and costs when
available. No API keys or encrypted key-store entries. Scripted runs have no LLM
calls; missing historical evidence and unknown cost remain distinguishable.

**Access:** public matches plus the requester's private matches. Exports are
on-demand snapshots, not a public corpus publication. The tool does not grant a
new license to third-party model output, uploaded code or prompts; downstream
redistribution must respect the source's terms. No redistribution claim is made.

**Limitations:** small synthetic combat domain, sparse rewards, scenario reuse,
no real-world capability inference, no safety or human-preference labels. Python's
default feature vector is partial; full protocol observations are in info. Split
training/evaluation by scenario and seed and report controller revisions; never
mix live/saturated records into official estimates. A matching hash establishes
replay consistency, not independent authenticity or absence of benchmark leakage.

**Recommended reporting:** engine version, manifest hash, scenario hashes, model
revision, seeds, match count, confidence intervals, pool, export date and exclusions.
Use the BibTeX on `/research/data`, adding exact suite and revision identifiers.

**Changelog:** v1 (2026-09-23), replay + I/O + call attempts + accounting, JSONL and
Parquet. Schema: DATASET_SCHEMA.md. No human subjects or paid provider calls were
used to validate this implementation.