data/protocol.md

CyberBench protocol v1

• Seed: 25
• Items: 160 (40 triage, 40 CTI, 40 Sigma, 40 RAG)
• Split: 96 development / 64 test, by family, not by rewriting one prompt
• Interactive: 20 ids listed in splits.json
• Gold: stored on the item record for scoring; omitted from visible_input and from prompts
• Judges: deterministic validators only. No LLM judge in the recorded scores
• Quality: 0.15 * json_validity + 0.40 * task_score + 0.25 * groundedness + 0.20 * abstention
• Latency and cost are reported beside quality and are not mixed into a hidden rank
• Unevaluated / unavailable is written not_evaluated, never 0
• IOCs use documentation ranges (192.0.2.0/24, 198.51.100.0/24, 203.0.113.0/24) and .example.test
• No exploit generation, no live SIEM, no real secrets
• Level-1 synthetic fixture. It does not prove operational accuracy