한국형 AI 벤치마크 프로토콜

한국어, 문서, 전문 업무, 배포, 비용을 평가하는 공개 벤치마크 프로토콜입니다.

Protocol published · Results not yet published

The site will not display synthetic scores before tasks, graders, prompts, versions, and run logs are reproducible.

Six evaluation tracks

Honorifics, ambiguity, spacing, dialect, long-form consistency, and KR↔EN translation.

HWP/HWPX, Hangul tables, government templates, PDF citations, and spreadsheet fidelity.

Legal research, public administration, customer service, finance, and clinical-document tasks with domain review.

Fresh Korean web search, source quality, citation correctness, and answer traceability.

Multi-step browser and office tasks in Korean interfaces, including recovery from tool errors.

KRW cost, latency from Korea, data residency, on-premise availability, context limits, and API reliability.

Prompts, source documents, expected outputs, and scoring code are versioned.

Professional tasks use at least two independent Korean domain reviewers.

Model version, provider, date, region, settings, latency, and KRW cost are retained.