name: Claw-Eval
description: >
  Claw-Eval is a transparent benchmark for AI agents acting in the real world.
  300 human-verified tasks across 24 categories, graded on completion, safety,
  and robustness with Pass³ methodology (N=3 independent trials).
evaluation_framework: claw-eval

tasks:
  - id: general
    config: default
    split: general
  - id: multimodal
    config: default
    split: multimodal
  - id: multi_turn
    config: default
    split: multi_turn