name: "SKBench-Env"
description: >
  SKBench repackaged as a Harbor agent environment: the agent fetches each
  cited source and judges whether the quoted text survives SKBench's
  mechanical normalization check. Eval-only.
evaluation_framework: harbor

tasks:
  - id: skbench_env
    config: default
    split: test