This is a gated artifact. Access is reviewed manually. Complete the access-request form on this page; the actAVA team typically responds within one business day. The handbook is not redistributed with the main actava/chi-bench dataset because of its sensitivity and clinical-author provenance.

What this is

The Managed-Care Operations Handbook is the agent-facing skill used by χ-Bench — a long-horizon, policy-rich healthcare-workflow agent benchmark. It is a lazy-loaded, reference-style book mounted read-only at /workspace/skills/managed-care-operations-handbook/ inside each task container. The agent reads the index (SKILL.md), opens the one sub-book matching its active role, then opens chapters on demand.

It anchors the clinical criteria, operational rules, and sequenced workflows for the three managed-care operations roles the benchmark simulates:

Plus shared appendices: a role-platform tutorial set and a deep clinical medical-library appendix.

Layout

managed-care-operations-handbook/
├── SKILL.md                       # index — role routing + how-to-use
└── references/
    ├── provider-pa/               # Prior-auth specialist sub-book (51 md + 8 pdf)
    ├── payer-um/                  # UM reviewer sub-book (8 md + 3 pdf)
    ├── care-manager/             # RN care-manager sub-book (8 md)
    ├── platform/                 # per-role platform tutorials (3 md)
    └── medical-library/          # deep clinical reference appendix (1,208 md)

~1,279 markdown pages + 11 PDFs, ~34 MB. The medical-library is lookup-driven — navigate by filename, never read linearly.

How χ-Bench uses it

1. The Harbor runner mounts this handbook read-only into every task container at /workspace/skills/managed-care-operations-handbook/.

2. The agent identifies its role from instruction.md, jumps to the matching sub-book's SKILL.md, and opens chapters on demand.

3. Skill ablation. χ-Bench's CHI_BENCH_SKILLS_ABLATE env var blanks out chosen sub-books at trial start (replacing each with a one-line stub SKILL.md) to measure how much the handbook contributes to agent performance. The four reported conditions:

   Condition	CHI_BENCH_SKILLS_ABLATE
   full (control)	unset
   no_domain	provider-pa,payer-um,care-manager
   no_medical	medical-library
   none	provider-pa,payer-um,care-manager,medical-library

Mount the handbook at data/skills/managed-care-operations-handbook/ for the ablation runs; see the main dataset card.

Access & license

• Gated, manual review. Submit the access-request form on this page. Each request is reviewed by the actAVA team before access is granted.
• License: actAVA clinical-collaborator license — non-commercial research use only, no redistribution. See actava.ai/benchmarks/contact.
• This artifact is distributed separately from the Apache-2.0 χ-Bench task fixtures precisely because of its clinical-author provenance and licensing terms.

Citation

@misc{chen2026chibenchaiagentsautomate,
      title={CHI-Bench: Can AI Agents Automate End-to-End, Long-Horizon, Policy-Rich Healthcare Workflows?},
      author={Haolin Chen and Deon Metelski and Leon Qi and Tao Xia and Joonyul Lee and Steve Brown and Kevin Riley and Frank Wang and T. Y. Alvin Liu and Hank Capps MD and Zeyu Tang and Xiangchen Song and Lingjing Kong and Fan Feng and Tianyi Zeng and Zhiwei Liu and Zixian Ma and Hang Jiang and Fangli Geng and Yuan Yuan and Chenyu You and Qingsong Wen and Hua Wei and Yanjie Fu and Yue Zhao and Carl Yang and Biwei Huang and Kun Zhang and Caiming Xiong and Sanmi Koyejo and Eric P. Xing and Philip S. Yu and Weiran Yao},
      year={2026},
      eprint={2605.16679},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.16679},
}