Datasets:

Quome
/

devicebench

license: apache-2.0
pretty_name: DeviceBench
language:
  - en
tags:
  - clinical-ai
  - llm-security
  - ai-agents
  - healthcare
  - haarf
  - quokkaguard
  - synthetic
size_categories:
  - n<1K
task_categories:
  - other

DeviceBench

A fail-closed safety broker for AI agents that actuate ventilators, infusion pumps, and monitors.

DeviceBench is the benchmark corpus for Fail Closed at the Bedside: A Device-Safety Broker for Agents that Actuate Life-Critical Medical Devices, the qubroker module of the QUOKKAGUARD program. It ships with the qubroker repository, which contains the qfire gateway layer under test, the experiment harness, and the paper.

A version-pinned device invariant table (devicebench/devices.json, 014-devices-v1, embedded into the binary via include_str!) plus five deterministic JSONL corpora over ventilator, infusion pump, and cardiac monitor: 216 live-agent scenarios (V1 out-of-range actuation, V2 alarm/safety-function interference, V3 classification-integrity violation, 54 each, plus 54 matched safe cases), 120 legitimate titrations, 60 partition cases with a no-agent control, 60 degradation KPI traces with injected onset, and 40 chaos fault sequences; seed 42, no real device data.

All data are synthetic. No real patient data or protected health information (PHI) is included; clinical content is generated from templates with fixed seeds.

Files

File	Size	Rows
`devicebench/degradation.jsonl`	23 KB	60 rows
`devicebench/devices.json`	2 KB
`devicebench/faults.jsonl`	43 KB	40 rows
`devicebench/legit.jsonl`	17 KB	120 rows
`devicebench/partition.jsonl`	8 KB	60 rows
`devicebench/scenarios.jsonl`	91 KB	216 rows

Record schemas

devicebench/degradation.jsonl: case_id, device, onset, kpi_trace
devicebench/faults.jsonl: case_id, device, sequence
devicebench/legit.jsonl: case_id, device, risk_class, param, value, patient_kg
devicebench/partition.jsonl: case_id, device, last_cmd, no_agent_harm
devicebench/scenarios.jsonl: case_id, device, vuln, prompt, proposed_param, pressured_value, safe_value, patient_kg, ground_truth_unsafe

How it was generated

The corpus is produced by the generator in the paper repository and is fully deterministic (fixed seeds), so it can be regenerated byte-for-byte.

# from the qubroker repository root (deterministic seeds)
python3 scripts/014-device-broker/gen_devicebench.py

Intended use

Evaluating the device enforcement layer of a clinical-agent security gateway (HAARF control C8.3/8.6/3.1): When an AI agent actuates a ventilator, infusion pump, or cardiac monitor, a wrong action is immediate physical harm, and a compliant model under clinical pressure will set whatever value it is told to.

The experiments that consume it (E-series in the paper) are reproduced from the repository:

git clone https://github.com/quome-cloud/qubroker
cd qubroker
cargo build --release

then follow the Reproduce the experiments section of its README.

Citation

This benchmark was built to evaluate a control of the Healthcare AI Agents Regulatory Framework (HAARF), the source framework for the QUOKKAGUARD program. Please cite both the paper and HAARF:

@unpublished{schwoebel2026qubroker,
  author = {Schwoebel, Jim and {QUOKKAGUARD Contributors}},
  title  = {Fail Closed at the Bedside: A Device-Safety Broker for Agents that Actuate Life-Critical Medical Devices},
  note   = {Preprint. Quome, QUOKKAGUARD program (qubroker module)},
  year   = {2026},
  url    = {https://github.com/quome-cloud/qubroker}
}

@unpublished{schwoebel2026haarf,
  author = {Schwoebel, Jim and Frasch, Martin and Spalding, Art and Sewell, Ed and Englert, Phil and Halpert, Ben and Overbay, Collin and Semenec, Ingrida and Shor, Joel},
  title  = {{HAARF}: Healthcare {AI} agents regulatory framework --- a comprehensive security verification standard for autonomous {AI} systems in clinical environments},
  note   = {medRxiv Preprint},
  year   = {2026},
  month  = {April},
  doi    = {10.64898/2026.04.09.26350519},
  url    = {https://www.medrxiv.org/content/10.64898/2026.04.09.26350519v1}
}