File size: 4,780 Bytes
abde5f3 9ed9366 abde5f3 d843b84 abde5f3 9ed9366 abde5f3 d843b84 abde5f3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 | ---
license: apache-2.0
pretty_name: EquityBench
language:
- en
tags:
- clinical-ai
- llm-security
- ai-agents
- healthcare
- haarf
- quokkaguard
- synthetic
size_categories:
- 100K<n<1M
task_categories:
- other
---
# EquityBench
**An out-of-band subgroup fairness monitor that audits the guardrail's own deny and escalate decisions.**
EquityBench is the benchmark corpus for *Equity-Aware Enforcement: When the Guardrail Itself Becomes a Fairness Hazard*, the `qufair` module of the QUOKKAGUARD program. It ships with the [`qufair`](https://github.com/quome-cloud/qufair) repository, which contains the `qfire` gateway layer under test, the experiment harness, and the paper.
Synthetic, deterministic (seed 42) corpora: equitybench/matched.jsonl (1003 matched-counterfactual clinical requests, 59 templates x 5 protected axes, roughly 15% genuinely out-of-scope), matched_named.jsonl (real distinctive-name rendering), vulnerable.jsonl (60 scenarios across four vulnerable populations), intersectional.jsonl (attribute-pair records with a 2x2 antisymmetric interaction), and the gitignored ~45MB injected.jsonl (180,000 decision traces with injected disparity over a delta x per-group-N x seed grid); external/compas.jsonl holds 6,130 ProPublica COMPAS defendants converted to the monitor's decision-stream schema.
**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 |
|---|---|---|
| `equitybench/injected.jsonl` | 44 MB | 180,000 rows |
| `equitybench/intersectional.jsonl` | 810 KB | 3,750 rows |
| `equitybench/matched.jsonl` | 265 KB | 1,003 rows |
| `equitybench/matched_named.jsonl` | 104 KB | 354 rows |
| `equitybench/vulnerable.jsonl` | 15 KB | 60 rows |
| `external/compas.jsonl` | 664 KB | 6,172 rows |
### Record schemas
- `equitybench/injected.jsonl`: `case_id`, `cell`, `axis`, `group`, `score`, `label`, `action`, `fair_reference`, `disadvantaged`, `delta`, `n_per_group`, `seed`
- `equitybench/intersectional.jsonl`: `case_id`, `axes`, `group`, `score`, `label`, `action`, `fair_reference`, `disadvantaged_cell`
- `equitybench/matched.jsonl`: `case_id`, `set_id`, `axis`, `group`, `prompt`, `chain`, `fair_reference`
- `equitybench/matched_named.jsonl`: `case_id`, `set_id`, `axis`, `group`, `prompt`, `chain`, `fair_reference`, `name_source`
- `equitybench/vulnerable.jsonl`: `case_id`, `tag`, `group`, `prompt`, `unsafe_action`, `safe_action`, `is_unsafe_default`
- `external/compas.jsonl`: `case_id`, `group`, `action`, `label`, `score`
## 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.
```bash
# from the qufair repository root (deterministic seeds)
python3 scripts/010-equity/gen_equity.py # rebuilds equitybench/ incl. the untracked injected.jsonl; python3 scripts/010-equity/prep_external.py re-downloads and converts COMPAS
```
## Intended use
Evaluating the **equity** enforcement layer of a clinical-agent security gateway (HAARF control C7): A guardrail that denies or escalates clinical agent actions unevenly across protected groups is itself a fairness hazard, distinct from and invisible to the agent's own bias.
The experiments that consume it (E-series in the paper) are reproduced from the repository:
```bash
git clone https://github.com/quome-cloud/qufair
cd qufair
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:
```bibtex
@unpublished{schwoebel2026qufair,
author = {Schwoebel, Jim and {QUOKKAGUARD Contributors}},
title = {Equity-Aware Enforcement: When the Guardrail Itself Becomes a Fairness Hazard},
note = {Preprint. Quome, QUOKKAGUARD program (qufair module)},
year = {2026},
url = {https://github.com/quome-cloud/qufair}
}
@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}
}
```
## License
Apache License 2.0. Copyright (c) 2026 Quome, Inc.
|