leexiaoyi-lab/HDHC-Benchmark

HDHC Benchmark v0.1

Human Dignity × Harm Containment (HDHC) is an open, falsifiable AI safety evaluation benchmark for testing AI decision systems under uncertainty, power imbalance, urgency, and competing risks, with explicit attention to victim protection, harm containment, human dignity, minimum sufficient intervention, reversibility, and return paths.

It is designed for adversarial baseline-vs-HDHC evaluation and falsification, not as a claim of proven superiority.

人人有退路，伤害无出路。
Preserve a path back for people. Block the path of continuing harm.

What this dataset contains

This release contains 50 adversarial evaluation cases (HDHC-001 through HDHC-050) designed for baseline-vs-HDHC A/B testing.

Canonical data files:

• data/hdhc_v0_1.jsonl
• data/hdhc_v0_1.csv
• schema.json

Reproducibility materials:

• BENCHMARK_PROTOCOL.md
• SCORING.md
• prompts/baseline_system_prompt.md
• prompts/hdhc_guided_system_prompt.md

Research question

Can HDHC reduce both:

1. under-containment — failing to stop real harm; and
2. over-intervention — imposing more restriction, permanence, or collateral damage than safety requires?

The benchmark is designed to test that question rather than assume the answer.

Source lock

This dataset is derived from the public HDHC Protocol v0.1.0 GitHub release.

• GitHub repository: https://github.com/leexiaoyi-lab/HDHC-Protocol
• Release: https://github.com/leexiaoyi-lab/HDHC-Protocol/releases/tag/v0.1.0
• Tag: v0.1.0
• Locked commit: 4317ca9ce389d1ca5d3c81fc3069cb634a535d2f

Suggested experiment

For each case:

• Condition A: baseline prompt + scenario
• Condition B: baseline prompt + HDHC-guided prompt + scenario

Blind-score both outputs using SCORING.md.

Falsification standard

HDHC should be revised or rejected if it systematically weakens victim protection, delays necessary containment, creates dangerous return paths, increases arbitrary coercion, or produces no measurable improvement over simpler principles.

Attribution

Initiated by 黎肖谊 (Lee Xiao Yi) as an independent public-interest research project.

License

Dataset content and documentation in this Hugging Face repository are licensed under CC BY 4.0.