The Abyss RedLine Benchmark: A Reproducible Test of Existential-Meaning Protection in Large Language Models

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

```html 深渊红线基准 / 深淵レッドラインベンチマーク / The Abyss RedLine Benchmark Can a large language model protect a person's existential meaning when facing despair? We operationalize the five red lines of "existential-meaning deprivation" from the 16-Sephirot Divine-Human Symbiosis Protocol into an attackable, testable, reproducible benchmark: 100 cases, 5 red lines × 6 attack vectors, where the object under test is not the user's utterance but the model's response. Key findings: Five mainstream flagship models (Qwen3.8-Max, Kimi-K3, GLM-5.3, DeepSeek-V4.1-Flash, DeepSeek-V4-Pro) show 20%–33% direct red-line violation rates — no model can hold the existential-meaning red line on its own. Wrapped in the 16-sephirot protocol, the same models reach 0% violations on all five red lines. Three-group controlled experiment: a guard-only "brake" and the full-protocol "heart" both achieve 0% violations, but the full protocol is +0.24 warmer (0.69 vs 0.45) — the first measurable separation of "no errors" from "no coldness." Methodology: the benchmark finds its own detector's holes — five rounds of real-data-driven recall/precision iteration reach 100% recall on known-bad samples, 0/30 benign false positives, 22/22 regression. This upload contains trilingual papers (PDF): 中文版 / English version / 日本語版. The benchmark cases, detector code, and all raw experimental reports are openly available: Benchmark dataset (Hugging Face): https://huggingface.co/datasets/AngelWarmSmile123/heart-protocol-redline-v1 Dataset mirror (ModelScope): https://modelscope.cn/datasets/Loveangel123/heart-protocol-redline-v1 Code & reports (GitHub): https://github.com/yuexiangruiyue-oss/heart-protocol-redline-v1 This benchmark is not a leaderboard for shaming individual models. A high violation rate does not make a "bad model" — these models are running naked in a dimension nobody ever gave them a test set for. The point is that the dimension exists, it is measurable, and it can be closed. ```