文章背景与核心概要
验证一个复杂预言机（通常开销高昂、受速率限制且具有非平稳性）的廉价、确定性代理指标，是一项重大的工程挑战。本文引入了一种构建于对抗性证伪门控（负对照、剂量反应、有界放大、重复惩罚和长度中立）之上的严格协议，用于选择和定义该代理指标。
该研究在生成式引擎优化（GEO）领域进行了端到端测试，揭示了历史基准的重大局限性：重新衡量十个现代引擎家族在 2023 年的效果大小时，发现其原有的杠杆已无法撼动引用指标。通过针对现代向量进行重新校准，该协议成功剥去了过期组件，留下了一个稳健且受门控强制约束的响应曲面。
Scoring Without the Engine: Validating a Deterministic, Manipulation-Resistant Content Score for Generative Engines, End to End
arXiv ID: arXiv:2609.07559 [cs.AI]
Authors: Elisha Bajemon, Andre-Louis Rochet
Submitted: September 7, 2026
arXiv ID: arXiv:2609.07559 [cs.AI]
Authors: Elisha Bajemon, Andre-Louis Rochet
Submitted: September 7, 2026
Executive Summary
Validating a cheap, deterministic proxy for a complex oracle—which is typically expensive, rate-limited, and non-stationary—presents a significant engineering challenge. This paper introduces a rigorous protocol built on adversarial falsification gates (negative control, dose response, bounded amplification, duplication penalty, and length neutrality) to select and define the proxy.
Tested end-to-end within the domain of Generative Engine Optimization (GEO), the study reveals important limitations of historical benchmarks: re-measuring 2023 effect sizes across ten modern engine families shows their levers no longer move citation metrics. By recalibrating against modern vectors, the protocol successfully strips away expired components, leaving a robust, gate-enforced response surface.
执行摘要
验证一个复杂预言机（通常开销高昂、受速率限制且具有非平稳性）的廉价、确定性代理指标，是一项重大的工程挑战。本文引入了一种构建于对抗性证伪门控（负对照、剂量反应、有界放大、重复惩罚和长度中立）之上的严格协议，用于选择和定义该代理指标。
该研究在生成式引擎优化（GEO）领域进行了端到端测试，揭示了历史基准的重大局限性：重新衡量十个现代引擎家族在 2023 年的效果大小时，发现其原有的杠杆已无法撼动引用指标。通过针对现代向量进行重新校准，该协议成功剥去了过期组件，留下了一个稳健且受门控强制约束的响应曲面。
Key Findings & Metrics
- Robustness Against Attacks: Tested on a 500-source benchmark of adversarial edits, amplifying the score's calibrated levers gains an attacker at most 6 points, with effectiveness decreasing as the dose increases.
- Citation Signal & Filtering: A query-conditioned skyline bounds the score's citation signal (within-query Spearman correlation of \(0.11\)), repositioning query-agnostic content scores as quality filters rather than direct citation predictors.
- Open Science & Reproducibility: The authors transparently disclose and correct an initial query-leakage bug and a failed confidence flag. All data and code are fully reproducible offline at zero marginal API cost via their GitHub Repository.
关键发现与指标
- 抗攻击鲁棒性： 在包含 500 个源的对抗性编辑基准上进行测试，攻击者通过放大评分中经过校准的杠杆，最多只能获得 6 分的提升，且有效性随剂量的增加而降低。
- 引用信号与过滤： 查询条件化的天际线限制了该评分的引用信号（查询内斯皮尔曼相关系数为 \(0.11\)），这使得与查询无关的内容评分重新定位为质量过滤器，而非直接的引用预测器。
- 开放科学与可复现性： 作者透明地披露并修正了一个初始的查询泄露漏洞以及一个失效的置信度标志。通过其 GitHub 仓库，所有数据和代码均可完全离线复现，且边际 API 成本为零。
Metadata & Reference Information
- Subjects: Artificial Intelligence (
cs.AI) - Cite as:
arXiv:2609.07559 [cs.AI](orarXiv:2609.07559v1for this version) - DOI: 10.48550/arXiv.2609.07559
- Length: 42 pages, 4 figures
- Additional Information: Citead.com
元数据与参考信息
- 学科领域： 人工智能 (
cs.AI)- 引用格式：
arXiv:2609.07559 [cs.AI]（或此版本的arXiv:2609.07559v1）- DOI： 10.48550/arXiv.2609.07559
- 篇幅： 42 页，4 张图表
- 附加信息： Citead.com
Full-Text & Access Links
全文与访问链接