遇见数据集

A Universal Cliff and a Design Fingerprint: Cross-Section Defect Detection Under LLM Orchestration — Data and Code

收藏
Zenodo2026-05-25 更新2026-05-26 收录
官方服务:

资源简介:

Data and code accompanying the preprint "A Universal Cliff and a Design Fingerprint: Cross-Section Defect Detection Under LLM Orchestration" (Hiroki Fukui). This deposit contains all runs behind the paper: the full run logs (final integrated reports and internal-state probes for every run), the four source documents with their embedded cross-section defects and their defect-free catch variants, the sixteen answer keys, the ETR (error-detection) and catch scores, the scoring and signal-detection code (including the verbatim ETR judge prompt), and the orchestration engine's section-assignment schemes — both the inherited and the corrected version — so that the protocol deviation described in the paper (and its confirmatory re-run) is fully inspectable. An exploratory fresh-read supplement is included but is not used in the paper's claims. Model outputs are released as evidence about system behavior, not as authored content; see the responsibility statement in README.md. Code is released under MIT and data under CC BY 4.0. Preregistration: OSF osf.io/fnbct (registrations Z9QW3, R7MH6, ZTDMJ). The prior single-paradigm stage of this research is arXiv:2605.13851.

提供机构:
Zenodo
创建时间:
2026-05-25
二维码
社区交流群
二维码
科研交流群
商业服务