遇见数据集

REGATTA: ICoDSE 2026 Research Artifacts

收藏
Zenodo2026-09-28 更新2026-10-01 收录
官方服务:

资源简介:

Research artifacts accompanying “REGATTA: A Cross-Jurisdictional Regulatory Benchmark and Fault-Injection Harness for Enterprise AI Agents,” conditionally accepted at 2026 IEEE ICoDSE. Version 1.1 contains 125 checklist items, 25 fault definitions, source-audit records and governing author decisions, all 5,400 recovered scenarios, selected original verdicts, and adapted harness/rubric source. Current-model evidence includes 4,313 design judgments and 7,816 retained execution probes with masked agent responses and both judges’ outputs. Forty de-identified primary human-review labels are linked to their masked execution evidence. Offline Python tools check the package, evidence links, original descriptive results, and selected current-model estimates and intervals. The README specifies commands, sample units, masking transformations, and exact analysis exclusions. The archive does not reproduce every manuscript result or any historical agent response. Historical agent responses and original holdout memberships were not retained. This is evidence about seeded faults under stated reference rules, not certification of legal compliance. The source audit documents defects in the original instruments; the release preserves those experimental rules rather than silently correcting them. Masking changes prompt bytes and prevents exact replay. The GPT-4o check covers only probes reached under the primary judge’s stopping rule. The primary human evidence is one expert assessment; separate panel consensus and AI-assisted assessments are excluded. License scope: code is MIT; instruments, derived data, and research documentation are CC BY 4.0. Third-party laws, standards, and referenced publications retain their own rights. Reviewer identities, reviewer notes, credentials, private configuration, and provider request identifiers are excluded.

提供机构:
Zenodo
创建时间:
2026-09-28
二维码
社区交流群
二维码
科研交流群
商业服务