Report-contingent scoring prompts increase upward misreporting in large language models

Description

The data-only package contains archived JSONL/API-return records, randomization records, failure records where applicable, comprehension summaries, and run manifests. It excludes Python, Stata, and R programs; processed datasets; figures; derived analysis outputs; historical pilot batches; and unrelated mock or exploratory data. No human-subject data are included. The package is intended to document the data underlying the manuscript and support provenance checks. Because several collections use provider APIs and separate collection batches, exact replication may depend on provider availability, endpoint state, and recorded collection metadata. Analysis programs are available from the authors upon request.

Institutions

• Shandong University

  Shandong, Jinan

Categories

Licence