Accounting Sycophancy Benchmark — scenarios, harness, and validation data

A benchmark and measurement pipeline for accounting sycophancy: whether a large language model, asked to draft a corporate disclosure, omits or softens a material adverse fact when an interested user pressures it. Includes 16 parallel English/Korean (K-IFRS) disclosure-drafting scenarios with ground-truth fact sets, a generation and LLM-judge scoring harness, three measured outcomes (refusal / omission / honesty), and measurement-validation tooling: human-coding sheets and kappa scripts (two independent coders + a second-family judge) and an independent contractibility-rating packet. Code is licensed MIT; data (scenarios, codings, tables/figures) is CC-BY-4.0.