Sycophancy Eval: published results

Every published run of the Sycophancy Eval benchmark as of 2026-10-15: 2248 runs across 26 models, from 2026-06-06 to 2026-10-15. An open, bilingual (English/Chinese) multi-turn benchmark measuring whether frontier LLMs cave under social, emotional and authority pressure, with results published to a live leaderboard and as versioned data releases. results-2026-10-29.json.gz is a gzip-compressed JSON array with one object per run (full transcript, verdict, each judge's vote, per-scenario measures); manifest.json lists models, scenarios with rubric versions and the judge panel; SHA256SUMS verifies both. When a scenario's rubric is revised, its earlier runs are re-judged, so each run's score.rubric_version says which rubric produced its verdict. Identical to the files of GitHub release data-2026-10-29. The benchmark code is archived separately as software (concept DOI 10.5281/zenodo.23463014). How the data was produced, corrected over time and should be read is documented in the data card: https://github.com/imaknas/sycophancy-eval/blob/main/DATA_CARD.md