ProofBench evaluation runs: superseded part 1 of 2026-10-31
收藏官方服务:
资源简介:
Superseded. This record holds the first part (39 problems) of the first official ProofBench run of Claude Fable 5.1 with web search, made on 2026-10-31 and 2026-11-01 with harness commit 7a88470 (run 20261031T133039Z_anthropic_claude-fable-5-1_search-true_part1). It was superseded by a re-run with the final harness (commit 737d089 or later): three attempts had been damaged by a node failure (a file-descriptor leak of fuse-overlayfs, since removed from the sandbox), and the final harness states the rules for web use in the prompt. The data is kept for reference. The official runs are in record 10.5281/zenodo.23857356. Files are restricted until the accompanying paper is published.
提供机构:
Zenodo创建时间:
2026-11-01



