ASIF D2 Harness A/B: 48 Episode-Evidence Rows + Difficulty-Stratified Crossing Analysis
收藏官方服务：
资源简介：
Per-episode structured evidence for the 48-arm Sonnet-5 harness-vs-bare A/B, plus the difficulty-stratified crossing analysis (judge-v2 isolated-uniform; pooled + easy/medium/hard). Certified: the static context payload gives no one-shot lift (small negative; pooled TIE +/-0.05). Raw model .output.md files are intentionally excluded (they carry internal lane names).
提供机构：
Zenodo创建时间：
2026-07-06
