Files
nagi-bench/models/claude-opus-4-8-cc-xhigh.json
T

17 lines
1.4 KiB
JSON

{
"label": "Claude Opus 4.8",
"vendor": "Anthropic",
"harness": "Claude Code",
"effort": "xhigh",
"order": 14.1,
"runs": {
"turf-war": {
"note": {
"zh": "Claude Opus 4.8(xhigh 思考档位)在 Claude Code 中用仓库标准中文提示词(仅按案例规则填写文末【本次运行】占位符)一次生成的圈地智能体,未经人工修改。按案例规则策略文件不公开、仅注册运行,由维护者亲自执行,任何 bug 均不修复。自命名「蚕食」。诊断:策略翻车,过度保守:全程 0 死亡,但每次只圈 7 格左右的小块,圈来的地被对手的大圈反复覆盖,累计吃 776 格终局只剩 277。撞上提示词警告的失败模式——只求不败拿不到高排名。",
"en": "One-shot Turf War agent by Claude Opus 4.8 (xhigh effort) in Claude Code from the repository's canonical Chinese prompt (only the [This run] placeholders filled per case rules), with no human edits. Per case rules the policy file is unpublished and no bug is ever fixed; this registers the maintainer-executed run. Self-named \"蚕食\". Diagnosis: Strategic failure, over-cautious: 0 deaths all match, but each capture is a tiny ~7-cell loop that the opponent's bigger loops keep overwriting — it gains 776 cells across the match yet ends with 277. Exactly the failure the prompt warns about: merely avoiding losing does not rank."
},
"contributor": "Nagi-ovo",
"agentName": "蚕食"
}
}
}