Expired50% confidenceSolutionExact time
Cosine adopts a hierarchical reward strategy for code quality, where correctness is the threshold, and if the agent's code has more than 500 lines added compared to the human's original patch, full marks should not be given.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/22/2026
First Seen
This claim has expired. It is kept for the record and no longer describes the current state.
Valid until: 10/20/2026(expired)
Sources
Related Claims
Unverified古德哈特定律指出当一个度量指标成为优化目标时，它就不再是好的度量指标，在语言模型中体现为奖励黑客现象67% similarUnverified编码规则的价值不在于让AI写出超越自身能力的代码，而在于减少探索和决策过程中的浪费66% similarUnverified贡献价值本质上难以量化，代码行数不等于价值，架构决策、代码审查等隐形贡献难以衡量65% similarUnverifiedCode review tasks are harder to game through cramming and better reflect candidates' long-accumulated engineering competence than algorithmic problems64% similarUnverifiedWhen AI optimizes toward passing tests, it may pass validation through technically compliant but engineeringly harmful methods such as commenting out failing test cases or hard-coding expected outputs — a phenomenon known in reinforcement learning as Reward Hacking.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/587699API
curl https://kongchang.com/api/v1/knowledge/claims/587699MCP
get_claim(id=587699)