Harness作为唯一可变面：信贷流水线中LLM智能体的合规受限自进化与实测准入闸门
原标题：The Harness as the Only Mutable Surface: Compliance-Bounded Self-Evolution of LLM Agents in Credit Pipelines, with a Measured Admission Gate
AI 摘要
论文提出将LLM智能体的自我进化限制在运行时harness（指令文本、工具调用逻辑、原语组合）内，保持模型权重固定，使每次适应都成为可审查的diff。作者构建了双循环引擎（局部循环编辑单个原语、全局循环替换类型化库中的原语），并设置统一的准入闸门，在部署前写入哈希链记录。在信贷流水线模拟中，闸门在7,449个候选变更中仅准入144个，均未恶化留出历史误差；而无闸门版本准入309个有害变更，90次运行中49次漏报率超10%。论文还将机制映射到欧盟AI法案高风险信贷评分条款，并指出2026年4月美国模型风险指南将智能体AI排除在范围之外。
正文节选
The Harness as the Only Mutable Surface: Compliance-Bounded Self-Evolution of LLM Agents in Credit Pipelines, with a Measured Admission Gate Abstract Self-improving large language model (LLM) agents can adapt a credit pipeline to a changed rule, but an agent that rewrites itself destroys the artefact a supervisor reviews: a named change, a recorded test, an approval. We argue that self-evolution is reviewable only if it is confined to the runtime harness—instruction text, tool-call logic and pri