证据优先Evidence first
从来源、数据边界与验收方式开始，而不是从模型名称开始。Begin with provenance, data boundaries, and acceptance criteria, not a model name.
关于 YiHarvestAbout YiHarvest
把 AI 能力放进可靠系统。Putting AI capability inside reliable systems.
当前专注多模态内容基础设施、Agent 执行内核、故障证据与开放网络工具。模型能力只是起点，边界、恢复和可验证性决定系统能否被真正使用。Current work spans multimodal infrastructure, agent execution kernels, failure evidence, and open-web tools. Model capability is only the starting point; boundaries, recovery, and verifiability make a system usable.
工作原则How I work
不追逐复杂本身。先理解失败方式、信任边界和真实操作路径，再选择技术。Start with failure modes, trust boundaries, and real operating paths, then choose the technology.
系统思考Systems thinking
把输入、路由、执行、确认、恢复和观测看成同一个系统。Treat input, routing, execution, confirmation, recovery, and observation as one system.
克制交付Useful restraint
清楚说明限制。界面、文档、默认值和失败信息都属于工程质量。State limits clearly. Interfaces, docs, defaults, and failure messages are engineering quality.
技术关注Technical focus
TypeScript 与 Python 为主，覆盖 Next.js、FastAPI、MCP、数据与任务系统、模型适配和可观测性。Mostly TypeScript and Python across Next.js, FastAPI, MCP, data and job systems, model adapters, and observability.