用 Jev 做编码工作流的路由与评分层
昂贵模型只负责写代码，Jev 用 Choice/Score 判断下一个 worker、给草稿打分、决定是否转人工，置信度低于 0.85 就升级给人类。
if your best coding model also decides which worker goes next, you hired a senior engineer to run a traffic light. that’s the JEV use case I care about. Give a backend workflow the repo, requirements and examples of the existing code. The expensive writing model builds the implementation. But choosing the next worker, scoring a draft and deciding whether a handoff needs review are different jobs. JEV takes the current state plus a narrow, typed question. It can return a choice, a score or a yes/no probability. Choice and Score also include confidence. Your code uses those answers to route the work: research → implementation → checks failed check → repair → checks again artifacts ready → human review The useful part is the evidence between steps. A changed file, the current revision and the latest test output go into the state. A test failure is a fact the runner can act on. JEV can judge the next narrow step; it doesn’t turn its opinion into proof that the code works. If the requirements are ambiguous, the workflow should expose that gap. In this replay, a Choice confidence floor of 0.85 routes uncertainty to a human. That’s an example policy, not an 85% correctness guarantee. The reviewer gets the implementation, tests, docs and summary. Merge and deployment remain separate approvals. The economics are worth understanding: TypeSafe publishes $42 per billion input tokens, or $0.042 per million. At an average of 1,000 input tokens per call, 10,000 calls would mean about $0.42 in JEV input charges. That’s arithmetic, not a measured run or the cost of the whole coding system; the writing model and tools still have their own costs. The attached video is a scripted HTML replay of this architecture. It shows the loop worth testing: expensive models create, JEV supplies narrow judgments, code enforces the workflow, and a human owns the final handoff. That’s the split I’m breaking down in my JEV article. Before adding another agent, decide which decisions actually need your best model.
这条是引用着另一条写的：x.com/0xkurt/status/2103410921892618651