JEV DESK:单页 HTML 里的五阶段决策流水线
把线索、工单、文档、记账四条队列接进同一个决策模型:模型只做 Choice/Score 选择,代码负责执行,0.85 置信度门控决定是否转人工。
HE LLM NEVER TOUCHES THE BUTTON. JEV DESK is a decision pipeline in one HTML: 5 stages, 30 nodes, 4 workflows most agent setups let the model write and act. JEV DESK splits it the idea: the model picks from a list, plain code does the rest i wired one decision model into 4 queues that eat most of a small team's week: leads, tickets, documents, bookkeeping. every case runs the same path STATE → ASK → JEV → GATE → CODE ▹ what runs under the hood 1. STATE: email / DM / ticket / pdf / receipt packed into 0.6–2.4 KB of json. no chat history, no prose 2. ASK: 3 typed question kinds · Choice / Score / Noul · 6 answers in parallel off one state 3. JEV: no generation, only a label + confidence · ~256 ms median reported on public demos · ≈ $0.42 per 10 000 decisions 4. GATE: hard line at 0.85 · lead 0.89 pass · document 0.94 pass · receipt 0.92 pass · ticket 0.71 → human gate 5. CODE: book_call / slack ping / file.move / ledger post. 4 plain functions, the model never calls them itself 6. GUARDRAILS: spend cap $1.00 · 10 actions max · never auto-post to X · every run logged lead DM → book_call · ticket #4812 → on-call · contract.pdf → /contracts · receipt $129 → 6100 software, draft alone the LLM writes nice text and guesses. alone the code has no idea what to do. together the model only decides, the code only acts, and the 0.85 gate decides who sees it before it ships ▹ prospects tile 10 on the board says trading. same pattern: state = mint + holders + dev %, Choice = ape / pass / small, gate, then a plain function sends the tx. the model never gets the private key bookmark this before your agent gets write access to everything edge is open until everyone else wakes up
这条是引用着另一条写的:x.com/vibelancer/status/2111786213191376899