--- title: "Jev will sort it in seconds. A refund still needs a name." subtitle: "Trial the classifier on fifty real threads. You write the folders. Auto-archive stays off." principle: "Signed bins" date: "2026-09-18T10:28:25.436Z" author: "Mikko Atienza" canonical: "https://ciel.codes/showcase/blog/the-label-did-not-finish-the-ticket" edition: "2026-09-17" source: "https://ciel.codes/showcase/pulse/2026-09-17" abstract: "Jev will sort a shared inbox for pennies. It will not refund a customer, sign a clause, or archive a thread. That split is the work this week. You trial the classifier on live mail. You keep the yes." cite: "Atienza, M. (2026). Signed bins. Pulse." citations: - handle: "@rileybrown" url: "https://x.com/rileybrown/status/2100404532119269426" --- Jev will sort a shared inbox for pennies. It will not refund a customer, sign a clause, or archive a thread. That split is the work this week. You trial the classifier on live mail. You keep the yes. ## Filing is a pick, not a paragraph Most shops already tried a chat box on the same pile. It wrote a draft. Someone still had to file the message. The draft was not the job. The job was putting the thread in the right bucket so the right person could act. Jev does not write. You hand it the thread and a closed list of folders you named. It returns one folder and a confidence number. It cannot invent a seventh folder. It cannot send. It cannot put money back on a card. That is why it belongs on triage and why a chat window never owned this queue. Riley Brown (@rileybrown), who shows operators what these tools cost when you actually run them, watched Jev classify 500 emails in seconds. It cost 3.5 cents. He is not handing you a new mailbox. He is putting a number on the old excuse. Cost is not why the pile sits. The pile sits because nobody signed the bins, a sample never tested them, and a label still gets treated like a close. You already have the inbox. You already have an office manager or an ops lead who lives in it. Jev is in early access at typesafe.ai. Get in, then open the playground. The week is fifty real threads through that classifier, not a demo ticket and not a rollout. ## Jev only sees the folders you signed Do not start from someone else's default folders. Start from the mail that actually arrives. For a team of twenty the list is short. Write it on a sheet. Put a person's name next to every bin that may never disappear on its own. 1. Order and delivery 2. Billing and refund 3. Quote and sales 4. Contract and legal 5. Vendor and supplier 6. Noise and FYI The owner signs that list. Not the intern. Not Jev. You are deciding what "refund" means in your building, and whether a delivery update may ever archive itself. If two people would file the same thread in two bins, the bins are wrong. Rewrite them before anyone pastes a message into the playground. Finance, or whoever issues refunds, is the closer on billing. Whoever signs paper is the closer on contract and legal. The inbox lead can accept Jev's labels on the rest once you have signed the definitions. They cannot close a refund. They cannot approve a clause. They cannot tell a customer the number that will hit the invoice. In the playground the left side is the message. The question is a Choice with those six bins as the only allowed answers. Add a second question: does this need a closer who can spend money or sign paper. Write both answers on the row next to the thread. Low confidence is a flag, not a close. You still read the ones you would not defend. > **Callout.** Do not let Jev auto-archive billing, refund, contract, or legal. A label on those threads is a route to a named closer. It is not permission to disappear the mail. ```mermaid flowchart TD A["Owner writes the bins"] --> B["Inbox lead exports fifty threads"] B --> C["Inbox lead runs those fifty through Jev"] C --> D{"Owner accepts Jev's labels?"} D -->|No| A D -->|Yes| E["Inbox lead routes the day's mail"] E --> F{"Refund, legal, or escalate?"} F -->|Yes| G["Named closer takes the thread"] F -->|No| H["Owner keeps auto-archive locked"] style A fill:#1e1b4b,color:#e2e8f0 style B fill:#064e3b,color:#e2e8f0 style C fill:#4c1d95,color:#e2e8f0 style D fill:#064e3b,color:#e2e8f0 style E fill:#1e1b4b,color:#e2e8f0 style F fill:#4c1d95,color:#e2e8f0 style G fill:#064e3b,color:#e2e8f0 style H fill:#1e1b4b,color:#e2e8f0 ``` Owner, inbox lead, named closer. On a small team those hats sit on two people. Write them anyway. Friday you need to know which hat failed, not which label ran. ## The first fifty are the policy Do not dump the whole mailbox into Jev on Monday. Export fifty recent threads. Oldest unread first, or the last two days, whichever is the real pile. Subject and body are enough. The inbox lead runs each one through Jev against the signed bins, then you sit with the fifty and disagree out loud. If you reverse more than a thin slice, the bins are not ready. You do not patch a sentence and hope. You change the written definitions. You run the fifty again. Only when you can sign the bins does the inbox lead sort the day's new mail the same way. Cold mail waits in a labeled pile. Hot mail goes to the named closer the same day. The pass Riley watched, 500 messages in seconds, is proof that volume is not your constraint. Your constraint is whether Jev's labels match how you actually work. A fast sort on the wrong bins is a faster wrong inbox. Once the bins are signed, Tuesday through Thursday look the same. The inbox lead runs the day's new mail through Jev. Hot threads leave the pile the same morning with a closer named on the row. Cold threads stay labeled and visible. Friday the owner reads two numbers: labels reversed, and hot threads with no closer. If both are quiet, you may unlock auto-archive on noise and FYI only. You still do not unlock billing or legal. That is the week. Not a rollout. A sample that earned a narrow archive rule. Keep an audit log a person can read. Message, Jev's label, confidence, who accepted or reversed it, whether a closer was named. A sheet is enough for the log. The classifier is Jev. The sheet is how you prove the sample held before anyone turns archive on. > **Aside.** The 3.5 cents is the cost of one Jev pass on 500 messages. It is not your close rate, and it is not a forecast for your inbox. ## The sheet you fill on Friday Nobody has a credible figure for how many of your open threads are labeled and still ownerless. Count yours. Empty cells are honest. | Measure | Count | |---|---| | Messages Jev classified in Riley's run | 500 | | Cost of that Jev run | 3.5 cents | | Sample threads you ran through Jev this week | | | Jev labels the owner reversed | | | Hot threads with no named closer | | The first two lines are someone else's pass. They exist so you stop treating cost as the reason you have not started. The last three lines are the week. If the reverse count is high, you rewrite the bins. You do not open auto-archive. If hot threads have no closer, Jev ran and the work did not. ## Run Jev on fifty live threads this week Do not boil the archive. Fifty messages from this week's real inbox. - [ ] Write the six bins on a sheet and sign them - [ ] Export fifty recent threads from the shared inbox - [ ] Open Jev and classify each against those bins - [ ] Reverse any label you would not defend to the customer - [ ] Name a closer on every refund, legal, and escalation thread - [ ] Leave auto-archive off until Friday's reverse count is a number you can live with You can run this yourselves on this week's pile.