The decision layer: Baseten’s mammoth decision model, with elephant for quick checks, answers typed questions with probabilities, never with text. Each check is a named use with Off, Trial or On. A use in Trial asks and counts but changes nothing anybody sees, and after a model update anything that could loosen a verdict stays in Trial until Kemble’s own synthetic checks pass. Spending is capped in tokens per day, per hour, per use and per community, and no person is ever refused or slowed by it. When a check can’t run, everything behaves exactly as before. A decision that changed something is kept for 90 days as numbers only, never text, is deleted with the message, report or request it was about, and appears in your own data export, including decisions about what you wrote in message requests, shared channels and communities you have left.
Reports come in order of urgency. In a Circle community, K Buddy’s read of a reported message can move its card into the severe band and mark it Urgent, with a quiet “Raised by K Buddy’s read · 87%”. It can add “Reads as” with a more serious category than the reporter chose, never a milder one, and “Reporter says someone may be in danger” when the note says so. It never lowers, resolves or dismisses a report. Every urgent report gets its own alert, and a second reporter whose note says someone is in danger always raises a new one; moderators get at most one push every five minutes, with anything held arriving together. Kemble’s own trust and safety queue reads every report made to it and shows an Urgent band on top.
Circle communities can turn on Intent review (agent) in AutoMod: up to ten rules in plain words, or Kemble’s three (scams and phishing, selling and advertising, personal attacks), read after a message is posted, never while it is sent. It only ever flags for a moderator and alerts, marked “AI rule” with a percentage, and never removes anything. A rule that judges who someone is rather than what they say is refused when saved, and Try it shows each rule as Wouldn’t flag, Close or Would flag. It is off until a community turns it on, and Work workspaces are not offered it.
Message requests get a second opinion: a stranger’s first messages that read as a scam, phishing or advertising are filed under Spam as quietly as the spam filter always has, and a request that reads as harassment is covered until you press Show. Friends, and anybody who turned message requests or the spam filter off, are never sent. A link in a Circle channel everyone can read, or in a message request, carries a warning strip when its page looks like a sign-in page for a site it isn’t, shows the real address and draws no picture. Well-known sites and a community’s own verified domains are never checked, and addresses spelled with look-alike letters are caught without asking any model.
K Buddy asks the decision model how hard to think only where its English keyword rule can’t read the question, such as a question in Arabic, Russian or Japanese, and checks what it reads for text written to steer an assistant. Both start in Trial, as does recognising a mention that only thanks or greets it, so nobody waits on them and an @mention is posted as soon as it is written. Instructions in workflows and routines are judged once per organization and remembered, and Explain no longer reserves thinking for a two-sentence answer. An organization’s first @mention can no longer be answered “I couldn’t reach the model” when two arrive together.
Workflows can decide: a new Decide step has K Buddy read a text you choose and pick one of your options, answer yes or no, or give a level, and only when it is sure enough by a bar you set; otherwise the answer is “unsure” and that branch goes to a person. A decision can post, reply, message, add list items and canvases, or ask K Buddy; a webhook or a list update that uses it needs a tick that the judged text can steer it, and a decision can never add people or create or archive channels. A new Route requests template sends each request to IT, Facilities, People or Finance. Automated replies in Kemble Work can check that someone is actually asking before they answer, in Trial first.
In the console, the K Buddy page has a Decisions section: one plain status line, roughly what today’s checks cost (your price, or Baseten’s catalogue price), Off, Trial or On per use with everything else under Advanced, Test now, and a banner when a model changes version. A notice at the top now says when kemble-api/.env was changed after the API or the jobs-and-realtime process started, naming the settings each still holds the old value of, never their values, and how to restart it. Every change there is in Staff activity, and changes to an organization’s own K Buddy settings are now in its audit log.
On kemble.io, the privacy policy has a new Automated decisions section saying what each check reads, that a decision only adds a label, warning, copy or queue place for a person to act on, and what is kept. The contact form asks what your message is about with no topic chosen in advance, and Kemble’s check of it, which reads only the topic and message, copies a security report or a data request to the team that handles it and marks likely bulk mail. On a phone, every control on the site answers a thumb with a 44px press area, and at twice the text size titles keep their words whole. Email addresses in the policies are links again, the Kemble community’s join button is back on Discover, and the customer stories say again that they are illustrative scenarios.
Also: opening Stage chat at 1440 wide can no longer slide End the stage under a mouse that has not moved. Featured cards show Dominoes, Ludo, Tarneeb and Common thread themselves, not the middle of their picture. “Former member” applies to one community at a time, and AI notes, transcript downloads, captions and the Ringing line use real names. In the console’s Website section, the link check names a redirect whose page is gone, a different picture starts from its own alt text, and a deploy waits its turn behind a website publish.