One decision model: elephant now answers every check. The larger model, mammoth, sleeps when nobody has asked it for a while, so its first answers took seconds and Test now mostly said it did not answer in time. Elephant answered every time in under a fifth of a second and matched mammoth on 7 of the 8 tasks we tested, so Kemble keeps the decision layer, runs every check on elephant and has retired mammoth. Its thresholds were checked again on elephant with Kemble’s own wording and synthetic examples: urgent reports now go first from 30%, look-alike sign-in pages are warned on from 70%, and Intent review stays at 80%. The console’s Decisions card has one model, one Test now, one price and one money line, and the privacy policy names one decision model, elephant-1.
Kemble now checks each new version of its decision model by itself. Baseten updates the model several times a day without notice; within about a minute of seeing a new version, Kemble runs its own synthetic checks once, in the background, and never while anybody waits. If they pass, nothing changes. If the model answers a check wrongly, everything that use does stays in Trial until a later check passes, and if the model can’t be reached nothing is held. The Decisions card and Staff activity say what happened in plain words, such as “Baseten updated the model at 09:12; Kemble’s checks passed”. Test now waits up to 25 seconds for a model that was asleep, says “Waking the model…” while it waits, and then how long it took.
What the decision layer reads has grown where it was missing. Intent review now reads a poll’s question and every answer, and each picture description, on send and on every edit; it stays Circle-only, opt-in and flag-only. A workflow can no longer route a Decide step’s answer through another step, or a branch that only runs when the answer went one way, to add people or create channels. With K Buddy limited to some spaces, automated replies in notes to self check “only when asked” like everywhere else. A stranger’s request that carries a link is never rescued from Spam by a model. Look-alike link warnings in Home message requests are kept for a day per sender, never shared between senders, and after the privacy policy is published every part of Kemble sees it within seconds.
Kemble Desktop 0.2.8: move your camera. Where a camera can pan, tilt or zoom, the camera settings offer Pan, Tilt and Zoom under “Pan, tilt and zoom” (just Zoom, under “Framing”, for a camera that only zooms). Each is a labelled slider that works from the keyboard, changes the live picture straight away, shows pan and tilt in degrees and is remembered for that camera on this device. In a browser or on a phone the camera is asked for exactly what it was before, and a camera that can’t move shows nothing new.
In the console, a new Model usage section shows every model Kemble uses — K Buddy’s providers, the decision model, speech-to-text and the picture check — per use case and model: requests, requests and tokens a minute (the average and the busiest minute, with its day and time), average tokens in and out per request, a rough cache hit %, and the traffic’s pattern in words with a Steady, Variable, Bursty or Spiky badge. Filter by time range, provider, model, use case, product or only failing; the filters stay in the link, and the table can be saved as CSV or JSON in the order it is sorted, or as the per-minute or per-hour series. Every export is in Staff activity. It is counts only: nothing about who asked, where, or what was said is recorded.
Also: /who, a list’s CSV export and a group link’s page name people by their name or @handle, and say “Former member” for an account that is gone. On kemble.io, the contact form’s “choose a topic” message can be edited in the console like its other messages, and Discover says in one quiet line that the communities shown are examples.