Decision Models
FRESH DEVELOPMENTS / PUBLIC CONVERSATIONS
Fresh headlines from publisher and primary-source feeds, alongside a ten-minute sample of public conversation.
Listening before labelling.
Topics appear after feed headlines or sampled posts have enough evidence and Clef has screened them.
Feeds retain headline metadata from the last 24 hours. Counts show source coverage, not popularity.
A bounded sample, not a measure of all Bluesky. One account counts once per topic; distinct accounts are not necessarily distinct people.
How the radar works ↗SAVED SOURCES
Source links you wanted to come back to. Saved in this browser.
OBSERVE / SCREEN / EXPLAIN
The radar reads fresh feed headlines and a bounded public conversation sample. Clef screens what is new before it appears.
Publisher and primary-source feeds supply recent headlines. Publication and indexing times are identified separately. Bluesky’s Jetstream supplies a bounded ten-minute sample of public posts. Conversation counts use distinct author accounts; feed source counts remain separate.
About Jetstream ↗Clef screens candidate headlines and phrases using structured choices. For developments, it compares recent accepted coverage to distinguish a new event, a substantive update and repetition. Labels come from source text. Without approval, a label stays off the map.
Meet Clef-flash ↗Select a development to see why it is new, publication and first-observed times, source headlines and the decision receipt. Conversations retain their author counts and activity history. Social momentum compares two complete two-minute windows and needs four continuous minutes of coverage.
More from TamTa Lab ↗AN HONEST EXPERIMENT
The map shows only approved, unexpired topics with current source evidence. A paused model can leave an earlier valid approval visible; its review time stays attached. Pausing visual updates does not stop shared collection.
Model confidence does not establish truth. Source checks concern availability and display policy; they are not fact checks. Social activity is a sampled English-language conversation, not a representative trend survey. Feed developments use headline metadata; full article bodies and linked pages are not read. Equal-size development rings do not imply equal importance, popularity or certainty.
The earlier Jev experiment remains a separate, frozen comparison. Its assistant-authored references are provisional; they do not establish independently verified accuracy on this feed.
KEEPING IT SMALL
This is a scenario, not measured spend. The radar screens topic candidates in batches, so its real prompt sizes and request counts differ.
Assumes 5,000 input tokens per batch and 30 days. A default 720 batches/day could represent 12 active hours at one batch/minute; actual scheduling differs. USD list prices; excludes hosting and free allocations. Total billed output includes reasoning tokens where charged. Providers may count tokens differently.
Rates checked 1 November 2026: Clef-flash / Clef · TypeSafe. OpenAI · Anthropic. Decision models use input-only rates of $0.038 / $0.042 / $0.10 / $0.24 per million. Astra and Fable use $10 input / $50 output per million. This compares token rates, not equivalent accuracy.