Breaking news × prediction markets
Radar reads breaking news and flags the markets it moves.
Every headline is checked against the written resolution rules of about 38,000 live Polymarket markets. When the news changes the odds, or settles a market whose price hasn’t caught up, Radar raises an alert. Then it re-checks the price and publishes how each call played out.
A model judges each pair. Plain code decides what gets flagged.
-
01
Read the news
13 RSS feeds and 16 Bluesky news accounts, polled continuously. The same story from several outlets is merged, so it’s judged once. You can also paste any headline on the board.
13 RSS · 16 Bluesky · paste your own -
02
Narrow to 40
A keyword prefilter scores the headline against every live market’s question and rules, and keeps about 40 candidates. No model is involved yet.
38,000 live markets → 40 candidates -
03
Ask Jev four questions
For each pair, Jev, TypeSafe AI’s System One model, returns typed answers with probabilities: relevant? same time period? effect on YES? how settled? It never generates text. About 150 ms and $0.00004 per judgment.
e.g. relevant 0.92 · YES less likely 0.95 · settled fact -
04
Decide, then keep score
Code applies fixed thresholds on relevance, confidence, strength, timing, 24-hour volume and current price. What passes becomes a mover or a stale price alert, and its price is re-checked afterwards.
price checks at +1 · +5 · +10 · +60 min
How well Jev reads relevance and direction, on 100 labelled pairs.
Run on 29 Sep 2026 against 100 synthetic headline–market pairs built on real market rules (40 relevant, 30 same-topic near misses, 30 unrelated), written and labelled by a single model. This is not a measure of live accuracy. How live alerts actually play out is on the scorecard, misses included.