How we judge
Not another leaderboard. RightSignal keeps a single, current answer per category — the model we’d actually reach for today — and shows its working.
One working answer
Each category has exactly one titleholder, judged against criteria published on its page. The board is reviewed daily against the leaderboards, fresh releases and what practitioners are reporting, and it changes the moment the evidence does — not the moment a press release lands.
Evidence rules
Nothing takes a title on launch-day hype. A change requires at least one independent, non-vendor evaluation, and a challenger must hold its case for a week before a title can move. Licences are read from repositories, not press releases; open-weight picks must be verifiably downloadable, linked to the original repository.
The radar
New releases are tracked from the day they appear — marked unverified while only the vendor’s own figures exist, and early data once the first independent numbers land. Radar entries either graduate into formal challengers (and sometimes titleholders) or age off the board after three weeks.
Every change on the record
Every title change is dated, cited and kept: full lineage tables on each pick’s page back to 2023, and a complete changelog. The board is also re-verified weekly against the live web, and the corrections are published like everything else — you can check our working, back to day one.
The sources we watch
Blind-vote arenas (Arena text, code, image and video boards), composite indexes (Artificial Analysis, LiveBench), task benchmarks (SWE-bench Verified, Terminal-Bench), domain leaderboards (Open ASR, RTEB, the speech arenas), release feeds, and primary sources — repositories, model cards and licences. Vendor numbers are reported as claims until someone independent reproduces them.