Pulling the wires…
Fetching ~60 newsrooms. First deck takes a few seconds.Deck's done.
You worked through every card in this beat.
Swipe, or use ← → to call it · Z undo · Space next
Pulling the wires…
Fetching ~60 newsrooms. First deck takes a few seconds.You worked through every card in this beat.
Swipe, or use ← → to call it · Z undo · Space next
Publishers you rated more or less generously than the score says. Positive means you trusted them more than the engine does.
Every publisher the engine knows, and the four numbers behind each score. Not in this list means the article falls back to an unrated baseline and is flagged as such.
Four dimensions, scored 0-10, weighted and scaled to 100:
Factual record carries the most weight because it is what readers actually care about. Transparency and independence matter because they are what keep the factual record honest over years rather than weeks.
Lean runs from -5 (left) to +5 (right). A mild lean costs nothing; the penalty grows quadratically so only the extremes are punished, up to 14 points. Direction is displayed but never rewarded. Being slanted is not the same as being wrong, so it is a modest thumb on the scale rather than a dimension of its own.
Clickbait formulas, conflict verbs, shouting caps, punctuation pileups and anonymous hedging pull the score down. Hard attribution and specific figures pull it up. The total is clamped to 10 points either way, so a careful headline can never launder a bad outlet and a lazy one never sinks a good one.
The engine's own call is a single line: 65 and above is "trust", below it is "sketchy". That is roughly where a standing corrections desk stops existing. Your calibration is how often your swipe landed on the same side as the engine.
About four in five real articles come from publishers above that line, so a deck drawn at random could be beaten by swiping right every time. Each deck is therefore dealt roughly half from each side, with a few satire wildcards. The scores are never touched, only which cards you are shown.