Candidate generation
Adjacent candidates come from live news: your top
tastes each pull the entities currently trending around them
(co-mentions, peers, domain stories) and graft in as
news candidates that expire with the news
cycle unless you engage. Liking a leaf still drills for more-specific
subtopics (also news-grounded). Inspect the raw signal in the
trending-adjacency tester →
Only likes count as interest. A card that was ≥60% visible for ≥500ms
then scrolled away unliked logs a skip — dwell time is never a positive
signal. Each card shows its fate in a status line once judged:
dimmed = seen & skipped, pink = liked. Unseen cards log nothing and
show no status. Each topic takes at most one skip per 30 min.
The ♥ → line on each card previews how a like's credit cascades — full
weight to the topic, and decaying roll-up credit to its parent
subjects (italic); after a skip it flips to the negative flow
actually applied.