A filtered catalyst tape, not a news archive. Several hundred crypto stories a day go in; a noise gate, an entity resolver and an impact scorer discard almost all of them, and what survives is bound to a Hyperliquid perp, scored for impact and given a signed direction. No LLM anywhere in the path — every gate is deterministic.
Impact is a bounded 0–1 score combining how much that class of event moves a market, how confidently the coin was identified, how many independent sources carried it, and how recent it is. The tick marks the 0.45 publish threshold — anything left of it never reaches the tape. Dir is the signed bias: ▼ risk-off, ▲ risk-on, ● direction-neutral. A +n next to the source is corroboration — how many other independent desks carried the same event. A story that resolves to more than one coin appears once per coin, scored and ranked independently — that is one event affecting two markets, not a duplicate.
Showing the 4 perps currently carrying news. The remaining 173 active perps are all news_pressure: 0 and are left off the board here — but the API returns them as real rows, because a cross-sectional ranking needs its zeros and a silently missing symbol is indistinguishable from a pipeline failure.
The pipeline re-scores its whole rolling window every 5 minutes, so newly qualified counts only what was not already on the tape — a story already published carries a novelty penalty on the re-run and lands in below threshold. It is normal for this row to read 0 while the tape above is populated. Counts are published on /news/sources as hourly buckets with no titles or URLs, so the filter can be watched over-tightening without us keeping a news archive to check it against.
Two numbers are the product. The headlines are the evidence.
News pressure (0–1) is how much qualified news weight a coin is carrying: a recency-weighted sum of its impact scores over the rolling 24-hour window, saturating so a burst compounds but stays bounded. News tilt (−1 to +1) is the direction of that pressure — the impact-weighted average bias. A coin can sit at high pressure with a tilt near zero when good and bad news land together, which is itself information.
The hard part of a news product is not extracting text, it is discarding 95% of it. Price predictions, listicles, sponsored posts, how-to guides, opinion pieces and weekly recaps are dropped before anything is scored. What is left has to name a catalyst class and resolve to a live Hyperliquid market.
Corroboration is the strongest free signal that a story is real. A genuine event gets picked up by every desk within hours; a press release does not. Events are clustered on the resolved (coin, category) rather than on title similarity — two desks describing the same outage in different words still land on the same event, which a word-overlap heuristic would routinely miss.
On latency, plainly: RSS lags the source by roughly 30 seconds to 5 minutes. This is a context, risk-off and event-study feed — it will not win a listing-sniping race and is not sold as one. The exception is Hyperliquid listings and delistings, which come from our own universe diff and are exact and immediate. This page is additionally a free preview cached for about 30 minutes; the API is real-time.
The coverage gap is published, not hidden. The perp universe contains tickers that are ordinary English words — TRUMP, MOVE, NOT, GAS, SAND — so those only ever match on a cashtag or their full project name, and a story about a politician never becomes a TRUMP event. /news/sources lists every symbol that can only be reached that way.
curl -H "X-API-Key: cdk_live_yourkey" \ "https://cryptodataapi.com/api/v1/news/pulse?hours=24"
/news/pulse returns one row per perp — news_pressure, news_tilt, event_count, top_impact, cross-sectionally ranked.
/news/market-moving is the tape itself, /news/coin/{symbol} narrows it to one perp, and /news/sources publishes feed health, the funnel and the coverage gap.
Free covers BTC, ETH and market-wide stories over 24h;
Pro opens the full perp universe over 7 days;
Pro Plus adds the score components and the measured market response at +15m / +1h / +4h — return, open-interest change, funding shift and volume multiplier, so a headline that moved nothing scores near zero.
Tiering changes coin coverage and score detail; it never changes the noise threshold.