The connector lives next door to the wire.
A press release is public the instant its CDN serves it. Everything after that is geography. We put a connector inside the same metro as each publisher's point of presence, so the first read costs single-digit milliseconds instead of a transatlantic round trip.
[01]The fleet
Six metros. Every connector within one hop of what it reads. Distance is the one latency cost you cannot optimise away in software. So we spend on it in hardware instead.
Washington, D.C.
first byte ~4 ms
New York
first byte ~3 ms
Chicago
first byte ~5 ms
London
first byte ~6 ms
Frankfurt
first byte ~5 ms
Tokyo
first byte ~7 ms
[02]Anatomy of a connector
Five stages between the wire and your socket. Every one of them is somebody's excuse for a second of latency. None of them get one here.
Poll on a jittered loop
Each source gets a cadence tuned to how it publishes: sub-second on the wires, long-poll on EDGAR. Jitter keeps the fleet from stampeding a single origin on the second boundary.
Detect change before parsing
Content hash first, parse second. An unchanged page costs one comparison instead of a full DOM walk, which is what makes a tight polling loop affordable at all.
Extract into a typed artifact
Boilerplate stripped, tickers and entities resolved, publish timestamp taken from the source's own clock rather than ours. The output is the same artifact shape the Verdict Engine scores.
De-duplicate across the fleet
The same release crosses three wires within a second. A content fingerprint collapses them into one artifact so a desk sees one alert, not three.
Hand off to scoring
The artifact enters the Verdict Engine in the same process boundary it was parsed in: no queue hop, no serialization tax, before it fans out to subscribers.
[03]When it breaks
Connectors break. The question is whether you find out. A feed that fails silently is worse than no feed: you will keep trading against it.
The page changes shape
A publisher reworks their newsroom template on a Sunday and every selector you own goes stale.
What we do
Parsers assert on their own output. A field that stops resolving fails the artifact loudly instead of emitting an empty string, and pages a human with a diff of what moved.
The origin pushes back
Rate limits, bot walls, and IP reputation checks are aimed at clients that behave badly.
What we do
We identify ourselves, honour robots directives and back off on 429s. Politeness is not altruism. A blocked connector is a blind one.
A region drops
A metro loses transit mid-session. The source is fine; your view of it is not.
What we do
Every source has a warm secondary in another metro. Failover is automatic and costs the extra network distance: degraded, not dark.
Silence that looks like calm
The most dangerous failure is a connector that returns 200 and nothing else. Quiet reads as a slow news day.
What we do
Each source carries an expected-volume envelope for its time of day. Falling out of that band raises an incident on the public status page, not a shrug.
[04]At a glance
The fleet, in four numbers. Each one is on the status page, where you can watch it move.
6
Metros
One connector cluster per publisher concentration.
< 8 ms
Fetch to parse
P95, measured at the edge node.
99.98%
Fleet uptime
Trailing 90 days, published on /status.
0
Aggregator hops
Nobody repackages a headline before you see it.
Stop paying for someone else's round trip.
The fleet is infrastructure you would otherwise have to build, colocate and babysit yourself. Point a socket at it instead.