How Arcline's Collection Layer Works

Arcline's collection layer is the part of the platform that keeps watching after the plan is written. It reads public sources continuously, cuts what it finds down to the geometry and window of an active mission, and traces every claim back to the source that produced it.

Most teams already know the problem it solves. You build a picture of the ground during the advance, and then the ground keeps moving while the picture sits still. We wrote about that gap separately. This is the piece about what runs underneath it.

What we watch

Collection is a layer, not a feed. Some of it runs continuously, some starts when a mission has an area and a window, and some fires only when an identifier enters the picture.

The continuous watch is tiered, and the tiers matter more than any single name in them.

Global. GDELT for news-derived events drawn from the world's press across dozens of languages, which is how something surfaces from a city where nobody on your team reads the local paper. GDACS, the disaster alerting system backed by the UN and the European Commission, for earthquakes, floods, cyclones, and other hazards at the moment they are declared. Maritime safety broadcasts for the coastal and port picture.

National. The National Weather Service and the National Hurricane Center for atmospheric hazards and active advisories. FAA temporary flight restrictions, which are frequently the first public signal that something has changed above a venue.

Regional and metro. A curated registry of wire services, government publishers, and metro newsrooms. Municipal crime reporting wherever the city publishes it. Traffic and work-zone data, which is the least glamorous category on this page and the one that most often changes an arrival plan.

Today the layer ingests from more than 80 public sources across those tiers. The number moves as we add and retire feeds, and it is not the interesting part of this post.

The registry underneath those tiers is curated, and the curation is where the time goes. Sources earn their place by proving out: a regional outlet with a working city desk stays, a portal that reposts press releases four days late gets retired. Anyone can subscribe to a wire. The registry is the accumulated record of which sources are actually worth a small team's attention, and it gets pruned as often as it grows.

Two other modes run alongside the continuous watch. Mission-demand collection starts when a mission has geometry and a window: atmospheric and space weather for that area and those hours, pulled because someone is going there. It closes when the mission does. Selector lookup runs when a domain, email address, phone number, or IP address enters the picture, querying public records against that specific identifier and attributing the answer to where it came from. A different pipe, the same discipline.

Every source above is public. Any operator with a browser and a free afternoon could reach all of it. The advantage was never access.

Four sources, one event

Here is what separates a collection layer from an alert feed.

A magnitude 6.1 earthquake gets reported by a disaster alerting system, three wire services, and a metro outlet. An unfiltered feed hands the operator five notifications and leaves them to work out that it is one earthquake. Most tools deduplicate by URL, which catches syndicated copies of the same article and misses entirely that five organizations described the same event in five different sentences.

Arcline resolves them. Records describing the same event, in the same place, inside the same window collapse into a single entity, matched on native identifiers where a source provides one and on space and time proximity where it does not. What reaches the operator is one event with every contributing source still attached to it.

That attachment is what makes corroboration visible. An event that four independent publishers reported is a different proposition from one that appeared in a single outlet, and the operator can see which one they are holding without opening anything.

Fusion is a word this industry uses loosely, usually to mean several sources arriving in the same inbox. Resolving them into one entity while keeping the provenance chain intact is a different piece of engineering, and it is the one that decides whether a four-person team can use what comes in.

The pipeline

Five stages, in order.

Poll. Event and hazard sources on a fifteen-minute cycle, news sources on thirty. Every response is stored as an observation before anything interprets it, so the raw record survives whatever happens downstream.

Extract. Structured content comes through as structured content. Prose gets read for the entities, places, and event types inside it, and anything that cannot be placed on the map with confidence is held rather than guessed at.

Geo-gate. An item that cannot be located near an active mission does not proceed. This is the first and largest cut, and it is why the layer can watch the whole world without drowning a small team.

Resolve. The step described above. Native identifiers, then space and time proximity, then corroboration against what is already in the graph.

Surface. What survives becomes a finding attached to the mission it touches, with its source and its rating, waiting on the triage queue for a human to acknowledge or dismiss.

Filtering happens at several points along that chain rather than once at the end. Where the thresholds sit is the product of running this against live missions and tuning what got through, and they are the difference between a system that surfaces four things a day and one that surfaces four hundred.

Every claim keeps its receipt

Each observation carries what it needs to be checked later: when it was observed, which source produced it, that source's reliability rating, the link back to the original item, and the geometry that put it on the map. When an observation becomes an entity in the knowledge graph, those stay attached.

Sources carry ratings on the Admiralty scale, the same reliability and credibility framework we walk through in how to rate an intelligence source. A UN-backed hazard alerting system and a regional outlet with a mixed record are not the same input, and the system does not pretend otherwise. GDACS sits high. GDELT, which is machine-generated and noisy by construction, sits lower, for reasons we get into in what GDELT is and isn't good for.

Where the underlying item has a URL, the operator can click straight through to it. Some hazard payloads do not carry one, and in that case the record says so rather than manufacturing a link. A finding you cannot check is a rumor with better formatting.

What it does not do

The collection layer surfaces findings and the sources behind them. The decision to reroute, delay, or stand down belongs to the team executing the mission. Today the engine does not make that call.

It also does not invent. No entity enters the graph that did not come from a source, and no finding appears without the record of where it came from. That constraint costs coverage in places where nobody publishes anything, and we would rather have a visible gap than a confident guess.

The honest summary of the whole layer: public sources, consumed continuously, cut to the mission, resolved across sources, traceable to the origin. None of those five words is a claim about secret access. All five together are why a team of four can hold a watch that used to take a room.

Arcline builds software for small teams running hard missions with little support. Request early access at arclinesystems.ai.