Source Classification

01

Everyone speaks at once

In a crisis, everyone speaks at once. Local media, eyewitnesses, propaganda accounts... they all land in the same stream side by side, describing the same event with different intents and in different ways.

Listening to this stream without sorting it means getting lost in the noise and picking up false information. This is why source classification is one of the main feeders of the early warning layer: every source's historical record and position is labelled automatically by source classification before it enters the system.

02

Trust is earned, not claimed

A source's reliability is the history of what it has said. What it shared, how it shared it, whether its details were accurate, whether its tone is biased; it is assessed and labelled across these and many similar metrics.

03

A fair labelling process

Labelling accounts correctly first requires a verified set of events to compare against. Building that set means weighing many considerations.

For example, the final version of an event cannot be taken as the baseline: events unfold gradually, details settle over hours, the picture is updated. A source that reported a smaller picture in the first hours cannot be marked unreliable because the picture later grew. Every post is assessed against what was knowable at the moment it was published.

A second example: the set cannot rest on the archives of international agencies alone; the events those agencies never covered would be left out. An event's absence from the major wires does not mean it did not happen; without this distinction, the most valuable local sources are filtered out of the system.

These and many similar possibilities are accounted for from the start as the event set is built. Accounts are assessed against this set and receive their labels; during data processing, the early warning uses these labels to weigh incoming information and to detect attempted manipulation.