1. 01CollectPublic posts and comments from Facebook groups and pages, TikTok, Telegram channels and groups, and X search. Author identities are hashed at capture.
2. 02ClassifyEvery item is classified automatically against a published codebook: one class, a severity, and who is targeted.
3. 03AdjudicateA human adjudicates a sample each round. Disagreements become written rulings in the codebook, and the agreement rate is recorded with the round that produced it.
4. 04AlertText rules — not the classifier — raise alerts: dated ultimatum, threat, call for violence, location exposure, statement attributed to an institution, volume spike.

Public content is collected by an instrumented browser reading as dedicated accounts, read-only: it never joins a group, follows a page or requests membership, and it carries no personal or organisational identity. Why not the platforms’ own interfaces?

1. 01Search on the campaign’s own language. A written, version-controlled list of the slogans the movement names itself with, and of individuals being talked about, is searched on every cycle. Because the list is a file under change control, what was searched in any given week can be reconstructed.
2. 02Links carried inside collected posts. The movement cross-posts between its own groups, so a group that matters is usually named by one already in scope. Those links are extracted and queued as candidates.
3. 03Every candidate is probed before anything is collected from it: public or private, how many members or followers, the date of its newest visible post, and whether its administrators have paused it.
4. 04A candidate is rejected at the probe, without an operator ever seeing it, if it is private, paused, below the size floor, empty, or has not posted in ninety days. It is enabled automatically only if it is public, above the floor, has posted within fourteen days, and carries a name that reads as the Libyan campaign rather than a neighbouring country’s movement or a refugee group discussing resettlement. Everything else waits for an operator’s decision, which is recorded against the source with its reason.
5. 05Sources leave by the same door. One that returns nothing across successive collections is visited less often and retired after sustained silence; the retirement and its reason are recorded against it.
6. 06The list of sources is never published or shown to the public.

The list is not a claim of completeness and nothing here presents it as one. Anything said outside the sources in scope is invisible to this platform: the figures above are what these sources carried, not what Libya said.

Everything a source in scope publishes is collected, not only what matches a word. There is no keyword filter at collection, and that is deliberate: a denominator is honest only if the material that turned out to be irrelevant is inside it. A large share of what is collected has nothing to do with the subject — a road closure, a condolence, a teachers’ strike — and that share is shown above rather than removed. On-topic means the item is about migrants, refugees, the United Nations or the campaign itself, decided by classification against the published codebook and not by the presence of a word.

Who is behind a figure is answered at three levels and never below: the group or page the item was published in, which is named; the number of distinct accounts behind it, which is countable because author identities are hashed at capture; and how many of those accounts appear to have been created recently. A rise in calls for violence against the United Nations is therefore reported as which pages carried it, across how many distinct accounts, how far it concentrated in a few of them, and whether the same text travelled between sources.

It is never reported as a named person. Author identities are replaced by a keyed hash at the moment of capture, and no individual is named in any brief, alert or export. That is a limit on the product rather than an omission from it: naming a person who is being targeted republishes exactly the exposure the service exists to flag.

1. 01Author identities are hashed at capture. The platform holds no identities and offers no way to resolve them.
2. 02No raw export. Subscribers see aggregates, labelled excerpts and alerts only.
3. 03Locations of individuals and organisations are always withheld — the occurrence is shown, never the content.
4. 04Every label carries the codebook version and the human/model agreement rate behind it.
5. 05Post text and author identifiers are deleted 24 months after capture. Only aggregate counts and the human-reviewed sample are kept beyond that.