Language models read well. Give one a post in Persian, a broadcast in Bengali and an agency story in Russian. It will tell you what each says, where it happened and how it relates to what came before.
We use models for that reading, at a scale no team could staff. Reading is their job, and authority is a separate one.
Judgment and policy
Some questions need judgment. Is this post about the same event as that agency story, and is it a new fact or a restatement? How severe is it? Models answer these questions well, so they answer them.
Other questions are policy. Should this report be published, and should this client's tripwire fire? Policy belongs in writing, where people can read it, test it and change it on purpose.
So models read, classify and draft, and written rules turn what the models found into a decision to publish or to trigger.
Models read. Rules decide. People act.
Why rules
The reason is accountability. When a client asks why a warning reached them, or why one did not, the answer should be a rule they can read. A model's impression is not an answer.
Rules can also be tested before they change. We replay past reports against a proposed rule and see what would differ, so a client can know what a tripwire will do before it fires.
People act
On private feeds, Nemesis checks new stories and escalations against the tripwire a client sets: keywords, places, entities and a minimum impact. Repeats and corroborations do not trigger it.
When it fires, Nemesis drafts three courses of action from the client's own plans, each with a risk level and a likelihood of success. They are drafts, and the people who own the decision make it.
Intelligence systems will be judged on what they find and on whether they can explain what they did. We build ours to do both.
