Stage one: collection
Each source is queried for the companies or topics being watched. This is ordinary engineering: fetch, parse, normalise into a common shape with a title, a link and the date the source itself gives.
The failure here is silent. A source changes its response shape or starts rate limiting, the connector returns nothing, and the night simply produces less. A system worth using reports per source what answered, what legitimately had nothing, and what failed, because those three look identical in the output otherwise.
Stage two: clustering
One event produces many documents. A funding round is covered by a dozen outlets within a day, and each copy looks like a separate finding to a naive system.
Clustering groups documents that describe the same event, usually by comparing overlapping word sequences rather than by exact matching, so a rewritten headline still lands in the right group. Done before scoring, it is what stops a quiet week reading as a busy one.
Stage three: scoring
Each cluster is scored so that only the top slice continues. The important design choice is whether the scoring is deterministic or a model judgement. Deterministic scoring - weighted trigger patterns, source authority, recency decay - is reproducible, explainable and cheap enough to run on everything.
This is also what makes the economics work. If ninety-nine per cent of collected documents never reach a language model, the model cost applies to the remaining one per cent, and the system can afford to read widely.
Stage four: corroboration
For the survivors, count how many independent sources carry the same event, by registrable domain rather than by outlet name, and discount copies that share wording with a press release.
The output of this stage should be a count and a list, not a verdict. Reported independently by three sources is a fact about the evidence. Verified is a claim about the world that no count can support.
Stage five: writing
Only now is a language model useful, and only on a small amount of material that has already been selected and bound to its sources. It writes what happened and what it means for the reader, and the link travels with the claim rather than being asked for afterwards.
The rule that keeps this honest: the model is given the documents and told to cite the ones it was given. It is never asked to remember a URL, because a remembered URL is a fabricated one.