Core concepts

Seven ideas explain almost everything Watcher does. None of them require reading the source code, and each page here ends with a link into the deeper technical documentation if you want it.

WatchesThe unit of monitoring: sources, a question, a schedule, a bar.

SourcesRSS, pages and search — what each is good at, and how text is extracted.

ConditionsThe plain-English question a model answers about your content.

Confidence and tiersConfirmed alerts, unconfirmed alerts, and silence.

Source trustWhy a site Watcher does not recognise cannot confirm anything on its own.

DeduplicationTwo layers that stop one event becoming five alerts.

SchedulingCron in UTC, due-filtering, and your real check frequency.

AccuracyHow Watcher tries not to lie to you, and where it still can.

The shortest possible summary

On every run, for every watch that is due, Watcher:

1. collects text from your sources,
2. discards anything it has already seen,
3. asks a model whether the remaining text satisfies your condition,
4. checks the answer against the collected text — the quote must be real,
5. decides a tier from confidence, corroboration and source trust,
6. suppresses duplicates of events it already told you about, and
7. delivers what survives.

If a step drops the candidate, the run ends there and records why. Those reasons are the run outcomes, and they are the fastest way to diagnose anything.