Glossary

Adapter : A concrete implementation of a port — the Telegram notifier, the RSS collector, the MongoDB repository. All input and output lives in adapters. See Ports and adapters.

Alert tier : How strongly Watcher backs an alert: strong (a confirmed alert), weak (labelled “unconfirmed / possible update”), or none. See Confidence and tiers.

Candidate : One collected item — a URL, the readable text found there, when it was fetched, and a trust tier. Candidates are what the model reads.

Cascade : The evaluator arrangement where a primary model’s verdict is checked by a second-opinion model from a different family, and where the secondary also acts as a failover if the primary is unavailable.

Channel : A delivery route for alerts: telegram, push, or email. Note that email is not implemented as an alert channel and logs to the console instead.

Collector : The component that turns a source into candidates. One per source type.

Condition : The plain-English sentence describing what would count as news for a watch. See Conditions.

Confidence : The model’s own 0-to-1 score for a verdict. It is not a calibrated probability; treat it as a dial.

Corroboration : How many distinct sources contributed fresh evidence for the same event. A watch can require a minimum with minSources.

Due : A watch is due when it has never run, or when the next time matching its cron expression after its last run has passed. See Scheduling.

Embedding : A numeric vector representing an event’s meaning, used to recognise the same story described in different words.

Evaluator : The component that answers “is this condition met?” for a set of candidates.

Event fingerprint : A stable identity for a notified event, computed from the evidence URL plus the normalised quote and date — not from the model’s summary.

Evidence quote : The exact sentence the model copied from a page to prove the condition. It must be findable in the collected text or the alert cannot fire at full strength.

Failover : Using the second-opinion model to produce a verdict because the primary model was unavailable. Such verdicts are capped at the weak tier.

Grounded : An evidence quote is grounded when it actually appears in the text Watcher collected.

Health alert : A ⚠️ operational message — a watch whose sources all failed, a crashed run — sent to the admin chat, and visually distinct from an event alert.

Key pool : A rotating set of API keys for one provider and use case. A rate-limited key cools down and the pool advances; an invalid key is benched.

Known state / event ledger : The dated list of the last few alerts for a watch, shown to the model so it does not re-announce something you already know.

Owner : The account a watch, trust rule, device token or run record belongs to. The API scopes everything to the signed-in owner.

Port : An interface the application layer depends on — Collector, Evaluator, Embedder, Notifier, Repository and others. Ports are how the core stays free of I/O.

Pending delivery : An alert that fired but could not be delivered on any channel, stored so the next run can replay it.

Run outcome : The single word recorded for every run, saying which stage it reached and why it stopped. See Run outcomes.

Run record : The persisted history entry for one run: outcome, summary, evidence, source errors, collection counts, models consulted, and any label you added.

Run state : What a watch remembers between runs: seen content, notified events, recent embeddings, the ledger, the last run time, and pending deliveries.

Source : Where to look — an RSS feed, a page URL, or a search query, plus optional collection hints.

Threshold / weak threshold : The confidence bars for a confirmed alert and for a labelled “unconfirmed” one.

Trust tier : A source’s credibility grade: official, trusted, unknown or low. Only the first two can support a confirmed alert. See Source trust.

Watch : The unit of monitoring: sources, a condition, a schedule, channels and thresholds. See Watches.