Run outcomes

Every run of every watch ends with exactly one outcome, recorded in the watch’s run history. Outcomes are the fastest diagnostic in the system: they say which stage the run reached and why it stopped.

Outcome	What happened	What was stored	What to do
disabled	The watch is switched off, so nothing was collected or evaluated.	Unchanged, apart from the run record.	Enable the watch if you expected it to run.
no-candidates	Sources ran without error but returned no items at all.	Unchanged.	Check the source still returns content — an empty feed, an empty search, or a page that extracted nothing.
collection-failed	Every source that was tried threw an error.	Unchanged, and the last-run time is deliberately NOT recorded, so the watch stays due.	Look at the recorded source errors. A health alert is also sent for this outcome.
nothing-new	Items were collected, but every one had already been seen on a previous run.	Unchanged.	Nothing. This is the normal outcome on a quiet day and costs no model calls.
not-triggered	Fresh content was evaluated and the model decided the condition was not met.	The fresh content is marked as seen so it is never re-evaluated.	If you disagree, the condition or the sources are usually the problem, not the threshold.
below-threshold	The model triggered, but the result did not clear the bar — either low confidence, or a quality check failed and there was no weak tier to fall back to.	Fresh content marked as seen.	Check the run record: an ungrounded quote or an ungraded source demotes an otherwise confident verdict.
already-notified	The triggered event has the same identity as one already alerted on.	Fresh content marked as seen.	Nothing. Deduplication working as intended.
semantic-duplicate	The event is a reworded version of one recently alerted on (embedding similarity at or above 0.92).	Fresh content marked as seen.	Nothing, unless you believe it was genuinely a different event.
notified	An alert fired and was delivered on at least one channel.	Seen content, the event fingerprint, the event embedding, and the event ledger are all advanced.	Read the alert. Optionally label it to feed the accuracy dashboard.
notify-failed	The alert was real and passed every check, but every configured channel failed to deliver it.	The formatted alert is stored in a bounded pending-delivery queue and replayed on the next run — no model call is re-spent.	Fix the channel. Delivery is retried automatically for up to five attempts.

Reading them as a sequence

The outcomes map directly onto the stages of a run, in order:

collection-failed / no-candidates   ← stage 1, collect
nothing-new                         ← stage 2, keep only new content
not-triggered                       ← stage 3, evaluate
already-notified                    ← stage 5, exact deduplication
below-threshold                     ← stage 5, tiering
semantic-duplicate                  ← stage 5, semantic deduplication
notified / notify-failed            ← stage 6, deliver
disabled                            ← never entered the loop

So an outcome tells you not just what happened, but how far the run got before it stopped.

The three you should act on

collection-failed

Every source errored. Nothing could be judged. A ⚠️ health alert is also sent. The watch stays due and retries at the next opportunity. Look at the recorded source errors — a moved URL, a portal outage, or a TLS problem. See Sources return nothing.

notify-failed

A real alert could not be delivered on any channel. It is queued and replayed on the next run, up to five attempts, with no model cost. If you see this repeatedly, the channel is misconfigured rather than flaky. See Alerts are not arriving.

no-candidates

Sources ran without error but returned nothing at all. Usually an empty search, a feed that moved, or a page whose text could not be extracted. Repeated no-candidates on a page source is the classic symptom of a JavaScript-rendered page.

The ones that mean “working as intended”

nothing-new on most runs is the healthy steady state. already-notified and semantic-duplicate mean deduplication earned its keep. not-triggered means the model read new content and correctly found nothing.

Where to see them

• Panel / app — a watch’s run history, newest first, with reasoning.
• API — GET /watches/:id/runs.
• CLI — pnpm watches history <id>.
• Scheduled runs — one line per watch in the job log.