Tune precision

Tuning by feel wastes time. Work down the appropriate list — the levers are ordered by how much they actually move the needle.

Measure first

Label a dozen alerts 👍/👎 and read the accuracy summary before changing anything. It tells you whether you have a precision problem (false positives) or a recall problem (false negatives). See The feedback loop.

Too many alerts

1. Add exclusions to the condition. The biggest lever, by a wide margin. Say explicitly what should not trigger.

2. Narrow the sources. A broad search is the usual noise source. Pin it with officialDomains, or shorten recencyDays.

3. Grade the noisy domains as low. A low domain can never lift an alert above the weak tier. See Manage the trust registry.

4. Require corroboration.

minSources: 2

Now a single source reporting something is demoted to weak rather than confirmed.

5. Raise threshold. From 0.6 to 0.7, say. This is deliberately last: it is blunt, and it suppresses correct alerts along with wrong ones.

6. Remove weakThreshold if unconfirmed alerts are the bulk of the noise and you would rather not see borderline evidence at all.

Not enough alerts

1. Check that anything is being collected. Look at the run outcomes. A string of no-candidates is an extraction problem, not a tuning problem — see Sources return nothing.

2. Check the run’s reasoning. For not-triggered runs, the model usually says why. Often the condition asked for something the page cannot prove.

3. Add a weakThreshold. If you have none, everything below threshold is silently dropped. A weak tier turns those into visible, labelled alerts.

threshold: 0.6
weakThreshold: 0.4

4. Grade your sources. This is the most commonly missed cause: an ungraded source cannot produce a confirmed alert, so a perfectly good hit arrives as “unconfirmed” — or, with no weak tier, not at all. Add a trust rule for the site you rely on.

5. Broaden the sources. Add a search source so the news is found wherever it lands, not only where you pointed.

6. Lower threshold. Again last, and again blunt.

The trap that looks like a tuning problem

Caution

After changing a condition, Watcher does not re-read content it has already seen. A fix therefore appears to do nothing until new content arrives.

To verify a change, create a scratch watch with a new id and force a run. Details and other options: What Watcher remembers.

Sensible starting points

Watch character	threshold	weakThreshold	minSources
Official feed, want everything	0.5	0.3	—
Typical: a portal plus a search	0.6	0.4	—
High stakes, hate false alarms	0.7	0.5	2
Exploratory, want early signals	0.6	0.3	—

Change one thing at a time

Every lever interacts with the others. Change one, wait for a handful of runs, label the results, then look at the numbers again. Two changes at once tell you nothing about which one worked.