Watching a source over months
Continuous monitoring — the agent researches which pages to watch, you confirm, and it reports only what actually changed and actually matters to you.
A follow-up fires once. Some questions never stop: has the regulation changed, has the price moved, has that approval been withdrawn. Answering them means reading the same pages every day for months and noticing the one day something is different — work no one does reliably, because most days there is nothing to find.
Tracking is that job. You describe what you care about; the agent works out which pages to watch, you confirm them, and from then on you hear from it only when something changed and that change concerns you.
Setting one up
You say what you want watched, in your own words: "changes to Brazilian glucose-meter regulations." That sentence is kept verbatim — it is what the relevance judgement reads on every round, not a summary of it.
The agent searches for candidate pages rather than recalling them. This distinction is the whole reliability of the step: a recalled URL that is one character wrong is nothing at all, while a URL picked out of real search results necessarily exists.
Every candidate is fetched once, and the first three entries it extracted are shown to you. If they read like menu links rather than announcements, the source is wrong — and this is the only moment that is obvious. A wrong source never errors; it just quietly watches a navigation bar for weeks. The agent looks at those samples too and flags the ones that look like navigation.
The first run records what is already there and reports nothing. Without that, day one would arrive as two hundred "new" items.
After that it checks on your schedule. Most days nothing happens, and on those days you hear nothing.
The confirmation step is not a formality — it is the only moment you can catch a wrong source.
You can also skip the research entirely and paste the pages you already read. That path calls no model at all — you already know where to look.
What counts as a change
Three kinds, and they are not the same event:
- A new entry appears on a list page.
- The text of a page moves. Timestamps, visit counters and session nonces are normalised away first; they are the entire source of false alarms for this kind of page.
- Something disappears or is marked repealed. Watching only for additions is half a mechanism: a rule being withdrawn while you still follow it costs more than missing a new one.
A change is not a report. Each one is judged against your sentence, and only the relevant ones reach you.
A site redesign is not a regulatory update. When most of a page changes at once, that is treated as a layout change and flagged for confirmation rather than sent as news — the most embarrassing false alarm this feature can produce is a serious-looking report about someone's new front end.
What arrives
A report, not a notification. It carries, for each change:
- the original wording, untranslated — you will take this to a lawyer or a supplier, and a translated sentence does not count in that room;
- the analysis in your language;
- publication date and effective date on separate lines. Where the text only says "within 90 days", no date is computed and the original phrasing is quoted instead. When that date is legally binding, arithmetic is not the right tool;
- a link back to the earlier report if this amends something you were told about before. That link is the difference between an alert feed and actually tracking something.
Reports appear in your conversation with the agent, and by email if you ask for it. Each change carries a button to file it into your knowledge base — into the review queue, never straight in. What was found is still someone else's page, and letting a human press that button is what makes it yours.
The quiet weeks are the hard part
A monitor that finds nothing looks exactly like a monitor that is broken. Two things address that.
You can run a check by hand, any time. The receipt says what it saw — how many entries each source returned and the first few of them — not just whether anything changed. On a quiet day "nothing changed" tells you nothing; "read 34 entries from ANVISA, here are the newest three" tells you it is working.
The weekly digest lists what was seen and judged irrelevant, by title, and invites you to correct it. False alarms are visible by nature — you see the bad report and say so. Misses are invisible: you simply never hear about the thing you needed. Printing the titles is the only way a miss can surface at all, so those numbers come straight from the run records and are never written by a model.
Every source is opened once while you are setting it up, and you see what came back. Checking later is too late: a page we cannot read looks exactly like a page where nothing was published, and a fault that is indistinguishable from normal is the worst kind.
Some pages need more than a plain request. If a direct fetch does not return the page, the setup step
escalates — through a proxy, then a real browser — and remembers which one worked, so the same page is
not re-tried from scratch every day. Two cases are called out rather than hidden: a site whose
robots.txt asks crawlers away is still watchable (you are asking us to check a page you can open
yourself, once or twice a day, not to index it), but you are told; and a site that blocks us outright
is reported as blocked, naming what is doing the blocking, because "we could not read it" is
something you may want to act on.