Skip to main content

​
Statuses

Every signal is in exactly one of these states. Use disable when you want to pause a signal you might bring back. Use deprecate when you’re retiring it for good. Observations are never deleted by a status change. Disabling a signal stops it finding new matches; everything it already found stays on your deals.

​
What the app calls these

The statuses above are the ones the MCP contract uses. The status pill on a signal card in the app uses shorter labels: So a signal these docs call “disabled” reads as Inactive on the Manage Signals tab. Nothing differs underneath; only the label is shorter.

​
Transitions

Editing configuration always returns a signal to disabled. If you change instructions or document types on an active signal, it stops evaluating until you activate it again. This is deliberate — it stops a half-edited signal running against live data.

​
Versions and concurrent edits

Each signal carries a version and a revision identifier that looks like my-signal-key:1. When you change a signal, you pass the revision you expect it to be at, and the change is rejected if someone else has edited it in the meantime. If you hit a revision conflict, re-list the signal to get its current state and try again. Don’t force past it without looking at what changed.

​
Backfill

Activating a signal only affects artifacts ingested from that point forward. A backfill runs it across artifacts you already have. Every backfill must be bounded. You must give a time window, an artifact cap, or both.
Backfills are queued and run asynchronously. They don’t block anything, and you can keep working while one runs.

​
Watching progress

Listing your signals reports three counts per signal:
  • Eligible artifacts — how many artifacts match the signal’s document types
  • Completed artifacts — how many have been evaluated
  • Observations — how many matches have been found
A signal that’s evaluated 681 of 1,330 eligible artifacts and produced 115 observations is mid-backfill and matching about one artifact in six.
Start with a narrow backfill — 30 days, a low cap — and look at what comes back before running it over everything. It’s much cheaper to fix instructions after 200 artifacts than after 10,000.

​
Deleting a signal

There’s no hard delete. Deprecate is the terminal state, and it preserves the observation history so past analysis stays intact.