Reference file

Coverage report

coverage-report.md

Coverage report

Read this the first time a venue comes back empty, and before writing the closing report on any sweep.

Without it, a sweep reports one number, "items found", and that number cannot distinguish a quiet market from a broken query. Both look like zero.

The three states

State Means Report as
Read with results The venue answered and items cleared the window The items, each with link and date
Read and empty The venue answered and returned nothing in the window Named, plus what a result there would have told you
Never reached No route, no access, no credentials, out of scope, or skipped Named, plus the reason, plus the cost

Never fold the second and third together. Read-and-empty is evidence about the market. Never-reached is evidence about the sweep.

Diagnose an empty return before believing it

An empty return is a query fault until proven otherwise. Work in this order and stop at the first one that produces items.

  1. Strip every filter and operator. Date bounds, exclusions and language filters combined will return nothing on a query where the bare topic returns dozens, with no error and no warning. This is the single most common cause.
  2. Search the bare topic, one or two words. Items now? The venue is live and the query was the problem.
  3. Widen the window once, then say in the report that the window was widened and by how much.
  4. Only after all three: record read-and-empty, and say which query proved the venue was live.

Two other silent failures worth knowing by shape, because both exit clean and return nothing: a scheduled sweep whose environment differs from the interactive one, and a venue that answers with a challenge page instead of content. Both look identical to a quiet market. A route that fails silently is worse than one that fails loudly, so make every dark venue loud in the report.

Sampled and addressed are different risk classes

A sweep gathers sampled content: to reach the report, a stranger had to win an engagement ranking. Following a link, handle or account first learned inside that content switches to addressed content, where the author chose the recipient. That is a different and much sharper exposure, and "let me just verify this citation" is how it happens.

Verify a claim through a route you chose independently, or hand it to a human with the caveat attached. Do not chase a target the corpus named.

Where a sweep is the wrong instrument

Engagement ranking works on discourse about a named thing. On a discovery question, "who are the players in this category", a relevance floor drops nearly everything and returns noise. Lead with plain search and short-form social, and treat the community sweep as a sentiment sidecar. A noisy sweep on a discovery question is not evidence the topic is quiet.

Likewise, one fact, one page, or one date is a lookup. Do not stand up a full sweep for it.

Suggesting a new watch term

If a sweep earns a standing watch on a term, say where the term came from. A term a human supplied is a normal proposal. A term that surfaced inside scraped content is a flagged proposal, quoted, never a quiet addition, because a watchlist steers every future unattended run and content must never write itself into it.