Page size

50 rows, fixed.

Ordering

Newest first, by the API's own ordering — posted_at DESC NULLS LAST, id DESC.

Canonical URL

Every filtered view declares the bare / as canonical, because a filtered view is the same page with a different slice.

Crawlers

robots.txt asks them not to walk the facet space at all. The canonical is the second line of defence.

Data source

Server-rendered from /public/jobs. The browser never talks to the API directly.

Cache

The page is regenerated every 300 seconds.

The * beside a remote status, and the footnote under the table, are the provenance model surfacing in the UI: the asterisk means we inferred the remote status from the listing text rather than reading it from a field the employer filled in. Salary is shown only where the employer published it. Both are the same distinction the API exposes as `field_sources`.

Picking a value on one side clears it from the other. Allowing and excluding the same value produces a query that can never match — the API ANDs them — and the page would then say “No openings match these filters” with no way to see why. This is the one piece of contradiction-detection the UI does that the API does not; the API returns the empty page faithfully.

Chip state

Solid = included. Struck through in red = excluded. Plain = neither.

Group header count

“n on” for included values, “n excluded” for excluded ones.

Open groups

Category is open by default; any group with an active value opens itself.

Mode in the URL

mode=exclude is UI state and is not forwarded to the API — sending it would be an unknown parameter.

Truncation

Country and City show up to 40 values; other groups up to 30. The full vocabulary is on /public/facets.

Every chip is an ordinary <a> carrying the whole query string, so the panel works with JavaScript disabled and every filtered view is a real, shareable, crawlable URL. That is also why the mode lives in the URL rather than in client state.

Two deliberate absences worth knowing about. The page emits JobPosting structured data, but without validThrough: no live row carries an upstream expiry date, and inventing one would publish a date the ledger cannot back on the one surface where a wrong expiry actively misleads a jobseeker. Instead a role that ends is delisted and its page stops being served. And when there is no advert text there is no structured-data block at all rather than an empty one.

Where the source stated no posting date, the row is labelled First seen with the note (the employer did not state a posting date) rather than showing a date under a “Posted” label. That is the same honesty the API expresses by leaving posted_at absent and populating first_seen_at on every row.

The advert body is shown as received, truncated at 12,000 characters. Job pages are indexable by search engines and closed to answer engines — see the crawler policy.

The directory shows employers with roles this site can actually serve, so a count on a company row and the roles you can click through to reconcile. Companies seen hiring only under a source we do not redistribute are not in the website directory at all; they are reachable through the API with include_discovered=true.