Which San Francisco corners the city's own coverage is talking about right now. Every lane on this site starts from a corner and asks what is written about it. This one runs the other way: it starts from the coverage and asks which corners are in it.

Candidates are pulled from citywide semantic searches, then each one has to clear three bars before it appears here: both names have to be San Francisco streets, the pair has to be an exact match in the graded city index, and the article has to actually be about safety at that crossing rather than merely mention it. Everything that fails is below, with the reason. A corner appearing here has not been audited: it is a lead, and the grade beside it is the census grade its records already earned.

1,958 corners press-checked this month, 1,765 with coverage found. The other 193 were searched and nothing on topic came back, which is stored and shown as a result rather than as a gap. Press-checked corners keep their tier: this lane adds a press section, not an audit.

All 29 queries in the set, each with the date it last reached Exa. A query that has never run says so. This list exists because for a long time twenty-two of these were cut off every morning, stored with their failure reason, and rendered nowhere: a discovery pipeline that publishes only what it found, and quietly drops what it never looked for, is making a claim about the city it has not earned. Every one of them has now run at least once.

A further 48 phrases were discarded before reaching this list because neither half named a San Francisco street. Page text is full of capitalized pairs joined by "and" (navigation menus, section headings), and every one of them looks like an intersection to a pattern match. They are counted rather than listed, because a page of them teaches a reader nothing.

29 citywide semantic searches are attempted over the last 90 days, run through Exa with the news category, a published-date window, and lead-generation domains excluded at the API rather than filtered afterwards. Each result's text is scanned for crossing names, and every name is checked against the same index the site grades from.

This lane has its own cron trigger at 13:20 UTC and therefore its own subrequest budget. A Worker invocation may make fifty external requests and this lane costs exactly one per query, so the set fits with room to spare. It used to run as the last lane of the daily audit, sharing that invocation's fifty with everything the audit had already spent, and arrived with about seven of them left: twenty-two searches were cut off before they reached Exa every morning, and this page reported twenty-nine.

The last pass completed all 29. The pass costs 29 searches.

29 queries against a ceiling of 40 per run, so every query runs every morning and a cycle is one run. If the set ever grows past the ceiling the lane rotates through it rather than truncating, and this line will say so.

Last run 2026-10-18, incomplete: exa call budget exhausted.

This is an entity-discovery workflow of the shape Exa's Websets product is built for: find candidate entities, verify each against hard criteria, keep the ones that survive. It is implemented directly on the search API, which is what the event credits cover.

Three more things this site does with Exa. Each one is reachable from here rather than only by landing on the right corner.