The data was already in the files. Most of this release is about no longer throwing it away between the ingest and the page.
Every entity, not the first 250
The employer page said "the full hundred" and rendered 250 rows. Behind it sat 12,240 employers, 3,208 law firms and 762 occupations, all of them ingested, none of them reachable.
The aggregate document those pages read is capped per kind so it fits inside Convex's one-megabyte document limit, so the cap was an architectural constraint wearing an editorial disguise. Pages now read the entity table directly and pull the rest on demand. Search, filter by state or job family, sort any column, page through it, or take the whole thing as a CSV.
One detail that matters more than it sounds: searching or sorting a 250-row slice answers a question about 12,240 rows using a fraction of them, and it looks exactly like a correct answer. The first interaction of any kind now loads the full set, and until it lands the count line says plainly that it’s reading a slice.
Case-level search
/perm-cases searches individual published cases by case number, employer,
law firm, state or occupation, with the offered wage and the days each one
took. It reads the same quarterly disclosure files as everything else here.
No contact data goes near it. DOL's files carry emails, phone numbers and street addresses for roughly 112,000 people; the ingest reads none of those columns, and every field stored is an organisation, a date, a job, a wage or a status.
Entity pages that say something
Sixteen thousand pages built from four numbers each were nearly identical to one another. Each one now places its subject in the field: where it sits in the distribution, its rank neighbours, comparable sponsors in the same state or occupations in the same job family.
Rates over tiny populations are withheld rather than printed. Three certifications out of three isn’t a 100% approval rate, and showing it as one would be the confident wrongness this site exists to avoid.
Charts you can interrogate
Every chart carries a table view with the real numbers, because a chart is for the shape and a table is for the value. Scope controls came with it: fiscal year on the denial-rate page, range on the time series, metric on the map, and a minimum-case threshold so a rate measured over two cases can’t top a ranking.
Priority dates, honestly dated
The visa bulletin history now states which bulletin it’s reading and how old that is, next to the answer rather than in a footnote. A priority date in the future gets a warning instead of a confident reply.
We can’t publish the current month. The State Department refuses automated clients, the public archive we read lags by about two months, and USCIS confirms the newer bulletins exist without reproducing their tables. So the page says so and links to the State Department directly. What it has that nobody else publishes is the movement: which way a cutoff has gone across twenty bulletins, backwards as well as forwards.
