Jev jobs

Build a Dive over the public share
md:_share/jev_jobs_public/b4600052-1467-4f33-87a3-7474c5644ca0 (attach it as
jev_jobs_public). It holds one table, `jobs` — 10,000 job postings with job_id,
listed_date, company, title, location, description — and a macro
prompt_jev(input, instructions, type := 'noul', criteria := NULL) that answers a
question about a text: a probability when type is 'noul', one label from
`criteria` when it is 'choice'.

Show the jobs as a table, 50 a page, newest first: listed date, title and
company, location, then one column per question.

Two actions above the table. "Extract features" runs five fixed questions over
the 50 jobs on the page in one statement — is this a data role (noul), role
family (choice), seniority (choice), does it state a base salary (noul), does it
manage people (noul). Render a probability as a percentage with a small bar, a
choice as a label chip. "Ask" takes one free-text yes/no question, runs it as a
noul over the same 50 jobs, adds a Match column, and sorts in SQL with
ORDER BY try_cast(matches AS DOUBLE) DESC NULLS LAST; add a "matches only"
checkbox that keeps rows at 0.50 and above.

Every click is LLM spend, so confirm first and state exactly how many
prompt_jev() calls it costs. Paging clears both results. Clicking a row opens
the full description with its per-question answers.

Style it quiet and editorial: off-white paper, near-black ink, one accent color,
no card chrome, small uppercase labels.