← Back to the article

The postings behind “Two in Three Engineering Job Postings Outlast Six Weeks. The Oldest Are Likeliest to Stay”

The public Greenhouse, Ashby and Lever job boards of the same 387 company names were read on 15 September, 13 October and 27 October 2026 (and on 29 September, a read that kept one posting ID per role and is not used here). Only boards that answered on all three days count. A posting qualified when its title described a software, data, machine learning, security, platform, mobile or infrastructure engineering role, with the filter of the September count; on 27 October that was 4,972 of 23,242 postings on 186 boards.

A posting is the board’s own ID: the Greenhouse job number, the Ashby or Lever posting UUID. It is “still up” when the same ID is on the board on a later read. Its age is counted from the published date the board gives that day: first_published on Greenhouse, publishedAt on Ashby, createdAt on Lever. Each read’s values (title, location, level, work mode, date) are kept under that read’s key in postings.json, so a split by level or mode uses the 15 September values for the whole cohort.

Sentences. Every sentence of every 27 October posting that mentions a vacancy, a pipeline, a future opening, several seats, team matching, a deadline or rolling applications was kept, and for postings naming an Ontario location also every sentence about AI near applicants, screening or interviews: 156 distinct sentences, each with the IDs of the postings it appears in (sentences.json). A language model (Claude) labeled every one twice under written instructions, in runs that did not see each other (first, second). They agreed on 145/156. The 11 disagreements were all sentences saying a company uses AI somewhere in hiring without saying it screens with it; they were settled by hand as ai_hiring_general, a label the instructions did not have (labels.json). No person checked every label, which is why every one is printed below.

Every figure in the article, with the checks behind it, is in results.txt, the unedited output of analyze.py. lifetimes.py joins the reads and extract.py picks the sentences. The Ontario and New York rules as read on 27 October 2026, with sources: legal-timeline.md. Long dashes inside quoted text are printed here as hyphens; the data files keep them.

The 15 September cohort

By company

Engineering postings at each company: open on 15 September, still up on 27 October, open on 27 October, and of those how many were published more than 180 days and more than a year before.

Ashby postings whose published date moved

Same ID on 15 September and 27 October, later date on 27 October. The last column is the datePosted in the live page markup when it was opened on 27 October.

Every sentence, with its label

Both runs and the label used. A sentence repeated across a company’s postings is one row.

Every posting

Every posting seen on any of the three reads at the boards that answered all three days. Dates are the published date each read gave; a dash means the posting was not on the board that day. “Ontario” marks a 27 October location naming an Ontario place. Sentence keys refer to the table above.