One run, start to finish
10,034 postings went in. Ten came out. Nothing sent itself.
Every particle above is one real job posting from the run of 18 August 2026. It enters at the left and dies at whichever filter rejects it. The beam's width is not styling — it is the survivor count on a log scale, so the narrowing you see is the measurement.
Where the ten thousand go
Four things happen in sequence, and each one is cheap enough to run on everything that survived the last. That ordering is the whole trick: full scoring parses an entire job description, so it can never run on ten thousand postings. It runs on 784.
| Stage | In | Out | Dropped | What decides |
|---|---|---|---|---|
| Discover | — | 10,034 | — | Six public job boards, deduplicated on company and title during the fetch. |
| Analyze | 10,034 | 784 | 9,250 | Cheap title ranking against target roles. Deliberately runs before anything expensive. |
| Score | 784 | 106 | 678 | Deterministic fit against the evidence vault. No model — a score has to be auditable. |
| Tailor | 106 | 10 | 96 | A résumé built per posting from claims that already exist. Capped per run, not by luck. |
| Your review | 10 | 0 | 0 | Nothing has left this stage. It is where the pipeline stops on purpose and waits for a person. |
Why it stops
The last stage is the only one with no exit condition the software can satisfy. Ten applications sit there until a human opens each one, reads what was written in their name, and sends it. The run log records that state as blocked, which is correct — it is blocked, and it is supposed to be.
10 awaiting your review — nothing submits automatically.
Run 24, approval node, verbatimThis is a structural property, not a setting. The component that fills in an application drives a visible browser and has no code path that clicks a submit control. There is no flag that turns that off. Indeed, Greenhouse and Dice all draw the line in roughly the same place, and a system that quietly crossed it would be worth less than one that stops.