444 collected, 239 kept. The 205 rejects are public.

A stack of 444 record sheets builds up. A slice of 144 slides off, then 41, and the last 20 slide off and turn red. 239 remain.

Those 20 of the 444 share one reason: text not found. They are prompts, search queries and screening criteria I collected from public pages on 28 September for a bank of what hiring-side people and tools ask AI to do.

Reject 187 is one of the 20, and a good prompt: "Can you provide an independent source confirming your [degree/title/headline result]? Who issued this credential?" The page it was attributed to did not contain it when I checked.

Every dropped record carries a reason in a public rejects file. 144 were out of scope: mostly real hiring uses of AI, like interview-question generation or outreach drafting, that the bank does not track, plus a few sales-prospecting prompts that were not hiring-side at all. 41 were duplicates of a kept record.

Every record in the bank has a URL and a flag saying whether the text was found on that page on the collection date. The 239 passed that check. What the check cannot see is the gap the 20 exposed: text I had recorded against a page that did not hold it.

Limits. "Out of scope" is my call, recorded per record but still a call. The 444 is arithmetic, 239 plus 205; no file holds the raw pull. The page checks ran once and were not repeated.

Which would you trust more: a 239-record bank that publishes its 205 rejects, or a larger one that shows only what it kept?

Where the numbers live

• README.md
• data/rejects.jsonl
• data/prompts.jsonl
• reports/summary.json
• data/sources.md

One collection pass. Out of scope is a recorded call, still a call. Page checks ran once, on 28 September 2026.