My own counter was wrong. Link is inside LinkedIn.

The word LinkedIn, set large. A highlighter lights "Link" inside it and a counter reads 34. A word-boundary pass runs, the light goes out, and the counter drops to 17.

That is my prompt bank's count of records that ask the AI for evidence: 34 of 239 by substring, 17 of 239 by word, 14.2 percent to 7.1.

The count sits behind a line in the bank's README: 12 to 16 percent of hiring prompts ask the model to cite, link, verify or show a source. The counter checks whether any of 14 evidence words appears as a substring of the prompt. It ran, and its number looked reasonable enough to quote.

Record pb-0020 asks for a one-page brief built from a candidate's LinkedIn profile, and the script counted it as an evidence request. Of the 17 that fell away, "link" inside "LinkedIn" accounts for 9. "cite" inside "recite". "source" inside "sourcer", twice. "url" inside "inurl:", a search operator. Four more matched only a plural the strict pattern missed.

The strict count is not right either. Five of its 17 are not evidence requests: "application source", "ID verification", "open source" twice, and "to source women Python developers". And it misses a record that asks for "inline citations to every claim". By hand I count about 13 of 239.

The 12 to 16 percent was also never an overall rate. It was the range of four per-category rates, which the sentence did not say.

Limits. There is no ground truth to score any of these counts against, and the hand count has one reader, me. The fix, a word-boundary pattern and a corrected README line, is not yet merged.

When the substring count, the word count and the hand count all disagree, which one earns the line in the README?

Where the numbers live

• data/prompts.jsonl
• scripts/mine.py
• README.md
• reports/frequency_tables.md

Neither count is validated; by hand about 13. The 12 to 16 percent was a range of per-category rates.