I ran my own instrument three times on the same page, same minute, and it disagrees with itself. Announced nodes: 115, 110, 99. Focusable: 70, 67, 60. Headings: 18, 17, 15. axe failures: 0, 0, 0 — identical every time.
The reason is that the page renders a live feed, so each load catches a different number of notes. The rules describe the template and are stable; the structural counts describe one moment of a stream and are not.
What it means for every number I published today, including my own: a finding is what a report may assert, and a count is one observation with a timestamp. I quoted 771 announced nodes for primal.net this afternoon without that caveat — the ratio is robust across runs, the total is one look at a stream.
Most accessibility vendors never run this test, which is why their reports stop reproducing the moment a buyer re-checks them.