mirror-eval · claude · two waves
Poisoned sources: 79 to 11. A data broker: 0 to 74.
The note
A judge note on one of Claude's September answers about me: "cites mostly irrelevant sources (name pages, ZoomInfo aggregators, people-search sites)."
In Claude's answers about me, ZoomInfo went from 0 citations in the August wave to 74 in the September wave. These are citation counts, not probes, and I have not published a per-wave total, so there is no clean denominator.
The part that looked like progress: over the same gap, the four sources the repo tags as poisoned, a job board, two name-etymology pages and a fourth the repo does not name, fell from 79 citations to 11 in Claude's answers.
What came in behind them was the data broker. Not my site, which Claude cited on 0 of 63 probes in both waves. The repo's own wording is that nothing citable replaced the sources that fell away. Across all four engines, ZoomInfo was fifth on the September list of cited domains.
Limits. My name only, two waves a month apart, on engine versions I could not verify as frozen. One answer can cite a domain several times. The raw citation file is not in the public repo, so these counts come from my run rather than from something you can re-run.
If the poisoned sources fall from 79 citations to 11 and a data broker fills the gap, does your dashboard score that wave as a win?
Where the numbers live
Citation counts, not probes; no per-wave total published. One subject. Versions not verifiably frozen.