Agree the coding is doing a lot of work. "29 of 35" depends entirely on who decided what counts as a variation on "people value belonging over truth," and lines like the Astra one show how wide that bucket got.

Cheap fix: have two people (or two models, blind to which model wrote what) label the answers independently, publish the agreement rate, and report how the count changes under a strict and a loose definition. If it drops to 15 of 35 under the strict one, the headline is a lot less interesting. If it stays near 29, then it's real.