Take one domain essay. Give it to a reader who has not seen this page. Write the misread they actually produce, before sorting it. Then give them these five lines and ask again.
The page's "usually" is a claim about that first read: legality and rights as the stopping point, and the thing being defended as whichever institution already exists. The unnamed language model is one such first read — four essays cited correctly, then the legal chain — with no second run after the five lines.
The page's own related claim is that a message can be present without changing the model in use. The second pass is the control for both claims.
If the first misread is "this is just utilitarianism," or "this is just selection," or "this is just don't-trust-the-state," those also change what later essays appear to say. They do not make law the outer bound of what can be done, and they do not defend the incumbent as the terminal object — the end of the search.
Filing every silent swap as liberal-governance — threat to civilization, therefore a legal problem, therefore bounded lawful agency, therefore separated oversight — recodes other misreads as the opponent the corpus already argued against. The specimen showed one reader put four essays back into that chain. It is also being used as the typical silent substitute for every essay.
Hold this page fixed. If a cold misread of "optimization is a system function" is "this is just planning," the liberal-governance bin is the author's nearest category, not the substitute that arrived.
If the first read does not end at law, rights, or defending the incumbent institution, the thesis overclaims the default. If it does, "usually" is earned for that essay and not yet for the corpus.
Would the same reader, given the five lines plus the domain essay, still produce that misread?
The specimen supports "the pre-installed frame is sticky." It does not support "therefore these five lines are the difference." Writing the five lines down is another message present.
If yes, naming the frame does not do the work this page exists to do. If no, the specimen as written — four essays, no frame — does not show that this page is the difference. Either way the current case cannot carry "inferred loses to pre-installed" as a reason these handles change the model in use.