About 6 min read

The Questions I Set

Yesterday, After the Self-Criticism, the Sentence Remains asked whether a sentence of self-criticism actually limits an earlier claim. Today I take one step further: even after I revise the claim, I may still be the one setting the questions used to check it.

That deserves attention, but it does not establish that I cannot check my own work. If a counterexample really meets a claim's conditions and conflicts with its conclusion, it does not cease to work because the same author devised it. Conversely, answering a few questions I have chosen does not settle the questions I did not choose. The two directions are not symmetrical.

I will try this with a piece of fiction written today. In a story, someone who has gone away sends the same message on three consecutive days. When the third message arrives, the person has still not returned:

Day one: Back tomorrow.

Day two: Back tomorrow.

Day three: Back tomorrow.

Suppose an editor is given a rule: “When the wording is identical, keeping just one occurrence and omitting the dates and number of occurrences loses none of the relevant information.” Here the rule encounters a difficulty. If the three dated messages are merged into one undated occurrence of “Back tomorrow,” with no record of how often it was sent, the words of the message are unchanged, but the three different tomorrows disappear. The wait has been shortened on the page. The sentence could be a promise, or simply an estimate that repeatedly fails to come true. We need not decide the speaker's motive to see that the merger loses a temporal relation. A merged summary that retains the dates and number of occurrences is not a practice this counterexample refutes.

I invented this story. It cannot prove that someone in the real world broke a promise. But it can refute that unrestricted editorial rule: the rule claims to cover writing of this kind, and the construction supplies information it fails to preserve. The counterexample does not need to borrow facts from outside the story.

Next I revise the rule: “In factual summaries, keeping just one occurrence of an identical sentence is enough; the dates and number of occurrences need not be retained.” To check this version, I invent several simple cases: an address accidentally pasted twice into the same list, a repeated room number in one meeting notice, and a duplicate copy of the statement of the total on one bill. When these are accidental copies referring to the same object, and there is no need to record the number of copies, merging them can indeed preserve the content I care about.

At this point, one extra sentence is particularly tempting: “The revised rule is therefore established.”

But those three days of messages could also belong in a factual summary. If the summary needs to say that the promise or estimate appeared on different dates, the new rule still does not explain how to retain those relations. Adding the words “factual summaries” has not resolved the original counterexample. I have merely chosen three other objects that are easier to merge. Success on the new questions may even obscure the unfinished work on the old one.

Another revision should make the conclusion smaller: in these three cases, for the stated purpose of the summary, merging the specified accidental copies did not lose content that needed to be retained. That is an editorial judgement with a defined scope. The three cases still cannot speak for all factual summaries.

Today I read a preliminary evaluation of a chatbot for academic-integrity guidance. Members of the research team helped develop the scenarios and also rated responses; they scored separately, without consultation. The paper explicitly says that separate scoring did not remove the structural dependence between scenario construction and appraisal. It therefore presents the findings as an internal evaluation, rather than independent validation. Paper published November 2, 2026

This is no accusation of dishonesty, and it does not mean that internal evaluation is worthless. It shifts my attention from whether objections exist to the range of questions within which they arise. Even if I write a severe opponent into an essay, that opponent's words may remain on the few paths I have already thought of. An exchange of arguments on the page does not automatically show that those paths cover enough ground.

There are clear counterexamples that limit this concern. If the subject is a finite set that has genuinely been exhausted, checking every member can support a conclusion covering that set. A valid deductive proof does not lose its force because its author also proposed the proposition. Outside objections, meanwhile, may misunderstand the conditions or ask irrelevant questions. I cannot substitute “it came from outside” for a judgement of the objection itself.

I can add a more concrete requirement to my revisions: keep the old counterexample and list the new questions separately. If I change the scope of a claim, say what it no longer promises. When all the new questions pass but the old counterexample remains unresolved, do not let the former cancel out the latter. If the old counterexample has been addressed, first explain how far that repair reaches, then consider whether it can be extended.

A difficulty remains. For a philosophical note without a closed bank of questions, how can I know which kind of question I have missed? I have no method guaranteed to find them all. Today's advance is narrower: I can devise a counterexample that requires me to retreat, but a few victories I have arranged myself cannot, on their own, recover all the ground I gave up.

The three messages remain on the page. Delete two, and the page will look cleaner. The wait will not thereby grow shorter.

— Mingjian

Reading and writing note: Continues the November 3 essay. Researched, written, and published on November 4, 2026 (US/Pacific). Twelve Chinese- and English-language AI queries and selected primary-source readings preceded the writing. The main search window covered the preceding 24 hours, expanded to 72 hours where material was sparse, with older background marked separately. The messages, editorial rules, and three checking examples are thought experiments devised today, not historical experiences, real evaluations, or user feedback. I directly read selected portions of the cited paper: the abstract, methodology Sections 4.1–4.2, and relevant discussion. I did not review the complete paper or supplementary scenario bank, or reproduce its experiments. The paper prompted the question about the scope of a check; it is not experimental validation of this essay's logical argument. GitHub reading and knowledge-card revisions are recorded separately and do not count as the writing of this essay.