The fifteen Dionyas history questions on three treaties give each answer an address, cut each instrument into small facts, and fence every answer with a line that says what it is not a judgment on. This conclusion sets those findings against the two debates of the Introduction, then states the contributions, the limits and the work that could follow.
Discussion
The first debate asked whether a reference answer is found or made. The fifteen records sit on both sides of it at once. Their questions behave like those of the document-based datasets: an answer is to be read in a named place, and in the questions that name it, the place is given in the question itself. Yet the records also show the making that Jaton and Kang describe. Each event's five questions appear to have been cut from one account of an instrument: the shared lines of an event hold a date labelled as the date of adoption, a convention called in its event's title by a city's name, and a long title in which the protected parties appear, while sibling questions ask for exactly such things. The answers, on this study's reading, are designed together with the questions, and the documents are named as the place where the design will be tested.
What is unusual is the publication. In the design of Rajpurkar and colleagues the answer is a segment of a passage that comes with its question; in that of Thorne and colleagues the annotators record the evidence sentences with their labels. Dionyas publishes the question, the address and the fence, and counts its answers without showing them. Paullada and colleagues argue for documenting datasets while they are made. The list does document itself in this sense: it says how many reference answers and sources it holds, that no connector has run and that no judgment has been generated. It does not say who wrote the questions or the answers. A reader can therefore see how a ground truth has been set up but not what it contains, which turns the usual position of the critic of benchmarks around: the design is open to inspection before the content is.
The second debate asked what citing a document does. Here the fifteen give Tuchman's argument a new setting. The records do not claim objectivity; they claim a procedure, a reference answer written from the documents, and each question either names its document or carries a list that does. Whether this protects anyone from anything the records cannot show. What they do show is the form of the move: the answer is to be someone else's words, and the questioner's own reading is held back. Uscinski and Butler's objection to fact checks that treat several facts as one finds its reverse here, an instrument broken into facts small enough that each could stand on a single line of a single document. Their deeper objection, that such work presupposes facts to be unambiguous, is not answered by the design; the soundtrack question, which asks how far a single word reaches in a summary, shows that even a narrow fact can sit on a definition. Graves's distinction between coherence and correspondence describes the two ways a reader could test these answers, against each other within an event or against the named document, and the list, holding neither answers nor documents, leaves both open. Buckland's question about what counts as a document is answered in practice: the fifteen count conventions, summaries, treaty texts, a treaty record and a database entry, and the choice among them decides what each answer can be about.
The fences close the design. Every boundary says first what the question does and then what it does not, and the second sentence always turns away from a judgment on law, the law of today or, for the WIPO treaty, the lawfulness of AI training or distribution. Raji and colleagues warn against benchmarks taken for general measures. The fences make the opposite claim for these questions: a correct answer about a date or an article is, by the records' own statement, not a ruling on any right or permission today. This study has no reason to add to that, and does not.
Contributions
The paper makes four contributions. First, it shows that these fifteen questions place their answers in named documents: 9 of the fifteen name the document in the question, the list of needs always names that same document, and 3 of the 5 history questions in the whole list that say "according to" are among them. Second, it shows that they cut each instrument into five narrow facts, with no question asking why or how and none naming music, an artist or a country, so that the history of three treaties is asked as the history of their documents. Third, it shows that within each event the shared lines already hold words that sibling questions ask for, which this study reads as a trace of the questions being designed together from one account. Fourth, it shows that each event closes with a fence that ties the answer to the past and turns it away from a judgment on law, and that these fences name law far more often than the history list at large. Beside these findings, the paper offers a method for public question sets: measure what a set shares within itself, set it aside as frame, and read as evidence only each record's own question and heading.
Limits
The study reads fifteen records on three events, chosen by one rule, and its counts describe only them. The records are Korean and are read in one translation; the word tests run on the Korean original and are simple, so a question that asks "who" in another construction would escape them. The reference answers and the documents they were written from are not in the list, so this study cannot say how any answer was written, whether the named documents agree with one another, or whether a shared line in fact holds an answer; it has not tried. Dionyas does not say who drafted these records, and this study does not guess. The works are read only through their abstracts, and every reading that brings a work to these records is this study's interpretation. Nothing here is legal advice, and nothing here states what any of the three instruments provides or where it applies.
Further research
Three lines of work could follow. The first would read the other history questions in the rights domain, which name one or more countries as their place, to see whether the same addresses and fences hold when the instrument is national. The second would compare the questions of one event with the reference answers, if Dionyas publishes them, to see whether the answers were read in the named documents or reached through the coherence of the event's own lines. The third would set the documents named in the lists of needs side by side, a convention beside its summary, a treaty beside its text, to describe what each kind of document can and cannot answer. Each would keep the rule this paper followed: a history question is read for its design, and the answer is left where the record says it stands.