❦
Every margin note in the current annotation sets was read and sorted into a small set of categories: what kind of claim in the essay it examines, and whether that claim survives. This page counts the results. The skeptical sets are Opus (Book V), Opus (old) (Books I–IV), Fable, Opus 5.5 and Skeptic Astra. The rebuttal sets are Rationalist Opus and Rationalist Astra. Opus (old) is off by default for readers but is the only Opus pass in Books I–IV, so it is counted there. Superseded passes (Old Rationalist Opus, and Opus (old) in Book VI, where Opus 5.5 replaced it) are left out so that no finding is counted twice. How the sorting was done is described at the end.
Loading…
The most objective question the notes can answer is whether the research the essays lean on still stands. Only studies the essays themselves cite are counted, in a footnote or by name in the text; studies that the notes bring in to argue against an essay are not. Each study is counted once, however many notes discuss it. When several notes disagree, the study takes the verdict most of them give. A study an essay cites that no note examines is shown as not assessed.
The table lists every study, with the verdict each skeptical set gave it. Click a column heading to sort.
Most notes are not about a cited study. They examine unsourced generalizations, history, attributions, predictions about AI, and the reasoning itself. Splitting the verdicts by kind of claim separates two different failures: a study that did not hold up, and an essay that misreported something.
The five skeptical sets were written with different briefs. Fable was told to be as skeptical as possible. Opus (old), in Books I–IV, was a gentler first pass; Opus in Book V and Opus 5.5 in Book VI were maximum-skepticism passes. Comparing their verdicts shows how much of each book's damage depends on who was reading.
In Books I and II, every Fable note has a reply from Rationalist Opus written against the same passage. In Book I, it also has a reply from Rationalist Astra. Because the replies are paired note for note, it is possible to count how often the defense concedes, and which of Fable's verdicts it contests.
The sharpest disagreements are notes where Fable finds the claim damaged, and the reply rejects the objection outright and argues that the claim stands. These are where a reader's judgment is most needed.
The two Rationalist passes answer the same 161 Fable notes in Book I independently. Where they reach different conclusions, the question is probably not as clear as either one makes it look.
Some studies are judged by more than one skeptical set. The list below shows the ones where one set thinks the study stands and another thinks it does not.
Each note in the seven current sets was read in full by a Claude model working from a fixed rubric. The model recorded what the note is about (a cited study, an unsourced empirical claim, a historical fact, an attribution, an AI prediction, the argument itself, or uncredited prior work). It also recorded the note's verdict on the essay's claim, the short name of the study if there was one, and a one-line summary. The studies on this page are a separate list, analysis/essay-studies.jsonl: every study an essay cites in a footnote or names in its text, with the notes that assess it and each note's verdict on that study. For rebuttal notes it recorded whether the reply concedes, partly concedes or contests the Fable note, and the reply's main device. The classifications are in analysis/findings.jsonl. scripts/build_findings.py checks them against the annotation files and computes everything on this page.
These are the annotators' verdicts, not independent ones. A note that calls a study "weakened" is counted as weakened, even if another reader would weigh the evidence differently. The numbers measure what the annotation passes concluded. They are not a ruling on the essays. Studies are matched by name, so the same study written two different ways can appear twice. The classification was done by a model and will contain some errors, but the totals should be stable to within a few points.
❦