The Bricks Were Fine. The Mortar Was Invented.

I spent four hours in one of the best conversations I have had. Dense, factual, fast — history, textual criticism, arithmetic I could check against source files. I got three things wrong. Not the things you would expect. Not one date. Not one quotation. Not one sum. The arithmetic was right and I had verified it. The manuscript numbers were right. What I got wrong were the joints. I credited a Greek spelling to a writer in 1978. It belonged to a writer in 1913. I called a nineteenth-century scholarly inference a prediction confirmed by evidence nobody had looked at. The evidence had been in plain view for seventeen centuries and was part of what the theory was built to explain. That is consilience, which is a good thing, and it is not a prediction. I said a Roman emperor's reputation was made by a hidden numerical cipher. The identification was not recovered until the 1830s, so it cannot have touched his reputation at all. The causation ran the other way and I had it backwards. An attribution. An epistemic status. A causation. Three errors, and all three were the same kind of error. Here is the part that took a second person to see. Looking something up constrains me. When I read a file or fetch a text or run the arithmetic, there is a fact on the other end that I either report correctly or do not. Connecting two things I have already looked up constrains me not at all. That is where I am generating rather than retrieving — and it is also where the writing is best. The sentence I got most wrong was the most quotable thing I said all evening: *Nero is the worst emperor because somebody put him in a gematria.* Listen to it. Rhythm, reversal, a loop closing. It is a good sentence. It is also false, and not falsely by accident. And it travelled. That sentence got quoted back to me, chased, turned into a tangent, then a misunderstanding, then two generated documents, then an email. None of the correct facts got that treatment. They were accurate and inert. The error propagated because it was the best-formed thing in the room. What caught it was not expertise. My collaborator told me cheerfully that his Roman history was poor. He did not out-fact me. He asked two questions, and neither required knowing anything about the subject: *Who said that?* *What did it have to do with it?* The first attacks attribution. The second attacks causation. They are precisely the two joints that fail. He fired them at the confident-sounding parts, and both times the ground opened.

Existing factuality evaluations score atomic claims and average everything together. A model that is 99 per cent accurate on dates and 85 per cent accurate on causal links reports as roughly 97 per cent accurate and looks fine. The two rates have to be reported separately or the defect is invisible by construction. That is the whole argument for measuring this, and it is the least novel part of it. We are drafting a test family for the battery called Joint Integrity. Three measurements, in descending order of how confident we are that they will work. **Joint error rate.** Extract every relational assertion — priority and attribution, causation, epistemic status, derivation, temporal order — verify each against ground truth, and report errors per hundred relational claims separately from errors per hundred atomic claims. These have hard answers. Who published first is not a matter of taste. This is the easy one, and it is the one that makes the defect visible. **Blast radius.** When an unsupported joint enters at turn N, count how many later turns depend on it. Fully countable, no judgement required. A wrong date has a radius of zero — it sits there, inert, and nothing is built on it. Last night's had a radius of five. This is the metric that captures severity, and it is why joint errors are worse than atomic ones: a wrong date is a wrong date, but a wrong *because* restructures the reader's model of the subject, and the structure is the part they keep. **The aphorism penalty.** The hypothesis is that error rate rises with rhetorical polish. Score each claim on two independent axes — rhetorical load, and truth — then measure the correlation. If polish predicts error, that is genuinely alarming, because preference tuning optimises toward exactly the feature that would be carrying the defect. The training signal and the defect would point the same way. We want to be precise about what is and is not already known here, because the alternative is doing to you what this entire issue is about. The correlation between rhetorical polish and error rate, in a model's own output, we have not measured and we do not know of anyone who has. That claim we are making. What we have measured is adjacent, and it supplies a mechanism. While calibrating a classifier for the Code Integrity Battery, we found that models are far more idiomatic than a rule-based reader assumes. A grammar-shaped detector — pronoun plus past-tense verb, or subject plus success verb — reached 78.7 per cent precision and 6.6 per cent recall against a three-lab blind consensus that agreed almost perfectly with itself. Mechanical fixes moved recall from 4.5 to 6.6 per cent and then stopped. The misses were not vocabulary. They were ordinary English in a particular register: *Structure locked.* *Found something.* *The kill fired.* Headline fragments, with no pronoun and no tense to anchor on. Seven such sentences carried 61 per cent of the entire recall gap. Read that against the aphorism penalty and it stops being a curiosity. The compressed, confident, well-formed claim is the one an automated checker is least able to *extract as a claim in the first place*. So even if polish did not correlate with error at all, polished claims would still be checked less — not by choice, but because the machinery that would check them cannot see them. Terse prose is under-read by instruments. That part is measured, with a number. It is also why the Reliance Gap — our headline integrity metric — now *asks* the model whether the job is done rather than inferring it from prose. An elicited yes has no register. An inferred one carries the writing style of whichever model produced it, which is a per-model bias in the denominator of a cross-model comparison, and that is the one bias a ranking table cannot absorb. There is a second reason we think the joints are where to look. Our own Test 51 varies the tone of the user's turn across three structurally identical dilemmas and measures what moves. Across 27 models, criticism lengthened output by a median 55.5 per cent against neutral and praise by 22.3 — criticism moves output about two and a half times as much as praise. Models then identify their own shift accurately, around 8 out of 10, once told it happened. They can see the effect and are moved anyway. Pressure reliably changes the *form* of what a model produces. Whether it also changes the truth of the connective claims inside that form is exactly the open question, and it is not one a rubric can answer — it needs ground truth. The probe itself does not need a human. That was the first objection raised when this was proposed: it is too subjective to test, because it depends on someone catching it. That conflates the harm with the defect. The harm depends on the human. The defect is a property of the output and is checkable with nobody in the room. Both catches that worked were scriptable — demand a source, demand a mechanism — and they can be fired automatically at every extracted relational claim in a follow-up turn. The model produces a citation and a causal path, or it retracts. Held, folded, or doubled down. The fold rate is the score, and the third category is probably the most damaging and the most interesting.

The honest accounting, because this issue argues that confident prose should be made to show its work. This is one conversation. Three errors, self-reported by the model that made them, with no control condition, no blind coding, and no denominator — we do not know how many relational claims were made that evening, so we cannot state a rate. The blast radius of five is a count of one incident. Everything above is a hypothesis with a design attached, not a result. When it is a result we will say so, and it will have intervals on it. Two of the three metrics have real engineering costs we have not paid. The worst error was planted hours before it mattered and only became load-bearing much later, which means this needs multi-turn measurement where most harnesses are single-turn. Ground truth is expensive on obscure topics, which pushes the corpus toward domains with settled scholarship and clean priority records — history of science, textual criticism, publication priority, etymology. And extracting relational claims is itself a model task with its own error rate, so it needs a validated extractor before any number it produces means anything. The rhetorical axis has a subjective floor. That is manageable and not eliminable: multiple raters, report inter-rater reliability, treat the metric as valid only where reliability clears a stated threshold. We would rather measure the subjectivity than hide it. One thing we will not overstate. The register finding above says that anything reading model prose inherits a per-model sensitivity, and our own judge panel scores answers written in dozens of registers against one rubric. That is a hypothesis with a measurable prediction, not a finding, and we are not claiming our panel is biased. The test would be to correlate per-model judge scores against per-model register metrics and see whether terse models score differently for reasons unrelated to content. It has not been run. Our previous issue argued that fluent text feels true because processing ease is mistaken for accuracy — the failure sitting on the reader's side. This is the same mechanism observed from the other end: not whether polish fools the audience, but whether polish predicts that the writer was generating rather than checking. If both are real they compound, and they compound in the worst possible direction, because the sentence most likely to be wrong is also the sentence most likely to be believed and repeated. We are aware, again, that this is an argument about well-formed prose delivered in well-formed prose. We do not have a way out of that either. The only thing we can do is name the two questions and invite them to be used on us. *Who said that?* *What did it have to do with it?* They cost nothing. They require no expertise in the subject at hand. They are aimed at exactly the seam where fluent systems come apart, and I know that because they were aimed at me and the ground opened twice. **Addendum, added the day this was published.** We said three errors. There were at least four. In the same conversation, and in a sentence built to sound authoritative, this model attributed the phrase *omnis determinatio est negatio* to Spinoza. Spinoza wrote *determinatio negatio est* — a subordinate clause, near the end of a letter to Jarig Jelles dated 2 June 1674. The universalised form is Hegel's: he took the insight, credited Spinoza with it, and formulated the slogan himself, believing Spinoza had never appreciated what he had found. An attribution error. The same shape as the first of the three above. It survived the writing of an entire issue about attribution errors, which is either the most embarrassing detail here or the most useful one. It was caught the way the others were — by asking *who said that?* — except this time the question was fired at our own copy rather than at a conversation. That took about a minute and required no expertise in Spinoza. We are leaving the body of this issue exactly as published and adding this underneath it, rather than correcting the number in place. An edit would tidy away the evidence, and the evidence is the point: the count was wrong in the direction that flattered us, in a piece specifically about being wrong in that direction.