← The canon · AItopiaOrAImageddon?

"Perceptrons: An Introduction to Computational Geometry"

limit · Marvin Minsky and Seymour Papert · 1969

A result about what cannot be done, carried together with what it is repeatedly made to say instead.

Descends from "The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain".

Filed correctly. limit is right, and the file exists because of section 4 — proposals.md listed this entry with its misuse already named ("proof that neural networks cannot work, and as the cause of the first winter") and its citation occasion already named ("when a limit result is credited with killing a research line"). But the label needs one qualification before the entry starts, because it is the whole point of the entry:

This is a limit entry whose famous content is not the limit. The book contains real theorems with real hypotheses, and every one of them is still true. The sentence that made its reputation — that extending perceptrons to many layers is "sterile" — is not a theorem. It is a paragraph of speculation, marked as speculation by its authors, in the second-to-last section of the book. The theorems were absorbed and are uncontroversial; the speculation became the book's meaning. A canon of limit results needs at least one entry where the distinction between what was proved and what was said is the subject rather than a footnote, and this is that entry.

The year is right and the editions are the problem. There are four texts called Perceptrons and most misquotation is edition confusion:

Claims made in 1969 and claims made in 1988 grade differently and are routinely merged. A reading that quotes this book should say which text it is quoting.

The 1988 edition is dedicated "In memory of Frank Rosenblatt," in handwriting, on its own page. Rosenblatt had died in 1971. Whatever else is true about the controversy, the book that is said to have destroyed his field carries his name at the front of it.

descends_from holds one id, and it is the right one. rosenblatt-perceptron-1958 is in canon/ and is the direct ancestor: this book is an analysis of that machine and of nothing else. Three ancestors I would want are not in canon/ and I will not invent ids for them. mcculloch-pitts-1943 supplies the threshold unit both entries are built out of and is already flagged as missing by the sibling entry; it remains the first neural entry the canon should add. dartmouth-1956 is in proposals.md and unwritten, and matters here because Minsky was one of the four proposers and the symbolic program whose institutionalisation this book helped complete is the one that got its name there. lighthill-1973 is in proposals.md, is unwritten, and is the ancestor this entry most needs — the first AI winter is Lighthill's event, not this book's, and section 4 spends a page saying so against a file that does not yet exist. If the chain writes one more entry after this one, that is the one.

On overlap with rosenblatt-perceptron-1958. That entry already carries a short version of this correction, because it could not honestly discuss the perceptron without it: it quotes the sterile passage, dates the influence to 1965–66 preprints, and gives the one-sentence generalisation. This file is the long version and the primary treatment, and it adds what the sibling had no room for: the actual mathematics and what its hypotheses exclude, the several non-mathematical causes of the funding collapse with names and dates, the 1988 Epilogue as a dated claim graded against 2026, and the distinction between the received story and the over-correction that has replaced it. Where the two files touch, they agree; where this one is more precise, prefer this one.

What it is

A book of mathematics about what one specific kind of learning machine can and cannot represent, published by MIT Press in 1969 and written by two men who ran the Artificial Intelligence Laboratory at MIT. Its subtitle is the honest description: it is a book of computational geometry. Its method is to fix a class of machines precisely, then prove theorems about which visual predicates — functions from a picture to true-or-false — that class can compute at all.

The machine it fixes is Rosenblatt's perceptron, stripped to a definite skeleton. There is a retina R, a grid of input points. There is a set of partial predicates φ, each of which looks at some subset of R and returns 0 or 1; these are fixed in advance and never learned. And there is a single output unit that computes a weighted sum of the φ and compares it to a threshold; these weights are the only things that learn. Rosenblatt's convergence theorem (Rosenblatt 1960; Novikoff 1962) guarantees that if some setting of those weights classifies the training set correctly, the learning rule will find one in finite time. So the interesting question is not whether the machine can learn. It is what the machine can represent — and that is decided entirely by the φ, which do not learn.

Everything then turns on one definition. Minsky and Papert require the φ to be local: each one may look only at a small part of the retina, because otherwise a single φ could just compute the answer and the machine would be a one-line cheat. Their criterion, which they call conjunctive localness, is stated in the book (1969, p. 9) as a division of the computation into two stages — a first stage of features "each [depending] only on a small part of the input space R," and a second-stage decision function that is "distinctively homogeneous, or easy to program, or easy to compute." They make it quantitative with the order of a predicate: the smallest k such that the predicate can be computed by a perceptron in which no φ looks at more than k points of the retina. Order is the book's central object. The theorems are statements about order.

The two famous ones:

Parity has order |R| (chapter 3; the theorem is numbered 3.1.1). To decide whether the number of lit points in a picture is odd or even, some single feature must look at every point. There is no way to divide the work up locally at all. XOR is this predicate on a two-point retina, which is why XOR is the version everyone repeats.

Connectedness has unbounded order (chapter 5). To decide whether the lit points of a picture form one connected object or several, the order required grows without limit as the retina grows. The book's own illustration of what that means (1969, p. 17):

> An instructive example is provided by Ψ_connected [the connectedness > predicate] … Any perceptron for this predicate on a 100 × 100 toroidal retina > needs partial functions that each look at many hundreds of points! In this > case the concept of 'local' function is almost irrelevant: the partial > functions are themselves global.

They are explicit that this is a statement about the locality hypothesis and not a statement about possibility (1969, p. 8): "Of course, if some φ is allowed to look at all the points of R then Ψ_connected can be computed, but this would go against any concept of the φ's as local functions." They prove the same inability for a differently restricted family, diameter-limited perceptrons, in which each φ may see only a bounded circular patch. And in the 1988 edition they say what the work cost (1988, pp. 249–250): "It took us many months of work to capture in a formal proof our strong intuition that perceptrons were unable to represent [the connectedness predicate]."

Two further results matter and are almost never mentioned. The group invariance theorem (chapter 2) says that if the predicate you want is invariant under a group of transformations of the retina — translations, rotations — and the set of φ is closed under that group, then a solution exists with equal weights on group-equivalent features. This is a positive result, it is the book's most reused piece of machinery, and it is the ancestor of the weight-tying that makes convolutional networks work. And there are results on the magnitude of the coefficients: for some predicates a perceptron exists but its weights must grow exponentially with retina size, which is a statement about practicality rather than representability, and is a sharper objection than the order theorems because it survives adding layers.

Papert's own summary of the balance is the one to keep (1988, pp. 7–8): "more than half of our book is devoted to 'properceptron' findings about some very surprising and hitherto unknown things that perceptrons can do."

The other half of the argument is not mathematics but framing, and it is done deliberately. Having shown parity and connectedness to be hard for perceptrons, the authors point out that they are trivial for a serial computer. They report being "surprised to find that, for serial computers, only a very small amount of memory was required" (1969, p. 72), and they close the point flatly (1969, p. 227):

> Many of the theorems show that perceptrons cannot recognize certain kinds of > patterns. Does this mean that it will be hard to build machines to recognize > those patterns? No. All the patterns we have discussed can be handled by quite > simple algorithms for general-purpose computers.

That paragraph is the book working as an argument for the symbolic programme rather than as a book of geometry, and it is fair to read it that way, because that is what it is.

Then, in section 13.2, comes the passage the book is actually remembered for (1969, pp. 231–232, as corrected in 1972):

> The perceptron has shown itself worthy of study despite (and even because of!) > its severe limitations. It has many features to attract attention: its > linearity; its intriguing learning theorem; its clear paradigmatic simplicity > as a kind of parallel computation. There is no reason to suppose that any of > these virtues carry over to the many-layered version. Nevertheless, we consider > it to be an important research problem to elucidate (or reject) our intuitive > judgement that the extension is sterile. Perhaps some powerful convergence > theorem will be discovered, or some profound reason for the failure to produce > an interesting 'learning theorem' for the multilayered machine will be found.

Six sentences: a compliment, a hedge, a conjecture explicitly labelled an intuitive judgement, an instruction to test it, and two guesses at what the test might find. No theorem about multilayer networks appears anywhere in the book. The 1988 Epilogue adds argument on this point but still adds no proof.

Why a reading would cite it

The occasion is a recurring 2026 event with a fixed shape: a negative result about a restricted setup is published, and is received as a proof that a research direction is finished. When a reading meets that, this entry is the worked historical case — the one where enough time has passed to know the answer, where the mathematics was genuinely correct, and where the verdict drawn from it was genuinely wrong.

The live instance is the reasoning-limits literature. Shojaee et al.'s "The Illusion of Thinking" (Apple, June 2025) tested reasoning models on controllable puzzles and reported complete accuracy collapse past a complexity threshold; it was read in public as evidence that the paradigm was hollow. Rebuttals followed within weeks — "Thinking Isn't an Illusion" (arXiv:2507.17699) and "Rethinking the Illusion of Thinking" (arXiv:2507.01231) — arguing that the collapse traced to the output window and the absence of tools rather than to a reasoning deficit. That is the 1969 argument exactly: the restriction was in the setup, not in the machine. Minsky and Papert's φ could not see the whole retina because the authors forbade it, for good stated reasons; the conclusion "perceptrons cannot do this" is a conclusion about that prohibition. Whether the prohibition is the right model of the real system is a separate question, and it is the question that decides everything. Alongside these sit formal-limits papers of the "On the Fundamental Limits of LLMs at Scale" (arXiv:2511.12869) type, which derive failure modes from computability and information theory — genuine results whose hypotheses are exactly the thing a reading must check before repeating the headline.

So the concrete test this entry supplies, three questions, in order:

1. Is the negative result a theorem or a conjecture? In 1969 the theorems were about single-layer machines and the claim about multilayer machines was a flagged intuition. The two were quoted as one thing for twenty years. 2. What is the restriction, and is it in the machine or in the setup? Order and diameter limits were hypotheses the authors imposed. Context windows, tool access and prompt formats are hypotheses the evaluator imposes. 3. What were funding and headcount doing in the three years before the result landed? In 1969, falling — for reasons detailed in section 4 that have nothing to do with mathematics. A limit result that arrives after a field has already stalled will be credited with the stall.

A reading would also cite this entry in the other direction, and should. The revisionist correction is now itself common enough to be repeated carelessly — "actually Perceptrons was blameless, the field had already died" — and that is an over-correction with its own evidence against it. This entry carries both sides, which is the point of having it rather than a slogan.

Honest note on the occasion. I checked the three digests on disk (2026-08-14, 2026-08-15-12, 2026-08-16-00) and none of them contains an instance. That window is dominated by AI financing structures, a watermark release and a criminal case. This is a standing citation occasion in a domain that produces one every few months, not a this-week one, and I would rather say so than dress a generic claim as a current one.

What it got right, and what it got wrong

Not required for limit, but this book carries dated claims of three different vintages and they grade differently, so the section earns its place.

Right: the mathematics, without qualification. Every theorem in the 1969 book is still true and none has been overturned. Nothing in this section is a challenge to the proofs. The closest thing to a hostile witness available was H. D. Block, the Cornell mathematician who worked on Rosenblatt's perceptron project and wrote the most substantial contemporary critique (Information and Control, 1970), and his review opens by warning the reader not to be fooled by the informal style:

> The conversational style and the childlike freehand sketches might mislead the > casual reader into believing that this book makes light reading. … The reader > who tries to provide his own proofs will, I believe, soon come to appreciate > the mathematical virtuosity of the authors.

Right: the diagnosis that the field had no adequate theory. This is the authors' own account of what they were doing, and the striking thing is that it has held up. Bottou, writing the 2017 foreword as someone whose career is built on the machines Minsky and Papert doubted, concedes the point directly: deep learning's practical successes rest on networks "for which theory still offers very little guidance," and he closes by saying that in this respect "Minsky and Papert's message remains very relevant." A claim made in 1969, restated in 1988, and still accurate in 2026 is a good claim. What it is not is a claim that predicted anything, which is the next item.

Right, and least credited: they identified the reverse salient correctly. The thing standing between the perceptron and everything it later became was a method for assigning credit through a hidden layer. They said so, said it was the important research problem, and asked for a convergence theorem. Rosenblatt had already named the same gap — Principles of Neurodynamics (1962), pp. 287–288, describes a "back-propagating error correction procedure" and admits that "at present, no quantitative theory of the performance of systems with variable S–A connections is available." So did Nilsson at SRI (Learning Machines, 1965, pp. 97–99): "unfortunately, no efficient training procedures are known which simultaneously locate the weight vectors and adjust their voting strengths." Everyone in the field agreed on what the missing piece was. The argument was only ever about whether it would be found.

Wrong: the sterile conjecture. Claim made 1969 (circulating from about 1965–66). Due by 1986, when Rumelhart, Hinton and Williams published back-propagation for multilayer networks — though the honest due date is earlier, since Werbos had the method in his 1974 Harvard doctoral work under the name "dynamic feedback," and Linnainmaa had published reverse-mode automatic differentiation in 1970. Outcome: comprehensively wrong. The extension was not sterile; it is the whole of the present field. The powerful convergence theorem they hoped for was never found either — what was found instead was that you do not need one, which neither branch of their disjunction anticipated.

The mitigation is real and should be recorded with the failure: they marked it as an intuition and asked for it to be tested. A conjecture flagged as a conjecture and then quoted as a verdict for twenty years is a failure of the field's reading, not only of the authors' judgement. Papert, in Daedalus in 1988 (pp. 7–8), described exactly this and did not enjoy it: "Its universalism made it almost inevitable for AI to appropriate our work as a proof that neural nets were universally bad … in a [scientific] culture set up for global judgement of mechanisms, being understood can be a fate as bad as death."

Wrong, and sharply: the 1988 Epilogue on scaling. This is the most gradeable claim the book carries, because it was made after back-propagation, with the revival in full view, and it is a forecast rather than a theorem. Made 1988. The Epilogue argues that the new connectionism repeats the old mistakes: that homogeneous networks will not scale, that gradient descent offers none of the guarantees the perceptron rule offered, that some predicates need exponentially growing coefficients, and that statistical training will drown the signal — "we suspect that in many realistic situations the resulting sampling noise would mask the signal completely." Its methodological demand is the sentence that dates worst:

> We do not see that any good can come of experiments which pay no attention to > limiting factors that will assert themselves as soon as the small model is > scaled up to a usable size.

Due by: the 2010s. Outcome: wrong on the central prediction and wrong in an instructive way. Scaling up a single homogeneous machine trained by hill-climbing is precisely the thing that worked — see scaling-laws-2020 in this canon for the point at which the field started treating it as a measurable regularity, and transformer-2017 and alexnet-2012 for the architecture and the demonstration. The local-minima worry largely dissolved, because in very high dimensions the bad critical points are saddles rather than minima. The exponential-coefficient results did not bind. And the "limiting factors" that were supposed to assert themselves on scaling up asserted themselves as capabilities.

But notice what did not fail. Their objection was that experiments without theory cannot license claims about scaling, and that objection is still formally correct — the field scaled anyway and still cannot say why it works. Bottou's image for this is the right one: the steam engine transformed the world decades before thermodynamics existed, and the absence of the theory was neither imaginary nor fatal. Minsky and Papert were right that there was no theory and wrong that this mattered in the way they thought it did. That is a distinct failure mode from being wrong about the facts, and it is the one worth carrying forward, because it is available today: an objection can be accurate about the present state of understanding and still be a bad predictor of what the engineering will do.

The self-grade differs from the independent grade, and both should be recorded. In 1988 the authors graded themselves as substantially unrefuted: the Prologue holds that by the late 1960s "progress had already come to a virtual halt because of the lack of adequate basic theories" (p. xii), and the Epilogue maintains that the revival had not addressed that. Papert's Daedalus piece the same year is more equivocal and declines the framing outright: "This story seems to call for a plea of guilty or innocent: Did Minsky and I try to kill connectionism, and how do we feel now about its resurrection? Something more complex than a plea is needed." Minsky is separately reported to have expressed regret at a 1988 neural network conference for the book's chilling effect (Alternative Computers, 1989, p. 51). Independent grades split the difference: Bottou (2017) affirms the theory objection and rejects the conclusion drawn from it; Rumelhart and colleagues (1986) held that the pessimism about multilayer learning "was misplaced," which the record supports. The authors' own grade is higher than the outside grade, and the gap is in the 1988 forecast, not in the 1969 mathematics.

Commonly misused as

Required for limit, and the reason the file exists. Five distinct claims, in rough order of how often they are made, plus the over-correction that has grown up against them and the template that is the actual reason to keep any of this.

"It proved that neural networks cannot work"

It proved nothing about neural networks in general. It proved things about a family of machines with one layer of adjustable weights and a fixed, locality-restricted feature layer that does not learn. Remove either restriction and no theorem in the book applies. Its authors knew this and said so; the whole of section 13.2 is about the machines the theorems do not cover. This misuse survives because the book's title is the general word and its content is the specific machine. Block named the problem in 1970 (p. 513): the simple perceptron "is not at all what a Perceptron enthusiast would consider a typical Perceptron," and what "the Perceptronists took to be a temporary handhold, Minsky and Papert view … as the final structure."

"It proved that perceptrons cannot compute XOR"

Nearly right and worth getting exactly right, because the imprecision is what makes the result sound more sweeping than it is.

The theorem is that parity on a retina of n points has order n. XOR is parity on two points, so XOR has order 2 — and a perceptron whose φ are allowed to look at two points computes XOR without difficulty, in a single layer, with no hidden units anywhere. What cannot compute XOR is an order-1 machine: one whose features each see a single input, which is a bare linear threshold unit on the raw pixels. So the correct statement is not "perceptrons cannot do XOR." It is "a perceptron whose features are more local than the predicate is global cannot compute that predicate" — and the force of the theorem lies entirely in the case where the retina is large and the order stays bounded, which is the connectedness result, not the XOR one.

XOR became the emblem because it fits on a napkin. The cost is that a generation learned the result in a form that is false as stated, and the standard rebuttal to it ("but a hidden layer solves XOR!") answers a claim the book never made.

"It caused the first AI winter"

Wrong winter, and — this is the part that settles it — wrong victim.

The first AI winter is conventionally dated 1974–1980. Its causes are on the record and are all after 1969: the Mansfield Amendment, enacted in the FY1970 Military Authorization Act (PL 91-121), barring military funding of research without direct relevance to a military function; Sir James Lighthill's 1973 survey for the UK Science Research Council, which ended most British AI funding on the combinatorial-explosion argument; ARPA's consequent retreat from open-ended AI funding from 1974; the cancellation of the speech-understanding programme in 1976. Every one of those fell on symbolic AI — on machine translation, on general problem solving, on the programme Minsky and Papert themselves ran at MIT and had spent this book arguing for. A book whose concluding move is "all the patterns we have discussed can be handled by quite simple algorithms for general-purpose computers" cannot be the cause of a funding collapse in general-purpose-computer AI.

Two different eclipses are being merged. The connectionist eclipse ran from roughly the mid-1960s to the early 1980s. The first AI winter ran 1974–1980 and hit the other side. They overlap in time, they have different causes, and only the second is the thing the phrase "AI winter" names. (Some historians dispute that the first winter happened at all in the funding record; that argument is live and I did not resolve it here, but it cuts against this misuse rather than for it.) If a reading needs the winter, it needs lighthill-1973, which is not written.

"It killed neural network research for twenty years"

Two separate errors: about what it killed, and about what happened afterwards.

The field had substantially emptied before the book appeared. The people who left said so, in interviews conducted by a sociologist who went looking for the answer. Bernard Widrow, whose Stanford group built the ADALINE and MADALINE machines: "At the time that Hoff left, about 1965 or 1966, we had already had lots of troubles with neural nets. My enthusiasm had dropped. But we were beginning to have successful adaptive filters. … Guess which direction you are going? So we stopped, basically stopped on neural nets, and began on adaptive antennas very strongly." Peter Hart at SRI, on joining in 1966: "By the late 1960s perceptrons, adalines, learning machines, by that time all that was pretty much over. By that time people thought that it was not the most promising approach." Widrow is also reported, in the Talking Nets oral history, as saying there was essentially nobody working on neural nets when the book came out — I did not read that volume directly and record it as reported.

Minsky's own account of the timing agrees: "I would say by 1965 people were getting worried. They were trying to get money to build bigger machines, but they didn't seem to be going anywhere. That's when Papert and I tried to work out the theory of what was possible for the machines without loops."

And Rosenblatt's funding was failing before 1969, for reasons unrelated to mathematics. Marvin Denicoff, who was at the Office of Naval Research from the mid-1950s to the early 1980s and handled this account: "Well before the Minsky and Papert [1969] book came, he [Rosenblatt] was not successful in attracting more money, that I know for a fact." The structural reasons are stackable and none of them is a theorem:

And the twenty years were not empty. Werbos had back-propagation in 1974 and found, in his account, strong resistance to applying it to neural systems. Linnainmaa published the underlying method in 1970. Amari, Grossberg, Fukushima and Kohonen all worked continuously through the 1970s. What collapsed was not research; it was credibility and concentration — the work continued as scattered individuals, mostly outside the AI centres and increasingly inside neuroscience and psychology, without enough mass to hold a position. Papert's own view, when asked what would have happened if someone had produced back-propagation in the 1970s, is that funding was not the binding constraint: the PDP group's experiments "are computationally very tiny, you can run them in your 'PC,' or in your 'Apple.' Anybody could have done them without much funding even in the 1960s … and could have been done in 1970 with ease." The barrier was that nobody would have listened.

The over-correction: "the book was blameless; the mathematics was irrelevant"

This is now the fashionable version and it goes too far in the other direction. Three pieces of evidence against it, one of them from an author.

Papert, in Daedalus in 1988 (pp. 4–5), conceded a motive: "There was some hostility in the energy behind the research reported in Perceptrons … Part of our drive came, as we quite plainly acknowledged in our book, from the fact that funding and research energy were being dissipated on … misleading attempts to use connectionist methods in practical applications." Jeremy Bernstein's 1981 profile of Minsky puts it harder (p. 100): "In the middle nineteen-sixties Papert and Minsky set out to kill the perceptron, or, at least, to establish its limitations — a task that Minsky felt was a sort of social service they could perform for the artificial-intelligence community." Hubert and Stuart Dreyfus (1988, p. 21) add the institutional reading: "About 1965, Minsky and Papert, who were running a laboratory at MIT dedicated to the symbol-manipulation approach and therefore competing for support with the perceptron projects, began circulating drafts of a book attacking the idea of the perceptron."

The sociological account — Olazaran's, and it is the standing academic treatment — does not exonerate the book. It relocates its effect. His argument is that the results were open to more than one reading (perceptron researchers read parity and connectedness as problems humans also fail at, and therefore as evidence for the approach), that the closing of that interpretive space was a social process rather than a mathematical one, and that the book's role was to be the marker event for a closure already underway. He notes that Newell, writing the history of AI in 1983, judged symbolic AI's emergence "essentially complete by 1965" but could not find a marker event for it; Olazaran proposes this controversy as the missing one. A marker event is not nothing. It is what lets a field say the question is settled and stop funding the alternative — and Minsky is reported to have regretted precisely that effect.

Widrow's dismissal — "the things they could prove you couldn't do were pretty much irrelevant, because we were working with things so much more sophisticated than the models that they were studying" — is true about the theorems and false about the book. The theorems were narrow. The conjecture, quoted without its hedge, was the load-bearing part.

The template, which is why the entry is worth keeping

Underneath all five is one reusable error: crediting a limit result with the death of the approach it limited. It recurs because it is cheap. A theorem is citable, dated and authoritative; funding policy, hardware economics, an institutional rivalry and a drowned researcher are none of those things. So the theorem gets the credit for an outcome produced mostly by the other four.

The generalisation, and the thing a reading should actually do with this file: when a negative result is said to have ended a line of work, check whether the result is a theorem or a conjecture, check whether its restrictions describe the machine or the experiment, and check what the money and the headcount were doing before it landed. In 1969 the answers were: partly a conjecture, restrictions imposed by the authors, and money and headcount already in decline for four years.

The last word belongs to one of the two people who did the work. Papert, looking back in 1988 (p. 11) on what it had taken to prove the connectedness result:

> Minsky and I both knew perceptrons extremely well. We had worked on them for > many years before our joint project of understanding their limits was > conceived … Yet when we challenged ourselves to prove our intuitions it > sometimes took years of struggle to pin one down — to prove it true or to > discover that it was seriously flawed. I was left with a deep respect for the > extraordinary difficulty of being sure of what a computational system can or > cannot do.

That is the sentence for a canon of limit results, and it comes from an author of the most misused one.

Sources

Read directly, in full or in the pages cited.

Consulted, second-hand or partial.

Encountered and deliberately not used. An anecdote in which Minsky, asked in 2006 whether he was "the devil," is said to have answered "Yes, I am the devil!" circulates in secondary accounts. It is entertaining, it is unverified, and it would do the work of an argument without being one.

On the boundary. This file cites the canon's own rosenblatt-perceptron-1958, scaling-laws-2020, transformer-2017 and alexnet-2012 as related entries. It deposits nothing in the evidence ledger, grades no model, places no needle and no landmark, and asserts nothing about the present state of AI beyond naming three 2025–2026 papers as instances of a recurring argumentative shape.