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

idea is right and the id was filed correctly, but two things about the label need saying before the entry starts, because both of them change how the entry should be read.

The date in the id is the publication, not the work. The Psychological Review paper is November 1958. The document that invents the perceptron is Cornell Aeronautical Laboratory Report No. 85-460-1, The Perceptron: A Perceiving and Recognizing Automaton (Project PARA), dated January 1957, prepared by Frank Rosenblatt as Project Engineer and approved by two department heads at a defence contractor in Buffalo. The funding proposal that turns it into a machine is dated 3 April 1957. Every quotation in this entry that carries a claim about capability, cost or schedule comes from those two 1957 documents, which I read directly, and should be dated 1957 whenever a reading uses one. The 1958 paper is the version the field cites; the 1957 report is the version with the promises in it.

The tempting alternative kind is moment, and it belongs to a different entry. The perceptron's famous date is 7 July 1958, the Office of Naval Research press conference, and its famous text is the next day's New York Times piece. That is proposed separately in proposals.md as rosenblatt-nyt-1958, filed under prediction, and it should stay separate. Splitting them is not tidiness. The whole use of this entry is that the technical claim and the press claim were different claims, made in different documents, and graded differently — and that the man who made the first one spent the rest of his life objecting to the second. Fold them together and the lesson is lost. Precedent is settled: eliza-1966 and mycin-1976 are programs filed as idea; deep-blue-1997 and chatgpt-2022, which are events with dates, are moment.

descends_from is empty, and that is a gap rather than a fact. The real ancestors are three, and none of them is in canon/. Warren McCulloch and Walter Pitts, A Logical Calculus of the Ideas Immanent in Nervous Activity (1943), supplies the threshold unit the perceptron is built out of. Donald Hebb, The Organization of Behavior (1949), supplies the rule that connections strengthen with use. Turing's Intelligent Machinery (1948) describes unorganised machines trained by reward and punishment, and is the nearest thing to a direct precursor written by anyone in canon/ — but the entry that exists is turing-halting-1936, which has nothing to do with this. None of the three appears in proposals.md either. mcculloch-pitts-1943 is the one I would write first; every neural entry the canon ever adds will want to descend from it. Until it exists, this header stays [] rather than pointing at something convenient.

What it is

A machine that learns to sort things into categories by being shown examples, rather than by being told the rule. It is the first system built on that principle that was specified precisely enough to analyse, cheap enough to build, and general enough that the thing you were sorting did not have to be decided in advance. Everything that has since been called a neural network is a descendant of it, and the phrase "multi-layer perceptron" is still the standard name for the feed-forward block inside a transformer.

The 1957 report states the problem it is solving in a form that has not dated:

> interest has centered on the idea of a machine which would be capable of > conceptualizing inputs impinging directly from the physical environment of > light, sound, temperature, etc. — the "phenomenal world" with which we are all > familiar — rather than requiring the intervention of a human agent to digest > and code the necessary information.

and then names the constraint that makes it hard, which is the constraint the whole field still lives under:

> In general, identities of this sort must be learned, or acquired from > experience, and if the system is to be economical, the number of functional > units in the storage system, or memory, should be much less than the number of > forms or memories to be retained. It is this last requirement which seems to > be incompatible with the nature of conventional computer systems.

That is the compression requirement, written down in January 1957, as a design constraint rather than as a discovery. A system that stored one entry per thing it could recognise would be a lookup table; the interesting machine is the one whose memory is smaller than its experience.

The architecture. Three layers, and the names have survived. The S-System (sensory) is a set of photocells or raster points. The A-System (association) does the work: each A-unit receives excitatory and inhibitory connections from a random subset of S-points, fires when the algebraic sum of its inputs exceeds a threshold θ, and carries a "value" that "will vary with its history, and acts as a counter, or register for the memory-function of the system." The R-System (response) fires when the mean value arriving from the A-System exceeds a critical level, and inhibits its rivals, so that a set of mutually exclusive responses "acts like a multi-stable flip-flop, in which only one stage can be 'on' at any one time." S-to-A connections are random and fixed; A-to-R values are what learning changes. Training is: show the machine an example, force the correct response, and increment the value of every A-unit that was active. That is it. The 1957 report works the whole thing through by hand on a 64-unit toy system in Appendix II and gets all six associations right.

The theory of statistical separability. The name Rosenblatt gave to why this works, and the reason the 1958 title says probabilistic. Two different patterns excite overlapping but not identical populations of A-units; with enough units, the overlap is small enough, reliably enough, that the mean value arriving at the right R-unit exceeds the mean arriving at the wrong one. Nothing in the system is guaranteed. The guarantee is replaced by a distribution. Rosenblatt is explicit that this is a purchase and states the price in the same breath:

> The penalty that we pay for the use of statistical principles in the design of > the system is a probability that we may get a wrong response in any > particular case — i.e., a wrong response that is inherent in the nature of > the system, rather than due to a malfunction of one of its components.

Read that sentence twice. In January 1957, in the founding document of machine learning, the inherent-error property is named, distinguished from a hardware fault, and accepted as the cost of the method. Rosenblatt then adds the optimistic half — "It appears that this probability may be reduced, however, to a quantity no greater than the typical probabilities of an error due to malfunctions in electronic equipment" — and that half is graded below.

The theorem that made it respectable. If the categories can be separated at all by a weighted sum, the training procedure will find a separating set of weights in a finite number of mistakes. Rosenblatt claimed a version of this; the proofs that survived are H. D. Block's, in Reviews of Modern Physics 34 (1962), and Albert Novikoff's, at the Symposium on Mathematical Theory of Automata in April 1962, which gives the bound of at most (R/γ)² mistakes and is the one still taught. It is the first convergence guarantee anyone had for a learning machine, and the reason the perceptron was taken seriously by people who did not care about brains.

The hardware. The Mark I Perceptron: a 20×20 array of 400 photocells, 512 association units, 8 response units, with the adjustable values held as potentiometers turned by electric motors. It was built at the Cornell Aeronautical Laboratory around 1959–60 and demonstrated publicly on 23 June 1960; it is now at the Smithsonian's National Museum of American History, having arrived in 1967 from Cornell as a transfer from the Office of Naval Research. It is worth being exact about one thing the retellings blur: the machine demonstrated to the press in July 1958 was not the Mark I. It was an IBM 704 running a simulation, and what it learned, after about fifty punched cards, was to tell a mark on the left from a mark on the right.

A note on what "the perceptron" means. Rosenblatt used it as a class name, not a product name, and said so with visible irritation in 1962: "The term 'perceptron', originally intended as a generic name for a variety of theoretical nerve nets, has an unfortunate tendency to suggest a specific piece of hardware." The 1957 report already distinguishes photoperceptrons from phonoperceptrons, momentary-stimulus from temporal-pattern, and floats "electro- or radioperceptrons." He meant a family of machines, and he built several.

Why a reading would cite it

proposals.md gives the citation occasion as "what training on data can and cannot produce is the question," which is right and is nearly always the question. Four concrete cases, all of which the current readings have already produced.

1. When the argument is about training data — what was bought, what was in it, and what the model kept. This is the live one. The 15 Aug midday reading has California's AB 412, which would have required documentation of copyrighted training material, held in committee with no recorded vote and dead for the session; the 16 Aug midnight reading has defunct startups' Slack archives and Jira tickets selling at $10,000–$100,000 per company; the 15 Aug reading has UK booksellers reporting months of incoherent bulk purchases delivered near Heathrow. Every one of those turns on a question the 1957 report poses as its central design constraint: the storage must be smaller than the experience that produced it. That is what distinguishes a model from a copy, and it is also, awkwardly, what makes it hard to say what a model contains. Rosenblatt did not resolve it. He wrote it down as the requirement that made conventional computers unsuitable, seventy years before it became a discovery question in front of a legislature. A reading covering the training-data fights can say that the compression-versus-retention property is not an emergent surprise of scale; it is the specification.

2. When a wrong output is being argued about as a bug. The inherent-error passage quoted above is the best single citation in this entry, because it is early, it is primary, and it is unambiguous. The founding document of the field says that a system built on statistical separability will sometimes be confidently wrong for reasons that are not faults, and treats this as the price of the capability rather than as a defect to be patched. Cite it whenever a reading has to referee between "the model malfunctioned" and "the model did what it does." It licenses neither complacency nor scandal: Rosenblatt also predicted the error rate could be driven down to hardware-failure levels, and that has not happened in sixty-nine years, which is itself the citation's second half.

3. When a lab says its instruments for measuring its own system have gone bad. The 15 Aug reading's strongest document is Anthropic raising its catastrophic-misalignment rating on the stated ground that safety benchmarks are saturating and R&D-capability measurement is degrading — an uncertainty adjustment, not a finding. The perceptron is where the measurement problem starts, because it is the first system whose behaviour is a distribution rather than a specification. You cannot read a learned weight vector and say what it will do; you can only test it and report a rate. Every argument since about benchmark saturation, held-out sets, contamination and evaluation drift is downstream of the moment the field accepted a probabilistic guarantee in place of a deterministic one, and that moment is dated January 1957 and has a price tag attached in the author's own words.

4. When the question is whether a limit result kills a research line. This is the entry's second-order use, and it is the one that will come up most often in policy and funding coverage. The perceptron is the standing case study in how a technical impossibility proof gets converted into a field-wide verdict, how long the verdict holds, and what it costs. The details are in "Commonly misused as" below, because the popular version is wrong in almost every particular. When a reading meets a claim that some result has settled whether an approach can work — and lighthill-1973, proposed and unwritten, is the case where a funder believed one — this is the entry that supplies the base rate.

There is a fifth occasion worth naming because it is small and exact. The 1957 report, at page 17, proposes adding "an independent R-set with feedback to a set of camera-aiming servos" so that the system can "'fixate' any pattern which might prove significant, in much the same manner as the human eye, limiting its 'recognition learning' to a relatively limited central field, analogous to the fovea in human vision." That is learned attention, proposed as a compute-economy measure, in January 1957. A reading that needs to show that an architectural idea now treated as recent has a dated ancestor has one here, and should resist the temptation to call it a prediction of the transformer. It is not. It is the same problem producing the same shape of answer twice, which is a more interesting claim and a defensible one.

What this entry should not be cited for: any claim that the perceptron shows machines do or do not think. Rosenblatt refused the question in print — "A perceptron is first and foremost a brain model, not an invention for pattern recognition" — and the work is weaker, not stronger, pressed into that service.

What it got right, and what it got wrong

Not required for idea. Included because this entry contains an unusually clean set of dated, numeric, gradeable claims — some in a technical report, some in a funding proposal, some in a newspaper — and because grading them separately is the only way to see that the technical document and the promotional document have very different scores.

Claims dated January and April 1957 unless noted. Graded 2026-08-16.

Right, and it is still the architecture. Random or fixed features, a learned linear read-out, an error-driven update, and a guarantee that holds when the problem is separable and not otherwise. That is the perceptron and it is the last layer of nearly everything shipped in 2026. The name survived intact: the feed-forward blocks inside a transformer are called multi-layer perceptrons, and nobody has felt the need to rename them. The 2016 Deep Learning textbook's formulation is the fair one — the perceptron "became the first model that could learn the weights defining the categories given examples of inputs from each category."

Right, and the least-quoted sentence in the file: inherent error. Graded above. The property is real, it is still with us, and it was named at the start.

Wrong, on the same page: that inherent error could be driven down to hardware-failure rates. Made January 1957, no due date given, and therefore gradeable only as an open claim that has now run sixty-nine years. Electronic components fail at rates that are engineered to be negligible; frontier models in 2026 do not. The 16 Aug reading's watermark item is a small, exact instance of the gap — a lab shipping a detection mechanism and publishing, in the same post, the four conditions under which it does not work. Rosenblatt's optimism here is the founding instance of a specific error: assuming that a statistical system's error floor is an engineering parameter rather than a property of the method.

Wrong by a very large factor, and this is the entry's best scaling claim. Report 85-460-1, page 17:

> A crude estimate indicates that a system with two to three thousand > A-units should be capable of maintaining a vocabulary comparable to the > English language, if connection parameters were properly optimized.

Made January 1957, no due date. What actually happened: no system with two to three thousand adjustable units has ever maintained anything resembling an English vocabulary. The systems that do are on the order of 10¹¹ parameters. The units are not strictly commensurable — an A-unit is not a parameter, and Rosenblatt is counting associative capacity rather than parameters — so the honest statement of the miss is that it is somewhere between six and eight orders of magnitude, and that the direction of the error is what matters. He was not wrong about whether the approach reaches language. He was wrong about the price, by an amount that made the difference between a machine you could build in 1957 and a machine nobody could build until roughly 2018. Cite this whenever a reading meets a capability forecast whose stated bottleneck is ingenuity rather than scale. The canon's other half of this lesson is bitter-lesson-2019, proposed and unwritten.

The proposal's four ultimate capabilities, graded. The 3 April 1957 preliminary proposal, Section I, Summary:

> Devices of this sort are expected ultimately to be capable of concept > formation, language translation, collation of military intelligence, and the > solution of problems through inductive logic.

Made April 1957, in a funding document, with "ultimately" doing the work of a due date — which is the tell, and the reason this list scores far better than anything in the next section. Language translation: achieved, by neural networks with a learned read-out, at production quality from about 2016, fifty- nine years later. Collation of military intelligence: achieved and deployed, and now a procurement category. Solution of problems through inductive logic: partially, and contested; the 15 Aug reading's Crouzeix's-conjecture items are exactly the class of claim this phrase would cover, and neither has passed peer review. Concept formation: unsettled, and arguably unsettleable as worded. Three of four, in about sixty years, with no date attached. A forecast with no due date cannot be scored as wrong, which is precisely why the canon should treat undated forecasts as worth less than dated ones, and why this list is recorded here as an example of the form rather than as a hit.

The schedule, which is the one claim with a real deadline. Same proposal: "The development and construction of a pilot model is expected to require the work of three professional people, a digital computer, and an associated technical staff for eighteen months." Made April 1957. Due about October 1958. The Mark I was demonstrated publicly on 23 June 1960 — roughly twenty months late on an eighteen-month estimate, which is to say the estimate was short by more than its own length. As overruns on frontier hardware go this is mild, and saying so is part of the grade.

The press claims of 7–8 July 1958, which are not this entry's and are graded here only to keep them from being attributed to the wrong document. The New York Times of 8 July 1958, under "NEW NAVY DEVICE LEARNS BY DOING," reports that the Navy expects a device "that will be able to walk, talk, see, write, reproduce itself and be conscious of its existence," that later perceptrons "will be able to recognize people and call out their names" and translate speech between languages, and that the machine was expected to be finished "in about a year at a cost of $100,000." Rosenblatt is separately quoted calling it "the first machine which is capable of having an original idea." Grading, briefly: see, yes, from 2012; talk and write, yes, from about 2020; walk, yes, in robotics, by systems sharing no lineage with this one; recognise people and call out their names, yes, from the mid-2010s; translate speech between languages, yes, from about 2014–17; reproduce itself, no; conscious of its existence, no, and not established as a question with an answer. Roughly four of seven, fifty to sixty years late, by machines related to the perceptron by descent rather than by design. The cost-and-schedule claim is graded above. The full entry belongs to rosenblatt-nyt-1958, and whoever writes it should note that the extravagant list is attributed in the article to the Navy, while the "original idea" line is Rosenblatt's own — the division of responsibility is the whole subject of that entry and this one should not pre-empt it.

And the thing Rosenblatt got right that the story about him gets backwards. He is remembered as the man who oversold a linear classifier and was corrected by people with better mathematics. The record does not support it. In Principles of Neurodynamics — completed as Cornell Aeronautical Laboratory Report VG-1196-G-8 on 16 March 1961 and published as a book in 1962 — he devotes Chapter 14 to "Summary of Three-Layer Series-Coupled Systems: Capabilities and Deficiencies", and the deficiencies he lists include excessive learning time, excessive dependence on external supervision, the figure-ground and connectedness problem, and the failure of two stimuli to be classified as similar unless they occupied nearly the same area of the retina. Then Part III goes past them: Chapter 13.3 is "Back-Propagating Error Correction Procedures", Chapter 15 covers multi-layer perceptrons with fixed preterminal networks, Chapter 16 covers four-layer perceptrons with adaptive preterminal networks, and Chapters 17–20 cover cross-coupled systems with closed loops. Between 1961 and 1967 he built the Tobermory, a speech perceptron with four layers and about 12,000 adjustable weights in tape-wound magnetic cores. He knew what the single-layer system could not do, published the list himself, and spent his remaining years attacking it.

What he did not have was a way to train the middle. His four-layer systems learned the second layer by a Hebbian rule and the third by the perceptron rule, with no derivative anywhere; the adjustable weights in Tobermory are all in the last layer. That gap is the whole of the delay. The field spent roughly two decades failing to find a method for adapting more than one layer at a time — Widrow and Hoff, having trained a single unit successfully by gradient descent, tried nearly everything except gradient descent on two layers and gave up — and the gap closed in 1986 with Rumelhart, Hinton and Williams. So the grade on the central bet is the same shape as logic-theorist-1956's and it is becoming a pattern rather than a coincidence: right about the destination, wrong about the road, vindicated by a mechanism the author could not supply. Rosenblatt did not live to see it. He drowned in a boating accident on Chesapeake Bay on his forty-third birthday, 11 July 1971, fifteen years before the paper that made his architecture work.

Commonly misused as

Not required for idea — the section belongs to limit — but this entry attracts more confident misinformation than anything else in the canon's technical range, and four of the five misuses below are load-bearing in arguments the readings will have to adjudicate.

"Minsky and Papert proved neural networks can't work, and their 1969 book killed the field for twenty years." Almost every clause of this is wrong, and the corrections are worth having exactly because the sentence is used as a template for present-day claims about limit results.

What they proved. Precise, correct, and narrower than advertised. Working with the order of a predicate — the largest number of retina points any single A-unit is allowed to see — they showed that parity requires order equal to the whole retina (Theorem 3.1.1), and that the order required to compute connectedness grows with input size (Theorem 5.5). XOR is the two-point case of parity and is the version everyone repeats. These are results about what a single layer of bounded-fan-in features can represent. They are true.

What they did not claim. They knew perfectly well that multi-layer networks compute XOR. The sentence that gets quoted as their dismissal is, in full, from Chapter 13 of the 1969 edition (pp. 231–232 in the text as corrected for the 1972 second printing):

> 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 judgment that the extension is > sterile.

An intuitive judgment, flagged as such, with a call to test it. It is routinely quoted as a verdict, routinely misdated to the 1988 epilogue, and routinely truncated before the word "Nevertheless."

Whether it killed the field. The people who were there say no. Bernard Widrow: "As far as I knew, there wasn't anybody working on neural nets when that book came out. I couldn't understand what the point of it was." Michael Arbib: the funding priorities and departmental fashions "had shifted the emphasis away from neural nets to the more symbolic methods of AI by the time the book came out." Minsky and Papert's own 1988 prologue says "progress had already come to a virtual halt because of the lack of adequate basic theories," and Papert, in Daedalus in 1988, declined the plea in both directions: "Did Minsky and I try to kill connectionism, and how do we feel about its resurrection? Something more complex than a plea is needed." Mikel Olazaran's 1996 study in Social Studies of Science is the standing academic treatment and argues that the "official history" is a product of symbolic AI's institutionalisation, later rewritten again by the winners of the 1980s revival. The influence that was real ran through preprints and conference talks from about 1965–66, not through the 1969 book — by which time, on the historians' account, there was little left to kill.

The correction is one sentence and it generalises: a limit result that lands after a field has already stalled gets credited with the stall. A reading that meets "result X ended approach Y" should ask what the funding and the headcount were doing in the three years before X, and this entry is the worked example.

"Rosenblatt didn't know about the XOR problem" / "he thought one layer was enough." Graded above and false. He published the deficiency list himself in 1961, including the connectedness problem by name, and Part III of the same book is his attempt on multi-layer and cross-coupled systems. What he lacked was credit assignment through a hidden layer, which nobody had. Say that instead; it is both true and more useful, because the shape of the failure — a working idea stalled for two decades on a missing training method rather than on a missing insight — is a shape worth recognising when it recurs.

"There was a neural network trained to spot tanks that learned to spot sunny days instead." Told constantly, often set "in the early days of the perceptron," and never once traced to a documented experiment. Gwern's investigation finds no primary source: every citation leads to another secondary. The earliest thread runs to Edward Fredkin's recollection of a conference around 1963 at which he speculated that a tank-detecting perceptron might be responding to overall brightness — a speculation, at a talk, never tested — and the first detailed written version is Hubert Dreyfus's, in 1992, thirty years later. Gwern's conclusion: "it is definitely not real as usually told: it is just an urban legend." The underlying phenomenon is real and has a proper name — shortcut learning — and there is no shortage of documented cases. A reading should use one of those. Using this one costs credibility for nothing, and the readings' own discipline about vendor-reported versus independently verified numbers applies with more force to a sixty-year-old anecdote than to a press release.

"Rosenblatt was the original AI hype man." The 1958 press coverage is genuinely extravagant, and one of its most extravagant lines is his. But the list about walking, talking and self-reproduction is attributed in the article to the Navy, and Rosenblatt's own retrospective judgment on that coverage, in print, in the preface to his 1962 book, is that the popular press "fell to the task with all of the exuberance and sense of discretion of a pack of happy bloodhounds," and that headlines such as "Frankenstein Monster Designed by Navy: Robot That Thinks" — Tulsa, Oklahoma Times — "were hardly designed to inspire scientific confidence." The same preface insists that "the perceptron program is not primarily concerned with the invention of devices for 'artificial intelligence', but rather with investigating the physical structures and neurodynamic principles which underlie 'natural intelligence'." That is a researcher trying to get his subject back. Whether he tried hard enough, and whether he could have, is a fair question and a good one for the rosenblatt-nyt-1958 entry. "He was the hype man" is not a fair summary of the documents, and a canon whose job is grading overclaims should be exact about who claimed what.

"The perceptron is the ancestor of every neural network since." True as a lineage claim, and it is how proposals.md words it, but it hides which part is inherited. What descends unbroken is the architecture — weighted sums, a threshold or other nonlinearity, a learned read-out layer, error-driven updates — and the vocabulary. What does not descend from Rosenblatt is the training method that made depth work. Backpropagation is not in his books; the systems that use it descend from Rumelhart, Hinton and Williams in 1986, standing on work in several other fields. A reading that wants to say "this is Rosenblatt's idea, at scale" should say "this is Rosenblatt's architecture, trained by a method he did not have," which is both accurate and the more interesting sentence.

Sources

Primary, read directly:

• Frank Rosenblatt, The Perceptron: A Perceiving and Recognizing Automaton (Project PARA), Cornell Aeronautical Laboratory, Inc., Buffalo, New York, Report No. 85-460-1, January 1957. Read as a page scan from the University of Massachusetts "Brain Wars" course archive at websites.umass.edu/brain-wars/files/2016/03/rosenblatt-1957.pdf (served via a bpb-us-e2.wpmucdn.com redirect). Every quotation in this entry of the "phenomenal world" passage, the storage-smaller-than-experience requirement, the definition of a perceptron as probabilistic, the S/A/R-system descriptions, the inherent-error passage and its optimistic sequel, the two-to-three-thousand-A-units vocabulary estimate, the 1000-A-unit worked example and the camera-aiming/fovea proposal is from this text, at the pages cited. The scan carries a TIB/UB Hannover stamp, consistent with the report's reputation for near-total scarcity.
• Frank Rosenblatt, Proposal for the Development of a Perceiving and Recognizing Automaton (Project PARA), Cornell Aeronautical Laboratory, "PRELIMINARY COPY", 3 April 1957, bound at the end of the same scan and stamped received by CAL's Systems Research Department on 4 April 1957. Section I, Summary, is the source for the four ultimate capabilities and the eighteen-month estimate. Only the proposal's first two pages are present in the scan; its cost figures and schedule detail were not readable and are not quoted here.
• Frank Rosenblatt, Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms, Cornell Aeronautical Laboratory Report No. VG-1196-G-8, 16 March 1961, under two Office of Naval Research contracts, one at CAL for Project PARA and one at Cornell University; published as a book by Spartan Books, Washington DC, 1962. The contract numbers on the title page are Nonr- prefixed but the digits are ambiguous between the report's OCR and DTIC's catalogue record, so they are not stated here. Read as the DTIC copy, AD0256582, from the Internet Archive full text at archive.org/stream/DTIC_AD0256582/DTIC_AD0256582_djvu.txt. Source for the preface — the "brain model, not an invention for pattern recognition" line, the "happy bloodhounds" remark, the Tulsa headline, the natural-intelligence statement — and for the chapter structure of Part III and the title of Chapter 14. Caveat: the fetch returns only the opening of a 616-page text; the preface and table of contents were readable, the body of Chapter 14 was not. The specific deficiencies attributed to Chapter 14 in this entry reach it through Yuxi Liu's quotation of it (below), not from my own reading, and the OCR of the Tulsa headline is garbled in the raw text — the wording used here is the one corroborated by Wikiquote's transcription of the same preface.
• Marvin L. Minsky and Seymour A. Papert, Perceptrons: An Introduction to Computational Geometry, reissue of the 1988 expanded edition with a new foreword by Léon Bottou, MIT Press, 2017 (ISBN 978-0-262-53477-2). Front matter and Bottou's foreword read directly as page images from leon.bottou.org/publications/pdf/perceptrons-2017.pdf. Source for the copyright page's account of the 1972 handwritten alterations and the 1988 preface and epilogue; for Bottou's statement of what the parity and connectedness results establish; for H. D. Block's 1970 review; for Papert's 1988 Daedalus passage and the 1988 prologue's "virtual halt" line; and for the citations to Block (1962) and Novikoff (1962) on convergence. The copy carries a handwritten dedication, "In memory of Frank Rosenblatt." Rosenblatt died in July 1971, two years after the first edition, so that dedication cannot be from the 1969 printing; I did not check a 1969 copy to establish when it was added.

Secondary and corroborating:

• Quote Investigator, "Quote Origin: (Multilayer Perceptrons) Our Intuitive Judgment That the Extension Is Sterile", 24 July 2026, for the full "sterile" passage, its location at Chapter 13, pp. 231–232 of the 1969/1972 text, and its history of truncation and misdating.
• Yuxi Liu, "The Perceptron Controversy", 11 January 2024, at yuxi.ml (formerly yuxi-liu-wired.github.io), for the Widrow and Arbib recollections, the Cowan account of the 1965–66 preprints, the Tobermory configuration, the quotation of Rosenblatt's own Chapter 14 deficiency list, and the argument that connectionism had already emptied out before 1969. This is a blog essay, not a peer-reviewed paper; it is heavily sourced and the quotations it carries are checkable, but everything credited to it here is at one remove.
• Mikel Olazaran, "A Sociological Study of the Official History of the Perceptrons Controversy", Social Studies of Science 26 (1996), pp. 611–659. Abstract read; the article itself was not, and its argument is summarised here from that abstract and from secondary description.
• Gwern Branwen, "The Neural Net Tank Urban Legend", gwern.net/tank, for the Fredkin c.1963 conference recollection, the 1992 Dreyfus version, the absence of any primary source, and the conclusion quoted.
• Ian Goodfellow, Yoshua Bengio and Aaron Courville, Deep Learning, MIT Press, 2016, for the formulation quoted in "What it got right". Second hand, via the University of Massachusetts "Brain Wars" course page that hosts the 1957 report scan; I did not read the textbook.
• Cornell Chronicle, "Professor's perceptron paved the way for AI — 60 years too soon", September 2019, for Rosenblatt's biography, the July 1958 IBM 704 demonstration and the fifty punched cards, the Cognitive Systems Research Program, the Tobermory's name, the Smithsonian, the IEEE Frank Rosenblatt Award (2004), and the date and circumstances of his death.
• English Wikipedia, "Perceptron", "Mark I Perceptron" and "Perceptrons (book)", consulted 2026-08-16, for the Mark I's configuration and 23 June 1960 demonstration, the Smithsonian transfer, the reported per-task accuracies, and the statement of Theorems 3.1.1 and 5.5. Tertiary; the hardware figures are consistent across the Smithsonian's own catalogue description and Cornell's account, and the theorem numbers are corroborated by Bottou's foreword.

Noted failures, per the discipline:

• I did not read the 1958 paper itself. Psychological Review 65(6), November 1958, pp. 386–408, doi:10.1037/h0042519, is behind APA's paywall, and eight open routes failed on 2026-08-16: a KU Leuven repository PDF returned an empty DSpace shell; the University of Pennsylvania and MIT Press copies returned HTTP 403; two university course-page URLs returned 404; a Semantic Scholar "PDF" resolved to an unrelated 2015 student slide deck; the Internet Archive's microfilm scan of the issue is present as sim_psychological-review_1958-11_65_6 but its full text truncates before page 386. Consequently the three opening questions, the sentence placing the theory in the "connectionist" tradition, and the closing line — "By the study of systems such as the perceptron, it is hoped that those fundamental laws of organization which are common to all information handling systems, machines and men included, may eventually be understood" — reach this entry through multiple independent secondary attestations rather than through the page. The closing line is quoted identically by Manhattan Rare Books' catalogue description of an offprint and by James Dobson's paper on memorization in early machine learning, which is why I have used it; a session with journal access should verify it and the three questions against the printed text. This is the single item that would most improve the entry.
• Rosenblatt's The Perceptron: A Theory of Statistical Separability in Cognitive Systems, CAL Report VG-1196-G-1, 1958 — the 262-page technical twin of the Psychological Review paper — was located in the Stanford and Wisconsin library catalogues but no digital copy was found.
• Rosenblatt's "Two Theorems of Statistical Separability in the Perceptron", in the proceedings of the NPL symposium on the Mechanisation of Thought Processes (24–27 November 1958), pp. 419–472, was not read. The Internet Archive holds the proceedings as mechanisationoft01nati and mechanisationoft02nati, but the full text did not come back through the fetch. It carries a recorded discussion, which would be the best available primary evidence of what Rosenblatt's contemporaries said to his face in 1958, and it is the second item that would most improve this entry.
• DTIC's own server returned HTTP 403 for AD0256582; the Smithsonian's object page for the Mark I returned HTTP 403; and pdftotext is outside this job's tool allowance, so scanned PDFs were read as page images rather than as text. Where a scan's OCR and my reading of the page image disagreed, the page image won.

This entry places no needle, no score and no landmark, and deposits nothing in the evidence ledger.