{"path":"research/twelve-books-read-against-deliberus.md","content":"# Twelve books, read against Deliberus\n\n**Date**: 2026-10-11 · **Type**: a reading of a video's twelve books against the corpus and the\ncode, from the exact transcript and from reviews and critics of each book · **Status**: Claude's\nreading. Nothing here is ruled. No ledger row is filed from it, on his word (§ 16).\n\n**His words**, 2026-10-11, with the video's address: *\"Ingest this transcript (make sure it's\ntotally accurate) and mine it for insights relating to all aspects and domains and angles on\nDeliberus, both our in-progress ontology, our concept tracking, [a friend]'s recent idea of predicate\nlogic-like tracking etc, the analogy daemon and the recent corpus graph with its many subagents\ncollecting analogies and connections, evidence as subgraph, and so on. Analyze every book mentioned\nin the video along with its unique ideas, and relate it to as many parts of Deliberus as possible,\nand the meta process of creating Deliberus etc. Ultrathink\"*. (One name is replaced in brackets: a\nfriend is not named on a public page.) Shown the plan (all twelve in one\nnote, a section each, read by the session with no helpers, proposals kept in the note), he wrote:\n*\"Go!\"*\n\n**The video.** Stephen Petro, *These 12 Books Will Permanently Change What Your Mind Is Capable\nOf*, 14 minutes, uploaded 2026-10-05.\n\n**Which purpose this serves.** By the unruled list in\n[what-deliberus-is-for.md](what-deliberus-is-for.md) § 8.6: that he sees the land more clearly. It\nis a side path beside the four open reviews, taken because he asked for it. It adds no row for him\nto rule on, and nineteen candidates that wait in § 16.\n\n**The same in plain words** (2026-10-12, on his ask to have the findings explained better): every\nlesson with everyday cases and the minute in the video, and ten places where this reading meets the\nreading of Wittgenstein's *Tractatus* merged that day:\n[twelve-books-in-plain-words.md](twelve-books-in-plain-words.md).\n\n## How it was read\n\n- **The transcript is exact.** The creator published no captions. Two machine readings of the audio\n  by two makers were compared word for word; they agree on 97.7% of the words and differ in 44\n  places. Names were set by the creator's own list of the books. The founder settled the last two\n  places by ear: at the eleventh book the creator says *Efficiency* and the title is *A Treatise on\n  Efficacy* (*\"he misspoke\"*), and *strict ideas* is what is said. The text is the creator's and is\n  kept privately; this note quotes it in short phrases.\n- **Four of the books were read first, earlier the same day**, from reviews and critics:\n  [four-books-the-corpus-had-not-read.md](four-books-the-corpus-had-not-read.md) (Ziliak and\n  McCloskey, Woodward, Paul, Jullien). Their sections here are short. They check that note against\n  the exact transcript and add what bears on the parts of Deliberus he named.\n- **The other eight are read here**: what the notes already hold (the index of notes was asked\n  first), then the publisher's page or an encyclopedia entry, reviews, and the strongest critic\n  found.\n- **None of the twelve books was read.** Every outside claim below comes through the sources at the\n  end, read as search excerpts and not as whole texts, except where a section says a page was read.\n  A detail from memory is marked *(recalled)*.\n\n---\n\n## 0. In short\n\n- **The strength number says how well a claim has stood. Three of the books name what it leaves\n  out.** How large the thing claimed is (Ziliak and McCloskey), how much the claim still says after\n  its conditions were added (Lakatos), and how much else rests on it, which is the one of the three\n  Deliberus computes (the hinge). A reader will take a strong badge for *this matters* (§ 13.1).\n- **A premise can move into a claim.** Lakatos's best response to a counterexample takes a step the\n  proof assumed and writes it into the theorem as a condition. That is a sixty-year-old account of\n  why *part of* and *reason for* were confused in the linking step, and of what the graph does not\n  record: how a claim changed when it was attacked (§ 3).\n- **W. Ross Ashby was not in the corpus, and his law is the theory under three things it holds\n  dear**: the *does not fit* answer on every closed list, variety of positions among contributors,\n  and reuse as the only way checking keeps pace with persuasion. It also explains why our own rule\n  files grow (§ 7, § 14).\n- **A word can have several models that usually agree.** Lakoff's *mother* is neither one sense\n  with a fuzzy edge nor two senses. The concept layer records senses and has no place for this\n  third shape, and his boundary cases are the probe it has been missing (§ 5).\n- **The friend's design of one proposition for each claim could propose kinships to the analogy\n  daemon.** It finds the same thing in the same relation by lookup. Turned around, the same\n  relation between different things is a candidate likeness, which nearness of wording cannot find\n  (§ 13.3).\n- **The rule written under a link is one of several that would do.** Hitchcock's defence of Toulmin\n  says a link claims that some general rule holds, and a named rule is support for that. So a rule\n  that falls does not take its link with it, which is how a machine's sentence is stored today\n  (§ 6).\n- **The video's last advice is a protocol for the two-person rehearsal**: one builds the strongest\n  case, another judges it without having to defend it, then they swap (§ 12).\n- **The making of Deliberus shows the books' patterns too.** The instruction files are the video's\n  company that adds an approval step after every mistake; a ruling kept in his words is the\n  opposite of Lakatos's tidy retelling; and he cannot rule on land he has not seen, which is Paul's\n  problem (§ 14).\n\n---\n\n## What the video is doing\n\nIts opening claim is that a good book should *\"force your mind to perform an operation it could\nnot perform before\"*, and it closes each book with questions to ask. That is this corpus's own test\nfor a borrowed idea: it earns its place by yielding a move\n([what-belongs-in-the-ontology.md](what-belongs-in-the-ontology.md) § 6b). § 15 lists the video's\nquestions beside where Deliberus asks each today.\n\nTwo more things about its form:\n\n- **The order is an argument.** Size of an effect, then cause, then the exception, then likeness,\n  then category, then argument, then system, then the chooser who changes, then the model that\n  leaves things out, then limits, then indirect action, then reasoning in company. It climbs from\n  one claim to a group, which is the scale ladder of\n  [fractal-scales-and-temporal-frame.md](fractal-scales-and-temporal-frame.md).\n- **Each book is handed to the next by a likeness**, for instance *\"just as decisions can conceal a\n  future self, models can conceal a present reality\"*. In Paul the one who judges changes. In Scott\n  the picture omits. The two share a verb and no mechanism. Hofstadter and Sander have a name for a\n  likeness made on the spot to carry a point: a caricature analogy (§ 4). The video's eleven joins\n  are a small test set for the daemon's checks, and most would not pass.\n\n---\n\n## 1. Ziliak and McCloskey, *The Cult of Statistical Significance*\n\n**In the video.** A study of 50,000 people finds a benefit that is statistically significant, and\nlater tests find it again. The average benefit is tiny. Ask how large, how uncertain, and whether\nthat size matters here.\n\n**Already read**: [four-books-the-corpus-had-not-read.md](four-books-the-corpus-had-not-read.md)\n§ 1. No critical question in any scheme asks the size of an effect; the size is stored on a claim\nand no strength reads it; *significant* has a technical sense and an ordinary one. The exact\ntranscript confirms that note's summary line.\n\n**What the exact transcript adds.**\n\n- **Replication is in the example, and it does not help.** Finding the effect again makes it surer\n  that there is one. It says nothing about its size. On the graph, several supporting studies raise\n  a claim's standing, and if the claim reads *the treatment helps*, the badge grows while the thing\n  established is *the effect is not zero*.\n- **That is two claims.** *An effect exists* and *the effect is large enough to matter for this\n  decision*. The second has a threshold that depends on the case. That is one of the eight seams\n  along which a claim can be cut, the one between a fact and the standard it is measured against\n  ([decomposition-axes.md](decomposition-axes.md)). The splitter's list names that seam and its\n  test is not taught. And *large enough to matter* is a weighing, so it belongs under the weighing\n  questions.\n- **Our own number has the same shape.** A claim's strength says how well it has stood against what\n  was said about it. It does not say that the claim matters. § 13.1 takes this up.\n\n**For evidence as a subgraph.** A study under a claim needs three things kept apart: what was\nmeasured, how surely, and whether that much bears on the claim as worded. Today the first two are\nfields and the third is nowhere.\n\n**For the friend's design** (§ 13.3). The value slot of a proposition is where a size would sit. Two\npropositions with the same sign and very different sizes fall on the same link, and for a decision\nthey are different findings.\n\n---\n\n## 2. Woodward, *Making Things Happen*\n\n**In the video.** Teenagers who use social media more report more anxiety. Does the one cause the\nother, the other the one, or does something else drive both? Ask what would happen to one if the\nother alone were changed, with the background held as it was.\n\n**Already read**: the four-books note § 2. Woodward names the theory under two open proposals (how\nwide a rule may be, ledger AD25; whether a finding carries over, Q11), and says exactly what the\nhinge measures: what a computed conclusion depends on, never what causes what.\n\n**What this reading adds.**\n\n- **One observation, three rival claims.** The video's three explanations all rest on the same\n  correlation. On the graph that is one piece of evidence supporting three claims that cannot all\n  hold. What tells them apart is never more of the same observation. It is what would happen under\n  a change. That is the difference the evidence note already draws between evidence of\n  difference-making and evidence of mechanism\n  ([what-counts-as-evidence.md](what-counts-as-evidence.md)).\n- **A likeness is a claim that something holds across a change of field.** Woodward's test for a\n  generalisation is the range of changes it survives. The analogy daemon's test for a kinship asks\n  whether the mechanism is the same in the second field. These are one test. The *relevantly\n  different* features the video asks for under Hofstadter are the things whose change breaks the\n  generalisation. So a counter-analogy is a case, found in the second field, where it breaks.\n- **For the friend's design.** *Is associated with* and *causes* are different relations, and a\n  causal proposition carries a change in it: which thing is changed, by how much, with what held\n  fixed. Two propositions *X raises Y* that mean different changes do not contradict. That is the\n  design's own second condition (scope) in causal form.\n\n---\n\n## 3. Lakatos, *Proofs and Refutations*\n\n**In the video.** A classroom argues over the claim that, for a polyhedron, vertices minus edges\nplus faces is two. A proof seems to settle it until students bring odd counterexamples. Each one\nmakes the class inspect the proof, change a definition, narrow the claim, or find a better theorem.\nWhen someone finds an exception to your idea, ask whether it shows a faulty step, a hidden\nassumption, a conclusion too wide, or a poor definition.\n\n**What the notes hold.** The hidden lemma, and writing a link out by refutation\n([philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md) § 10). The wager read\nas a research programme whose reformulations nobody tracks\n([threads-synthesis-and-gaps.md](threads-synthesis-and-gaps.md)). And one of the three kinships the\nreaders found: telling a refinement from a retreat to a safer definition\n([the-analogy-daemon.md](the-analogy-daemon.md) § 8).\n\n**The book's own ideas** (two parts of the journal version read in excerpt; a 2025 chapter by\nDutilh Novaes; a review by the Mathematical Association of America):\n\n- **Six things one can do with a counterexample.** Give the claim up. Call the case a monster and\n  keep it out by changing what the word means (*monster-barring*). List it as an exception and\n  narrow the claim to a safe region (*exception-barring*). Describe the case again so that it\n  complies (*monster-adjustment*). Find the step of the proof that the case breaks and write that\n  step into the claim as a condition (*lemma-incorporation*). Or look for a deeper proof to which\n  the case is no longer a counterexample.\n- **Two kinds of counterexample.** A *local* one breaks a step of the proof and leaves the claim\n  standing: the claim may be true and this proof of it bad. A *global* one breaks the claim. His\n  rule for a local one is to replace the broken step, not to drop the claim.\n- **A proof's use.** One student's line: its virtue is *\"not that it compels belief, but that it\n  suggests doubts\"*. Listing the steps a claim rests on is how one learns where to look for a case\n  against it.\n- **Words stretch without anyone deciding.** The refuters widen *polyhedron* to cover things nobody\n  first meant. The defenders look as if they keep changing the definition, and may be the ones\n  holding the first meaning still. A definition can cover far more than its author intended.\n- **Every condition added costs content.** A claim hedged until nothing refutes it says little. His\n  later rules push the other way: find the proof that explains more.\n\n**What its critics say.** The dialogue is a tidied history, and the real one is in the footnotes;\nhistorians (Koetsier; Musgrave and Pigden, as cited in the 2025 chapter) say it fits a handful of\ncases. The 2025 chapter adds that the book is written as a dialogue and is not about dialogue: its\nactors are the concepts, and the people are incidental. The association's reviewer finds him\nunfair to Cauchy and inclined to see in black and white. That Feferman judged the method to fit\nlittle of mathematics is *(recalled)*.\n\n**Where it touches Deliberus.**\n\n- **A premise can move into a claim.** Lemma-incorporation takes a reason for the claim and makes\n  it a part of the claim. In our terms, a claim on the supporting side of a link becomes a\n  condition inside the conclusion. The linking step stored *this is a premise of that* as a part-of\n  link for six months. The census of part-of links had a label for the confusion, *a reason filed\n  as a part*, and used it on 70 of 197 links; of the 113 written the right way round it read 59% as\n  not parts ([auto-connect-upgrade.md](auto-connect-upgrade.md) § 2c). Lakatos says\n  why the two are so easy to confuse: in live reasoning, the same sentence is first a step of the\n  argument and later a condition of the claim. The test that separates them is his: if this\n  sentence falls, is the claim false (a part), or is only this argument for it gone (a reason)?\n- **Local and global are already in the strength rulings.** A refuted supporter adds nothing and\n  takes nothing away (ruled 2026-09-30, for now), which is his local counterexample: the proof is gone and\n  the claim stands. A required part that falls caps its whole, which is his global one.\n- **The graph records attacks and not what the attacked did next.** His six responses are a type\n  the ontology lacks. The pieces exist: an attack, a qualifying link, a defining claim, a part-of\n  link, a superseded claim. Nothing says *this claim was narrowed in answer to that attack* or\n  *this word was redefined in answer to it*. The motte note's four tests for telling refinement\n  from retreat ([the-motte-and-the-periphery.md](the-motte-and-the-periphery.md)) need exactly that\n  record. So does the wager's own honesty: its ceilings must count against it, and a wager that\n  survives by narrowing should show its narrowings.\n- **Standing without content.** A claim can reach a high strength by shrinking. Nothing on a claim\n  page says how many conditions it has taken on since it was first asserted. § 13.1.\n- **Our own corrections are mostly his kind.** Of the 304 findings in the index of notes that say a\n  passage is corrected elsewhere, 191 narrow a claim and overturn nothing\n  ([corpus-graph/README.md](corpus-graph/README.md) § 22.6a). The corpus grows by adding\n  conditions, as he says knowledge does.\n- **Who moved the word.** For the concept layer: when a word's use changes during a dispute, the\n  side whose wording changed is not always the side whose meaning changed.\n\n---\n\n## 4. Hofstadter and Sander, *Surfaces and Essences*\n\n**In the video.** Analogy is central to thought: a new situation is understood by seeing a familiar\nstructure under an unfamiliar surface. A company that answers every mistake with one more approval\nstep is like a doctor who treats each side effect with one more drug. Ask where you have seen this\nrelationship before, then name what corresponds and what is relevantly different.\n\n**What the notes hold.** The 2001 essay and the Copycat program, with the book itself marked unread\n([perceiving-possibility-and-the-analogy-organ.md](perceiving-possibility-and-the-analogy-organ.md));\nthe psychology of comparing cases, including that people are reminded by looks and convinced by\nstructure ([what-a-rule-rests-on.md](what-a-rule-rests-on.md) § 4); what is left of two cases when\ntheir subjects are taken away ([corpus-graph/what-a-move-is.md](corpus-graph/what-a-move-is.md)).\n\n**The book's own ideas** (five reviews: Times Higher Education, the Irish Times, Bookforum, The\nPhilosopher, and a reading group's notes):\n\n- **Putting a thing in a category is making an analogy.** Words have their boundaries extended one\n  likeness at a time. A reviewer quotes the authors that behind every word *\"there lurks a blurry\n  richness\"*.\n- **Most categories have no name.** The authors' own example, as a reviewer tells it: someone picking\n  up bottle caps on a tour of a great temple brings back a small boy, years before, absorbed in\n  ants at the rim of a great canyon. Nothing names what the two share.\n- **How fine a category is depends on the purpose.** While drinking, my glass and yours are\n  different things. In the dishwasher they are the same.\n- **Caricature analogies**: likenesses invented on the spot to convince someone.\n- **Science as a chain of analogies**, with a long chapter on Einstein that reviewers praise.\n\n**What its critics say.** Bookforum's reviewer: once all thought is analogy, the word's power to\nexplain *\"comes to seem tautological\"*, and it is hard to see what would count as a wrong analogy.\nThe book has no notes and no bibliography, and does not place itself beside other work on the\nsubject. Others: several hundred pages of examples and little about how a mind does it.\n\n**Where it touches Deliberus.**\n\n- **The concept layer and the analogy daemon are one faculty at two distances.** Deciding that a\n  use of a word belongs to a sense is, on the book's account, the same act as seeing a kinship\n  between two debates. That fits the proposal to store uses and read a sense as a group of them\n  ([store-the-uses-read-the-meaning.md](store-the-uses-read-the-meaning.md)). It also explains a\n  count from the index of notes: asked for likeness across fields, readers returned mostly an\n  earlier ruling that bears on a passage (3 kinships among 42 findings, then 5 among 54). Seeing\n  that a ruling covers a passage is a near analogy. A kinship is a far one.\n- **The book gives the faculty and no test for a wrong likeness.** That is the daemon's reason for\n  being built in two passes. On the claim graph the test exists: the analogy scheme asks for\n  relevant differences and for a stronger counter-analogy (`deliberus/extraction/schemes.py`). On\n  the notes it does not: a kept kinship carries one reason and nothing set against it.\n- **Naming is the creative step.** A kinship kept in the index is a pair and a sentence. The\n  daemon's design mints a claim that says what the two share. That claim is a new category with a\n  new name, which is where a likeness becomes portable, and where it can be wrong.\n- **Sameness depends on purpose.** The glass example is the finding the sameness work reached from\n  philosophy: two claims are the same if they are used the same way\n  ([claim-sameness-philosophical-readings.md](claim-sameness-philosophical-readings.md)).\n- **The video's own example describes our instruction files** (§ 14).\n\n---\n\n## 5. Lakoff, *Women, Fire, and Dangerous Things*\n\n**In the video.** Most people take a category for a box with defining properties. Lakoff argues\nthat many are organised around prototypes, bodily experience, cultural models, and extensions from\ncentral cases. The title is a noun class in Dyirbal, an Aboriginal Australian language, that holds\nwomen, fire and various dangerous things, joined by a chain that makes sense inside one system of\nideas. *Mother* is his familiar case. When an argument turns on a label (*successful*, *natural*,\n*normal*), ask what the prototype is, which cases are central, and what happens at the edges.\n\n**What the notes hold.** Almost nothing by name: one mention of framing, and two source lists. The\nidea is present without him: a word has a shared core (reference, category, prototype) and a\nperiphery that differs by person and community\n([core-periphery-and-the-concept-layer.md](core-periphery-and-the-concept-layer.md)).\n\n**The book's own ideas** (the publisher's page, an encyclopedia entry, an excerpt of chapter 6 used\nas course material):\n\n- **Idealised models.** We organise what we know in simplified models of a piece of the world. A\n  category's better and worse examples come from how well a case fits the model.\n- ***Mother* is a cluster of models**: giving birth, supplying the genes, raising the child, being\n  the father's wife (the excerpt read names birth and raising; the rest of the list is\n  *(recalled)*). In the central case they all hold. *Birth mother*, *adoptive mother*, *foster\n  mother*, *surrogate mother* are the cases where they come apart, and each modifier names which\n  model holds or fails.\n- **Radial categories.** A central case, with extensions that are learned by convention. An\n  extension is *motivated* by the centre and cannot be predicted from it by a rule.\n- **Dyirbal.** Four noun classes. The linguist Dixon's account, which Lakoff builds on: things join\n  a class through the part of life they belong to and through myth, so no property is shared by\n  all members.\n\n**What its critics say.** Vervaeke and Green (*Women, Fire, and Dangerous Theories*, 1997; abstract\nread): he confuses concepts with ways of conceiving, showing that prototypes exist does not rule\nout classical accounts, *motivated* explains too little, and the Dyirbal classes may be symbolic\ngroupings and no general folk taxonomy.\n\n**Where it touches Deliberus.**\n\n- **A third shape for a contested word.** The concept layer asks whether a word's dispute is about\n  what it refers to or about its criteria, and it records senses. Measured in September, 49 of 90\n  concepts read as split in two, largely because the detector writes two senses for every contested\n  term. *Mother* is neither one sense with a fuzzy edge nor two senses. It is one word with several\n  models that agree in ordinary cases. Many of the words people fight over look like this:\n  *fair* (equal shares, equal chances, what was earned, what is needed), *natural*, *intelligent*.\n  A dispute then has a precise place: both sides agree which models hold in the case at hand, and\n  differ on which model governs the word here.\n- **Edge cases are the probe.** The core-periphery note proposes probing a person's sense with\n  quick yes-or-no cases. Lakoff says which cases to make: the ones where the models come apart. Ask\n  *is a surrogate a mother?* and the answer shows a weighting no definition would.\n- **Why a cut through a concept is an argument.** The resident line says that cutting *through* a\n  concept usually gives premises, not parts. A cluster concept shows why: no one model is required\n  (an adoptive mother gave no birth) and none is enough. Each model is a rule that holds unless\n  something defeats it. So a category is a bundle of Toulmin's warrants (§ 6), which is also the\n  order in which the video places the two books.\n- **Our own kinds are radial, by design.** The graph has one kind of node, the claim, and a field\n  for its kind. A sentence a machine writes under a link was ruled a claim on 2026-10-10: one that\n  nobody asserts. That is *foster mother*: a member that lacks one model of the central case, with\n  a modifier that says which. The test that reads every query for the kind\n  (`tests/test_machine_sentence_readers.py`) guards against treating it as central. Claims minted\n  from reported speech are the same shape.\n- **Every list of types we keep is radial too.** The five families of finding were named on the\n  same day one kind of finding was moved from one family to another. Lakoff's word for how members\n  join is *motivated*. The *does not fit* answer admits it.\n- **Chains.** A Dyirbal class holds together by links from member to member. Natural categories\n  chain, and a chain is no identity: the rule against merging by a silent chain of *same as* stands\n  for that reason.\n- **For the friend's design** (§ 13.3). Deciding that two mentions are one thing is cheap for an\n  individual (a named trial, a law, a person) and is a judgement of sense for a kind (*basic\n  income*). The second belongs to the concept layer.\n- **For lenses.** A chain that *makes sense inside one system of ideas* is what a worldview lens\n  would have to carry: the models a community brings to a word.\n\n---\n\n## 6. Toulmin, *The Uses of Argument*\n\n**In the video.** Formal logic says whether a conclusion follows from premises of a given form.\nReal arguments rarely work so. Students remember more when they take notes by hand, so\nuniversities should ban laptops: the evidence may support the claim about memory and does not give\nthe policy. A hidden warrant joins them, perhaps that keeping the most information should decide\nclassroom policy. The full model: claim, data, warrant, backing, the conditions under which the\nclaim holds, and rebuttals. The weakest part of an argument is often the bridge nobody examined.\n\n**What the notes hold.** A great deal, in 31 files. The rule under a link is a licence and no\nmissing premise ([philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md));\nwhere a rule comes from ([what-a-rule-rests-on.md](what-a-rule-rests-on.md)); his place among the\nancestors ([science-lineage-philosophy.md](science-lineage-philosophy.md)).\n\n**What the critics and defenders add** (Hitchcock's two papers and his seminar notes; a review of\nFreeman's book; the introduction to a 2006 volume on the model):\n\n- **The standing objection**: in real texts one cannot tell data from warrant (van Eemeren,\n  Grootendorst and Kruiger). Freeman rejects warrants altogether after the longest critique in\n  print.\n- **Hitchcock tested it.** He drew 50 arguments at random from a research library and could single\n  out a rule that licenses the step in 49.\n- **His repair.** A warrant is usually not a piece of the argument to be reconstructed. It is asked\n  for when the argument is judged, and the question is *how might you get there?*, not *how do you\n  get there?* What a step claims is that **some** general rule covering it holds. A named rule is\n  support for that claim. Several rules suit any one step, and each has its exceptions.\n- **Fields.** Toulmin held that what backs a warrant differs by field. Johnson's objection, which\n  Hitchcock accepts: many warrants belong to no field and are rough common sense.\n- **Reception.** Philosophers were hostile at first. Law, psychology and American speech teaching\n  took the model up.\n\n**Where it touches Deliberus.**\n\n- **A link outlives its rule.** Since 2026-10-10 a machine's sentence under a link is stored as a\n  part of the link's relevance claim and moves no strength until a person argues that the part is\n  required (`deliberus/machine_sentences.py`). Hitchcock gives the reason this default is right. A\n  rule written under a link is one way the step could hold. If it falls, another may do, which is\n  Lakatos's local counterexample again (§ 3). It follows that *required* can seldom be argued for a\n  rule, and can for a fact about the case. That bears on the open question of which kind of link\n  should join a step to what was written under it (ledger AG1 to AG5).\n- **It agrees with the count.** A person could find the licensing rule in 49 arguments of 50. A\n  machine told to write one wrote it for 29 links of 30\n  ([philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md) § 10). What neither figure shows is\n  whether two readers would write the same rule, and Hitchcock's point is that they need not.\n- **The video's warrant is a value, and Toulmin's one slot hides two things.** From *students\n  remember more* to *ban laptops* the bridge is a value (keeping information matters) and a\n  weighing (it outweighs access, differences among learners, and other uses). The video's\n  rebuttals are rival values. The ontology already keeps these apart: a link from fact to value\n  rests on a value, and its dispute goes to the weighing questions.\n- **The qualifier sits on the link.** Freeman argues that words like *probably* describe the step\n  from premises to conclusion and are no part of the conclusion. That is the two-factor rule of\n  2026-09-28: a link carries its own relevance, apart from the standing of what it joins.\n- **For evidence as a subgraph.** *Evidence* in Toulmin is three things: the datum, the licence to\n  use it, and what backs the licence. Each can be disputed apart. That is the founder's position of\n  March that evidence decomposes into its parts.\n\n---\n\n## 7. Ashby, *An Introduction to Cybernetics*\n\n**In the video.** The most technical book on the list. A system stays stable under disturbance only\nif its regulator has enough different responses: the law of requisite variety. A company with one\nrule for every unhappy customer cannot handle fraud, confusion, a broken product and an emergency\nalike. When something keeps failing, map the state you want to keep, the disturbances, the feedback\nthat shows them, and the responses you have.\n\n**What the notes hold.** Nothing. The earlier note counted one mention; that *Ashby* is a different\nresearcher, in a list of studies of categorisation. Stafford Beer, who built a practice of\nmanagement on him, is not in the corpus either.\n\n**The book's own ideas** (an encyclopedia entry, the Principia Cybernetica page, and three papers\non the law):\n\n- **The law.** The variety of outcomes cannot be brought below the variety of disturbances divided\n  by the variety of the regulator's responses. His phrase: only variety can destroy variety.\n- **Two ways to meet it.** Give the regulator more responses, or let fewer kinds of disturbance\n  reach it. Beer made a practice of the pair: amplify one side, attenuate the other.\n- **The good regulator** (Conant and Ashby, 1970): every good regulator of a system must be a model\n  of that system.\n\n**What its critics say.** The law is a necessary condition and not a sufficient one. Its strong\nform ignores buffering: something that passively absorbs a disturbance needs no response to it\n(Principia Cybernetica). Outside toy cases nobody can count the states. Glanville separates\nregulating from controlling by restriction, and notes a physical limit on how much variety any\nregulator can hold. That the good-regulator theorem proves less than its title says is\n*(recalled)*.\n\n**Where it touches Deliberus.**\n\n- **It is the theorem under the *does not fit* rule.** A classifier with seven answers, facing more\n  than seven kinds of case, must put some case in the wrong box. The resident line says every\n  closed list needs a value for what does not fit. By Ashby's law that value does not add an\n  answer. It passes the unhandled case to something with more variety: a person, or the place where\n  new schemes are grown. The law that every taxonomy breaks on the next register says that the\n  disturbances keep growing.\n- **It says what contributors are for.** The human-reserved classes are the considerations a corpus\n  silently lacks and the frame nobody inside can see. Both are variety the machine cannot make. The\n  group premise holds only if the minds differ\n  ([the-group-premise-evidence.md](the-group-premise-evidence.md)), and what the derivability probe\n  found people adding was evidence from a position. So the number that matters is how many\n  different positions contribute, and the count of contributors matters less.\n- **It gives the scrutiny gap its form.** Persuasion at machine speed is disturbance. Checking is\n  the regulator. Checking that keeps pace with persuasion is requisite variety held over time.\n  There are two ways, as above. No reader can match a million messages. So the way open is the\n  second: let fewer distinct things arrive, by deciding that two of them are the same claim.\n  **Deciding sameness is the attenuator that loses nothing when it is right. Summarising is the one\n  that loses what the regulator needed**, the rare claims\n  ([saturation-and-the-long-tail.md](saturation-and-the-long-tail.md)).\n- **Scott's disasters are attenuation done to the world** (§ 9). A state that cannot hold a\n  forest's variety can raise its own or lower the forest's. It planted rows. The rule for Deliberus\n  is the other way: when a form cannot hold what a person said, the form gives way, and the\n  person's words stay as the anchor.\n- **Buffers.** The critics' point has a use here. An unexamined claim moves nothing, however it is\n  linked. A machine's part weighs nothing. Splitting a claim into three parts or twelve gives the\n  same strength. Each absorbs a kind of junk without anyone having to detect it. That is the form\n  the corpus says to look for first against anyone who games a rule: take away the payoff.\n- **The good regulator and the person who checks.** To check a machine's proposal a person needs a\n  picture of the matter as rich as the proposal. That is the worry that review turns into approval\n  of whatever is well formatted ([cognitive-commons-and-deliberus.md](cognitive-commons-and-deliberus.md)).\n- **It explains our own rule files** (§ 14).\n\n---\n\n## 8. Paul, *Transformative Experience*\n\n**In the video.** Advice on decisions says to imagine each future and value it. Some experiences\ncannot be known until lived, and living them changes what you value. Ask who this choice might turn\nyou into, what you can know beforehand, and whether you would choose partly for the discovery.\n\n**Already read**: the four-books note § 3. She argues from decision theory for a line the founder\nratified (the experience is the person's alone; its causes and alternatives can be mapped), and\nadds one field the library of crossings lacks: whether the person reporting has made the choice.\nThe exact transcript confirms that note's summary line, and has her answer too: choosing for the\ndiscovery itself.\n\n**What this reading adds.**\n\n- **A person's sense of a word has a date.** What *tired*, *free* or *love* mean to someone changes\n  with a crossing. If the concept layer records whose sense a sense is, it should record when, as\n  [attunement-words-and-the-concept-layer.md](attunement-words-and-the-concept-layer.md) implies.\n- **A reviewer who has seen a machine's proposal is a changed judge.** The principle that nobody\n  audits their own frame notes that a person anchored on a proposal has taken on its framing. That\n  is Paul's structure in small: the one who judges after is not the one who would have judged\n  before, and cannot go back. It is why the protocol for the sameness sitting asks for a blind part.\n- **Deliberus is this kind of thing to a newcomer.** What it is like to see one's own quarrel laid\n  out cannot be told. Mercier and Sperber report that people underrate what arguing in company does\n  for them (§ 12). Together these are the case for the two-person rehearsal over any description,\n  and for the hand-made example on the landing page: Bykvist's criticism of Paul is that parts of a\n  new experience can be compared with known ones, and a thermostat quarrel is such a part.\n\n---\n\n## 9. Scott, *Seeing Like a State*\n\n**In the video.** Large institutions make populations, forests and cities understandable through\nmaps, measures and categories. That allows coordination, and it can erase the local knowledge that\nlets a system work. A forest reduced to a count of saleable timber is easier to measure, and the\ncount ignores soil, wildlife and local use. The map grows clearer *\"as the territory becomes\npoorer\"*.\n\n**What the notes hold.** The fullest adversarial reading in the corpus opens with it\n([legibility-under-power-red-team.md](legibility-under-power-red-team.md), *Seeing Like a Graph*):\nDeliberus proposes a land register of reasoning; practical knowledge that cannot be drawn does not\ncount; what extraction loses is unmeasured. Porter's warning that such tools get captured rides\nwith it ([fractal-priors-for-success.md](fractal-priors-for-success.md)).\n\n**What the reviews add** (an H-Net review, DeLong's review, Tania Li's article, an essay by Sustrik):\n\n- **Four conditions for disaster**: a society made legible, a faith in redesign by plan, a state\n  able to impose the plan, and people unable to resist it.\n- **His advice**, as the H-Net reviewer gives it: take small steps, prefer what can be undone, plan\n  on surprises, plan on people's inventiveness.\n- **DeLong**: the argument about local knowledge is Hayek's, unacknowledged.\n- **Li**: governing is done by many parties besides the state; ask what a scheme does, not only why\n  it failed.\n- **Sustrik's essay**: people often want to be legible, because living by practical cunning is\n  exhausting; and the book says little about why planning from above sometimes works.\n\n**Where it touches Deliberus.**\n\n- **His advice is, nearly word for word, how the graph is written.** What can be undone: a\n  machine's parts are marked and reversible, and a contribution is added and never overwritten.\n  Surprises: the *does not fit* answer. Inventiveness: a reader may say the whole frame is wrong,\n  and that counts as a contribution. The red team's first hole has a published remedy. Two\n  parts of it are built. The third is ruled and unbuilt: no page yet lets a reader say the frame is\n  wrong. The fourth, small steps, is how the work is done, and nothing enforces it.\n- **The four conditions are the missing constitution's four clauses.** Three notes ask for a\n  constitution and none writes it. Scott says what it must take away. Deliberus makes reasoning\n  legible (the first condition) and believes that doing so helps (the second). So the safeguards\n  have to sit in the other two: an operator with no power to impose a rendering, and contributors\n  who can refuse one, leave, and take the map with them. Two exist as principles: a person may\n  refuse a statement of their position, and the project aims to be a reference implementation and\n  no walled garden.\n- **Legible to whom.** His subjects resisted because the map served the tax collector. The\n  incentives note asks when being legible pays the person who is legible\n  ([incentives-analysis.md](incentives-analysis.md)). The essay's point supports its answer: people\n  seek legibility that works for them. The question from Jullien applies to every surface: does it\n  make a person more visible to us than we are to them.\n- **The map changes the land.** Scott's stronger claim is that the forest was replanted to fit the\n  count. Our form of it is people coming to speak in the shapes the graph offers, and a frame that\n  gets harder to leave as more is built on it.\n- **Standard measures and standard senses.** That local measures carried local purposes, and that\n  the metre replaced them with one that carried none, is *(recalled)* from the book's first\n  chapter. The concept layer's average sense, written by a model, is a metre for meanings: nobody's.\n  The proposal to treat it as a reference point from which each person's difference is recorded is\n  the remedy in his spirit.\n- **Our own index is a state's map.** Counted today: the view of what waits for the founder's word\n  reaches 68 of 441 notes ([corpus-graph/README.md](corpus-graph/README.md) § 22). The ledger makes\n  positions legible to ruling. A thought with no row is the forest floor.\n\n---\n\n## 10. Gigerenzer and Selten (editors), *Bounded Rationality: The Adaptive Toolbox*\n\n**In the video.** Theories of rational choice imagine unlimited information, time and computing.\nNobody has them. The book asks how simple rules can use the structure of an environment. The video\nadds that such rules must be used with care, and says the authors set out how. The book is a\nworkshop report with many authors, and a reviewer finds its group reports its best chapters.\n\n**What the notes hold.** Little: that his programme denies *error* is the right word for a\nheuristic ([cognitive-bias-codex-and-human-contribution.md](cognitive-bias-codex-and-human-contribution.md),\n[what-belongs-in-the-ontology.md](what-belongs-in-the-ontology.md)).\n\n**The programme's own ideas** (the publisher's page, three reviews, Gigerenzer and Brighton's 2009\npaper):\n\n- **A toolbox, with no master tool.** Each simple rule has three parts: where to look next, when to\n  stop, how to decide.\n- **A rule is good for an environment.** His term is ecological rationality.\n- **Less can be more.** *Take the best* looks at clues in order of how well each has predicted and\n  decides on the first that tells the options apart. In prediction across twenty data sets it\n  matched or beat regression. The reason given: with little data, a rule that ignores how the clues\n  depend on each other does better than one that estimates the dependence badly.\n\n**What its critics say.**\n\n- **People do not reliably use it.** Newell and Shanks (2003): use of *take the best* rose when\n  information cost more, and even then many kept buying clues after one had decided the matter.\n  Bröder found no support for the claim that everyone uses it.\n- **Choosing the rule is the hard part.** Bröder and Newell (2008): combining many clues is not\n  costly for people. Deciding how to decide is, and the toolbox does not say how a tool is picked.\n- **One process may do.** Gathering evidence up to a threshold, strict or lenient, covers both the\n  simple and the thorough strategies.\n- **Reviews of the book**: a war on two fronts, against both optimisers and bias-hunters; thin on\n  social and cultural rules.\n\n**Where it touches Deliberus.**\n\n- **Outside support for the weakest-part rule.** The strength rulings treat required parts by their\n  weakest, so one bad part decides and nothing else makes up for it. Corroborating reasons add up.\n  The argument for the first was that it stays right when parts share an origin and nobody knows\n  how much ([weakest-link-arithmetic-and-the-merge-hunch.md](weakest-link-arithmetic-and-the-merge-hunch.md)).\n  That is Gigerenzer's finding in another field: where evidence is sparse and dependent, a rule\n  that ignores dependence beats one that guesses at it.\n- **A caution for the evidence note.** It found that answers to critical questions are counted\n  where every tradition says to weigh them. This tradition says that counting with equal weights\n  predicts about as well as fitted weights when there is too little data to fit them. That argues\n  for keeping the count until a weight can be measured. It is a qualification, not a reversal.\n- **Our limits are his rules, and each has an environment.** Ask the question on which most hangs\n  first; stop when no answer could change what is shown; read each claim's twelve nearest partners;\n  show about ten boxes. The third was tested his way, against the full computation: it kept 20 of\n  24 attack links. It was measured on one pair of opposed texts. Across fields, nearness of wording finds\n  nothing (dogfood run 6). A limit should say where it was measured.\n- **His environments do not answer back.** A simple rule that becomes known can be gamed.\n  Advertising buys recognition, which defeats the rule *choose the one you recognise*. That is the\n  corpus's distinction between the three classes of adversary: free to carry a rule within a\n  class, a rebuild across one.\n- **Where the person's effort goes.** If choosing what to look at is the costly step, and combining\n  what one sees is not, then the system should choose the next question and the person should\n  answer it. That is the principle that the machine holds the user's hand through the descent, with\n  an experiment under it.\n- **Machine passes.** A closed question read the same on the cheapest models and a question of\n  degree on none ([pass-quality-and-the-choice-of-model.md](pass-quality-and-the-choice-of-model.md)). A small\n  task is a simple rule fitted to an environment, and an error rate belongs to the task as written.\n\n---\n\n## 11. Jullien, *A Treatise on Efficacy*\n\n**In the video.** A common Western model of action is to design an ideal outcome and impose it.\nAgainst it stand ideas he reconstructs from Chinese thought: the strategist studies the\nconfiguration and changes conditions so that the wanted outcome becomes easier to produce.\n\n**Already read**: the four-books note § 4. Jullien describes how this project acts in the world\n(cheap persistence, position before the avalanche), and his own last chapters show the same way of\nacting turning into manipulation, which Deliberus is built against. The exact transcript confirms\nthat note's summary line. The creator says *Efficiency* for the title; the founder heard it.\n\n**What this reading adds.**\n\n- **Three of the video's books say one thing from three sides.** The structure is in the situation\n  and not in the plan. Scott tells it from the side of those planned upon (§ 9), Gigerenzer from\n  the side of the one deciding (§ 10), Jullien from the side of the one acting. Ashby gives the\n  arithmetic (§ 7).\n- **It is how the ontology has been grown.** The corpus calls settling the philosophy first a\n  closed path, tests the ontology on built examples, and expects every list of types to break on\n  the next kind of material. Categories are left to come about from runs.\n- **It is how sessions came to be steered** (§ 14): by placing a cue where the act happens, where\n  earlier files commanded.\n\n---\n\n## 12. Mercier and Sperber, *The Enigma of Reason*\n\n**In the video.** We picture reasoning as a private path to truth. Their hypothesis is that reason\nevolved mainly to justify our claims to others and to judge theirs. The video says the hypothesis\nis disputed and that private reasoning keeps its value. Alone we build cases for our first\nintuition. Under the right social conditions others find the weaknesses we missed. So after an\nimportant decision do not ask whether someone agrees. Separate the roles: one builds the strongest\ncase, another judges it without having to defend it, then swap. A group whose disagreement is well\nstructured can be more intelligent than any member.\n\n**What the notes hold.** It is a foundation: reasoning is largely after the fact\n([academic-foundations.md](../academic-foundations.md)); the group premise holds under five\nconditions, of which the graph supplies three\n([the-group-premise-evidence.md](the-group-premise-evidence.md)); nobody audits their own frame.\n\n**What the critics and the authors' replies add** (Dutilh Novaes 2018; Dogramaci 2020; a review by\nNoveck with the authors' reply; a review by Chater and Oaksford):\n\n- **Dutilh Novaes**: the bias toward one's own side is hard to defend as an advantage to the\n  individual; the authors bring in a division of labour, which is an advantage to the group, while\n  denying selection at that level. People detect lies at about chance and believe by default, which\n  sits badly with the vigilance the theory predicts.\n- **Dogramaci**: it is not shown that getting others to believe what you believe is an advantage.\n- **Noveck**: say what would refute the theory, and which problems groups help with. The results\n  come from problems that have an answer which can be shown.\n- **The authors' reply.** When people who disagree share an interest in the answer, the best\n  argument tends to win; on problems with a showable answer one person who has it nearly always\n  convinces the rest. Their exception matters here: in one study people who had gone wrong were not\n  moved by the correct working until it addressed the step where they went wrong. They add that\n  people consistently underrate how well majority rule, averaging and argument work.\n\n**Where it touches Deliberus.**\n\n- **The video's advice is already the architecture, and it is a protocol for the rehearsal.** A\n  separate reader checks every split; the outside check has a reader who is not the author compare\n  a new proposal with what is settled; a reader of the notes is never their writer. For the\n  two-person rehearsal the same advice is a script: each writes the strongest case for the other's\n  claim, the other accepts or refuses the rendering, and they swap. Refusing a rendering is the one\n  act the corpus reserves to the person whose position it is.\n- **A correct answer convinces where it lands on the faulty step.** That is the case for\n  addressability in one finding: pointing at the step is what converts noticing into changing one's\n  mind. It joins the result that people detected a machine's error and deferred anyway because they\n  could not say what was wrong.\n- **The asymmetry holds for machines.** Models correct an error shown as someone else's far more\n  often than the same error in their own output, and the corpus's rule is that a pass never grades\n  itself. A judge that flips when argued with in conversation is the evaluator's role\n  collapsing into the conversation partner's, which is why a verdict here is a record and\n  challenges arrive as claims.\n- **A written-out step is a justification, and that is enough.** By this theory a reason given on\n  request is made for an audience and need not be what caused the belief. The design does not need\n  the cause. It needs something that can be answered.\n- **What may not be borrowed.** The strong results are for problems with a showable answer. The\n  wager is about values. *Truth wins* does not carry over to it, and the group-premise note already\n  holds the polarising result for arguments exchanged within one frame.\n- **Nothing here rests on the evolutionary story.** The claims the corpus uses are about what\n  happens when roles are separated, and those are measured.\n\n---\n\n## 13. Across the twelve, by the parts of Deliberus he named\n\n### 13.1 The ontology in progress\n\n**One number, and what it does not say.** A claim's strength is computed from what has been said\nfor and against it. Read with these books it measures one thing, standing, and is silent on three:\n\n| | What it is | Whose point | In Deliberus today |\n|---|---|---|---|\n| Standing | how well the claim has held against what was said | the strength layer itself | the badge |\n| Size | how large the thing claimed is | Ziliak and McCloskey | stored on some claims, read by no strength, asked by no question |\n| Content | how much the claim still says, after its conditions | Lakatos | not recorded |\n| Dependence | how much else moves if it falls | Woodward's test, turned on the model | the hinge score |\n\nA statistically significant result that is taken for an important one is the confusion of the\nfirst row with the second. A strong badge taken for *this matters* would be the same confusion on\nour own page. The hinge is the nearest thing here to what Ziliak and McCloskey ask for. It orders which\nquestion is asked next, and no page of the site shows it beside a badge (the site's code was\nsearched for it).\n\n**Other points for the ontology**, each argued above:\n\n- how a claim changed in answer to an attack is not recorded (§ 3);\n- a premise can become a condition of the claim, which is why *part of* and *reason for* blur (§ 3);\n- a rule under a link is one candidate among several, so it supports and is seldom required (§ 6);\n- the kind field on a claim is the modifier of a radial category (§ 5);\n- a closed list of answers has a floor on its error that only an exit lowers (§ 7);\n- taking the weakest required part is the rule that survives unknown dependence (§ 10).\n\n### 13.2 Concept tracking\n\nSeven of the books bear on how a word's meaning is kept:\n\n- **Lakoff**: several models under one word, and edge cases as the probe (§ 5).\n- **Hofstadter and Sander**: a sense is a group of uses, grown one likeness at a time (§ 4).\n- **Lakatos**: a definition can cover more than its author meant, and a word can stretch in a\n  dispute without anyone deciding it (§ 3).\n- **Ziliak and McCloskey**: *significant*, and every word that hides a threshold (§ 1).\n- **Scott**: a standard sense is nobody's sense (§ 9).\n- **Paul**: a person's sense has a date (§ 8).\n- **Toulmin**: the same word is backed by different things in different fields (§ 6).\n\nTogether they describe a layer that keeps uses, groups them, records under each word the models in\nplay and which hold in a given case, and dates whose sense is whose. The first two parts are the\nproposal of [store-the-uses-read-the-meaning.md](store-the-uses-read-the-meaning.md). The third is\nnew here.\n\n### 13.3 A friend's design: one proposition for each claim\n\n**What is meant.** The founder's words were *\"[a friend]'s recent idea of predicate logic-like\ntracking\"*. This note takes that to be the design a friend sent by mail, described in\n[store-the-uses-read-the-meaning.md](store-the-uses-read-the-meaning.md) § 4b: for each claim about\nwhat is the case, a model reads one proposition (subject, relation, object or value, qualifiers,\nand the time the content is about); mentions of one thing are grouped; and links between things are\ncomputed from the propositions.\n\n**Answered 2026-10-12.** Asked whether that was the idea he meant, he placed it: *\"it was [a\nfriend]'s idea from when I visited him last week iirc\"*. The record from that week holds one design\nthat fits, and it is this one. The mail's own word for the middle slot is *predicate* (this corpus\nwrote *relation*), and its predicates come from a short closed list with a *does not fit* answer.\nIt was written out by the friend's coding assistant on a question the founder put to it, in answer\nto the friend's wish for something exact and compact where a store of sentences would be a tangle.\nSo the subsection is aimed right. What the record cannot show is whether the friend said more in\nconversation than the mail holds.\n\n**What the books say to it.**\n\n| Book | For the design |\n|---|---|\n| Ziliak and McCloskey | the value slot should hold a size and its uncertainty; same sign with a very different size is a different finding |\n| Woodward | *causes* and *is associated with* are different relations; a causal proposition names the change it is about |\n| Lakatos | the qualifiers are where an added condition lands; a proposition read without them is the first, naive form of the claim |\n| Hofstadter and Sander | matching by identical slots finds the same thing said twice and cannot find the same relation between different things |\n| Lakoff | grouping mentions is cheap for individuals and is a judgement of sense for kinds; the list of relations will be radial |\n| Toulmin | a proposition is the claim alone; the data, the licence and the exceptions live in the argument around it |\n| Ashby | a fixed list of relations has too little variety, and the share that does not fit measures by how much |\n| Scott | a graph of what is the case is a land register: good for looking things up, dangerous as an authority |\n\nThree of these are already the design's own added conditions (scope, no number on a world link, an\nopen list of relations with a *does not fit* answer). Two are new:\n\n- **Individuals and kinds.** The middle layer should group individuals. Whether two kinds are one\n  is the concept layer's question.\n- **The design can propose kinships.** The readers of the notes found that nearness in meaning\n  cannot propose a likeness: the best ranking held 25 of 57 where chance gives 15. A second try\n  wrote a short card for each passage and let code propose pairs from the cards. Those 18 pairs and\n  16 of the readers' own went to outside judges, who called 18 of the 34 the same move\n  ([the-analogy-daemon.md](the-analogy-daemon.md) § 8). A proposition is a card\n  with fixed slots. Two propositions with the same relation and different things in the slots are a\n  candidate kinship. So the layer built to find one thing said in two places could, read the other\n  way, tell the daemon where to look. That is the structure-over-surface finding from the\n  psychology of analogy, used as an index.\n\n### 13.4 The analogy daemon\n\n- **Its test has a definition**: a kinship is a generalisation that survives the change of field,\n  and a counter-analogy is a case where it breaks (§ 2).\n- **Its two passes have a reason**: the one who proposes is a poor judge of the proposal (§ 12).\n- **Its output has a life**: a naive likeness, then a counterexample, then a condition added, then\n  a rule of known width. That is Lakatos's sequence (§ 3), and it ends where the question of a\n  rule's breadth begins (ledger AD25).\n- **It can borrow the claim graph's own questions.** The analogy scheme already asks for relevant\n  differences and a counter-analogy. A kept kinship between notes could carry one named difference,\n  which the rehearsal never produced (§ 4).\n- **Where it reads is a question of variety.** What gets found is where somebody looked, and the\n  notes a reader is sent to were chosen by hand (§ 7).\n- **A proposer exists in the friend's design** (§ 13.3).\n- **The standing caution holds.** A likeness may organise and may not justify. The video's own\n  joins are examples (*What the video is doing*, above).\n\n### 13.5 The index of notes and its readers\n\n- **The readers are the daemon's rehearsal**, as the founder saw; § 13.4 applies to them.\n- **Most findings are a ruling that bears on a passage, and that is a near analogy** (§ 4). The\n  families of finding were ruled today: corrected elsewhere, missing piece, same idea, at odds,\n  fact mismatch. On the book's account *same idea* is the far end of a scale whose near end is *corrected\n  elsewhere*; that is a reading, and nobody has tested it.\n- **Most corrections narrow** (191 of 304), which is Lakatos's observation about how claims\n  improve (§ 3).\n- **The index sees what has a row** (§ 9). The other notes are reached only when somebody opens\n  them.\n- **The reminder that reading is owed is a regulator with a sensor** (§ 7): the repository decides\n  when the queue is worth saying, and the hook says it at the moment a note is opened.\n- **A reader is never the author**, and every quoted passage is checked by code before a second\n  reader opens both ends (§ 12).\n\n### 13.6 Evidence as a subgraph\n\nRead together, the books give the parts a piece of evidence has under a claim:\n\n| Part | From | Kept today |\n|---|---|---|\n| What was observed, where and by whom | the evidence note's own finding that evidence ends in an encounter | source, passage, attribution |\n| How large and how sure | Ziliak and McCloskey | fields on some claims, unread by strength |\n| What it would take to tell rival explanations apart | Woodward | the causal scheme's questions |\n| The licence to use it for this claim, and what backs the licence | Toulmin | the scheme, the relevance claim, a written-out rule in trials |\n| Whether it strikes the claim or one argument for it | Lakatos | the strength rulings on parts and supporters |\n| Who chose to bring it | Mercier and Sperber | nothing: a claim's supporters are brought by its friends, so their one-sidedness is expected |\n| Whether the kind of evidence can be counted at all | Scott | the ladder of methods counts by type; evidence from a position has a note and no weight |\n| How to combine several | Gigerenzer | weakest part for what is required, a sum for what corroborates, a count for answers |\n\n---\n\n## 14. The making of Deliberus\n\n**The video's company is our instruction files.** A mistake by a session becomes a rule in a file\nthat every session loads. Each rule is locally right. Together they grow into something that causes\nnew misses, because a longer file is read less well. Ashby says why the rules multiply: the kinds\nof mistake are many, and a regulator needs a response for each. He and Beer also name the other\nway, which the corpus already uses for product hazards and less often for its own conduct: remove\nthe occasion. A guard that makes a mistake impossible needs no rule. The filter for what may stay\nresident (would a session without it assert the opposite) is attenuation on the regulator's side.\n\n**A directive has Ashby's four parts.** The state to keep, the disturbance, the feedback that shows\nit, the response. The corpus's finding that a rule cued on a word about to be written fires, and a\nrule cued on noticing a state of mind does not, is a finding about the third part: a regulator\nneeds a sensor. *A guard that never fires looks exactly like a guard that works* is a regulator\nwith none.\n\n**Steering moved from command to disposition.** Earlier files said *must* and *never*. The later\nones put a short cue where the act happens, and a hook where a cue is not enough. That is Jullien's\ncontrast (§ 11). His warning applies as well: arranging conditions so that someone does a thing of\nthemselves is how manipulation works, so the cues that steer a session are written down and the\nfounder can read every one.\n\n**A ruling kept in his words is the opposite of Lakatos's method.** Lakatos wrote the history as it\nshould have gone and put what happened in the footnotes. The rule here, made after 25 of 31 rulings\nwere found in Claude's paraphrase, is that a ruling is recorded in his sentence. The session\nhistory and the dated lines under corrected passages are the footnotes, kept in the body.\n\n**The roles are separated because the reasoner cannot check itself.** The record of this project's\nown errors has one pattern: a fluent sentence stood until the founder contradicted it from memory,\nor until a reader other than its author opened the source. That is Mercier and Sperber's asymmetry\n(§ 12), and the reason the outside check and the index of notes were built.\n\n**He is the regulator, and the good-regulator theorem says what that costs.** To catch a wrong\nsentence about the project he must hold a model of the whole project. The index of notes is an\nattempt to give sessions some of that model, so that fewer errors need him.\n\n**He cannot rule on land he has not seen.** The ledger had 348 open rows on the evening of\n2026-10-11. A ruling asked for before he has walked the ground is Paul's case: the one who would\nknow how to judge it is the one who has already been there. The walkthroughs exist for that, and\nwhat he said of them on 2026-10-08 is this list's first book applied to speech: they sounded like\na sports narrator *\"burping up individual numbers as though that's the point\"*.\n\n**The ledger is the legible part** (§ 9), and the session habits are fast rules whose environments\nnobody wrote down (§ 10): three helpers at most, two hundred words, one decision in a reply.\n\n**The transcript was itself a small case.** Two machine readings of one recording agreed on 97.7%\nof the words. Twenty-four of their 44 differences were names, which is where a reader of the video\nwould have been misled most. Agreement between two readers was high and said nothing about the two\nplaces both got the same way, which a person settled by ear. Steady is not the same as right.\n\n---\n\n## 15. The video's questions, and where Deliberus asks them today\n\n| Book | The question it trains | Asked in Deliberus today |\n|---|---|---|\n| Ziliak and McCloskey | how large, how sure, and does that much matter here | by no scheme; the size is stored and unread |\n| Woodward | what would happen to this if that alone were changed | the causal scheme asks about confounders, order and mechanism; *which change is meant* is not asked |\n| Lakatos | is the exception a bad step, a hidden assumption, a claim too wide, or a poor definition | an attack lands on a claim or on a link; which of the four it was is not recorded |\n| Hofstadter and Sander | where have I seen this relationship, what corresponds, what differs | the analogy scheme asks the last two on the claim graph; the daemon is designed; the notes have readers |\n| Lakoff | what is the central case, and what happens at the edges | contested terms are detected and given senses; no edge cases are generated |\n| Toulmin | what joins the evidence to the claim, what backs that, when does it fail | the scheme and its questions; the relevance claim; written-out steps in trials |\n| Ashby | what state, which disturbances, what shows them, which responses | every closed list has its exit; no surface states the four for itself |\n| Paul | who might this choice make me, and what can I know beforehand | the ruling on mapping the inside of a hard choice; nothing built |\n| Scott | what did the model leave out that the system lives on | the ledger of omissions in a synthesis; the map's row naming what kind was left out; loss in extraction is otherwise unmeasured |\n| Gigerenzer and Selten | which simple rule fits this environment | stop rules and limits exist; none names where it was measured |\n| Jullien | which condition can be changed so the outcome comes easily | how the project works; no surface |\n| Mercier and Sperber | who builds the case and who judges it | separate readers for machine work; for people, the rehearsal is planned and unbuilt |\n\n---\n\n## 16. Candidates, held in this note\n\n**None is filed in the ledger.** That was the plan he approved. The first six are carried from the\nfour-books note, which held them back until this reading.\n\n1. A critical question on the size of an effect.\n2. A check that *significant* keeps its technical sense through extraction.\n3. Woodward named as the theory under the proposals on a rule's breadth and on carrying over.\n4. The hinge described to readers as a change made to the model and not to the world.\n5. A field on a report of a hard choice: has its author made the choice and lived with it.\n6. The question for any surface that learns about people: does it make them more visible to us than\n   we are to them.\n7. One sentence, wherever a badge is explained: it says how well a claim has stood, and not how\n   large, how much it says, or how much rests on it.\n8. Record how a claim changed in answer to an attack: dropped, word narrowed, exception listed,\n   case described anew, condition added, or none of these.\n9. Show how many conditions a claim has taken on since it was first asserted.\n10. When a person takes up a machine's rule under a link, the default relation is support for the\n    link's relevance; *required* is argued for facts about the case. A note for ledger AG1 to AG5.\n11. A third record in the concept layer: the models in play under a word, and which hold in a case;\n    with edge cases generated as the probe.\n12. In the friend's design, group individuals only; send kinds to the concept layer.\n13. Add to that design's smallest experiment: list pairs of propositions with the same relation and\n    different things, and have a person read them as candidate kinships.\n14. A kinship kept between notes carries one named difference.\n15. Each stop rule and limit says where it was measured and which class of adversary it has met.\n16. Keep counting answers to critical questions until a weight can be measured.\n17. A script for the two-person rehearsal: each writes the strongest case for the other's claim,\n    the other accepts or refuses it, then they swap.\n18. Scott's four conditions as the checklist for the constitution: who reads the map, who can\n    change it, who can refuse a rendering, who can leave with it.\n19. Before a rule is added to an instruction file, ask whether its occasion can be removed. This is\n    already the practice for product hazards; Ashby is its reason.\n\n---\n\n## 17. What was not done\n\n- **The books were not read.** Each section rests on the sources below.\n- **Marked *(recalled)* and to be checked**: Feferman on Lakatos; the weakness of the\n  good-regulator theorem; Scott on local measures; and, from the earlier note, Woodward's phrase and\n  Pettigrew's later book.\n- **Not opened**: Toulmin's book against Hitchcock's account of it; Dixon on Dyirbal; the chapters\n  of the Dahlem report; Mercier and Sperber's book against its critics' summaries.\n- **Not run**: nothing here was tried on the graph. Every statement about what the code does is\n  taken from the instruction file's dated lines or from the files named beside it, and three were\n  opened for this note (`deliberus/machine_sentences.py`, `deliberus/extraction/schemes.py`, the\n  section list of the friend's design).\n- **The video's eleven joins were not put through the daemon's checks.** One was read (§ *What the\n  video is doing*).\n- **Corrected with this note**: the earlier note's count that Ashby was in the corpus once. He was\n  not; the name there is another researcher's.\n\n## Sources\n\nRead 2026-10-11 as search excerpts unless a line says otherwise.\n\n**Lakatos.** *Proofs and Refutations* parts III and IV, journal version (excerpts read):\nhttps://classes.matthewjbrown.net/teaching-files/hps/lakatos3.pdf ·\nhttps://classes.matthewjbrown.net/teaching-files/hps/lakatos4.pdf · Dutilh Novaes, *Proofs as\nDialogues* (2025): https://link.springer.com/chapter/10.1007/978-3-031-88213-5_3 · Mathematical\nAssociation of America review:\nhttps://old.maa.org/press/maa-reviews/proofs-and-refutations-the-logic-of-mathematical-discovery-0\n\n**Hofstadter and Sander.** Times Higher Education:\nhttps://www.timeshighereducation.com/books/surfaces-and-essences-analogy-as-the-fuel-and-fire-of-thinking-by-douglas-r-hofstadter-and-emmanuel-sander/2005717.article\n· Irish Times:\nhttps://www.irishtimes.com/culture/books/surfaces-and-essences-analogy-as-the-fuel-and-fire-of-thinking-by-douglas-hofstadter-and-emmanuel-sandler-1.1410120\n· Bookforum:\nhttps://www.bookforum.com/print/2001/two-books-seek-to-explain-how-our-minds-work-their-way-through-the-maze-of-consciousness-11211\n· The Philosopher: https://www.the-philosopher.co.uk/2013/09/review-article-surfaces-and-essences.html\n· a reading group's notes: https://massline.org/ScienceGrp/SurfacesAndEssences.htm\n\n**Lakoff.** University of Chicago Press:\nhttps://press.uchicago.edu/ucp/books/book/chicago/W/bo3632089.html · encyclopedia entry:\nhttps://en.wikipedia.org/wiki/Women,_Fire,_and_Dangerous_Things · chapter 6 as course material:\nhttps://dai.fmph.uniba.sk/courses/CSCTR/materials/CSCTR_05sup_Lakoff_women_fire_ch6_1987.pdf ·\nVervaeke and Green (1997), abstract: https://www.tandfonline.com/doi/abs/10.1207/s15327868ms1201_5\n\n**Toulmin.** Hitchcock, *Toulmin's Warrants*:\nhttps://www.humanities.mcmaster.ca/~hitchckd/Toulminswarrants.pdf · Hitchcock, *On the generality\nof warrants*: https://www.humanities.mcmaster.ca/%7Ehitchckd/generality.pdf · his seminar notes:\nhttps://www.humanities.mcmaster.ca/~hitchckd/toulmin.htm · review of Freeman's *Dialectics and the\nMacrostructure of Arguments*:\nhttps://informallogic.ca/index.php/informal_logic/article/download/2542/1983 · *Arguing on the\nToulmin Model* (2006): https://link.springer.com/book/10.1007/978-1-4020-4938-5\n\n**Ashby.** Principia Cybernetica, *The Law of Requisite Variety*:\nhttps://pespmc1.vub.ac.be/REQVAR.html · encyclopedia entry:\nhttps://en.wikipedia.org/wiki/Variety_(cybernetics) · Zwick, *Requisite Variety and the Second\nLaw*: https://pdxscholar.library.pdx.edu/cgi/viewcontent.cgi?article=1188&context=sysc_fac ·\nGlanville: https://asc-cybernetics.org/2004/GlanvillePaper.htm · a survey article:\nhttps://eapad.dk/article/requisite-variety-and-its-implications-for-the-control-of-complex-systems/\n\n**Scott.** H-Net review: https://www.h-net.org/reviews/showrev.php?id=3853 · DeLong's review, as\nposted to a list: http://mailman.lbo-talk.org/2000/2000-April/009001.html and quoted at\nhttps://www.econlib.org/archives/2017/09/delong_on_scott.html · Li (2005), *Beyond \"the State\" and\nFailed Schemes*: https://anthrosource.onlinelibrary.wiley.com/doi/10.1525/aa.2005.107.3.383 ·\nSustrik, essay (undated in the copy read):\nhttps://www.greaterwrong.com/posts/iiDzt5qhesmQZiNAj/james-c-scott-seeing-like-a-state\n\n**Gigerenzer and Selten.** MIT Press:\nhttps://direct.mit.edu/books/edited-volume/4232/Bounded-RationalityThe-Adaptive-Toolbox · reviews:\nhttps://academic.oup.com/ej/article-abstract/113/485/F189/5079528 ·\nhttps://ideas.repec.org/a/eee/jeborg/v53y2004i3p431-434.html ·\nhttps://exa.ai/library/publication/s2nyb0930bg · Gigerenzer and Brighton, *Homo Heuristicus*:\nhttps://pmc.ncbi.nlm.nih.gov/articles/PMC3629675/ · Newell and Shanks (2003):\nhttps://www2.psy.unsw.edu.au/Users/BNewell/NS2003.pdf · Newell, Weston and Shanks (2003):\nhttps://ideas.repec.org/a/eee/jobhdp/v91y2003i1p82-96.html · Bröder and Newell (2008):\nhttps://www.cambridge.org/core/journals/judgment-and-decision-making/article/challenging-some-common-beliefs-empirical-work-within-the-adaptive-toolbox-metaphor/3586A89055EB355EBA20330BFE425458\n\n**Mercier and Sperber.** Dutilh Novaes (2018), *The enduring enigma of reason*:\nhttps://onlinelibrary.wiley.com/doi/10.1111/mila.12174 · Dogramaci (2020):\nhttps://www.cambridge.org/core/journals/episteme/article/abs/what-is-the-function-of-reasoning-on-mercier-and-sperbers-argumentative-and-justificatory-theories/214AA03FC41B89E8861EF85EBFB8E03B\n· Noveck's review: https://dialnet.unirioja.es/descarga/articulo/6865386.pdf · the authors' *Replies\nto Critics*: https://dialnet.unirioja.es/descarga/articulo/6865387.pdf · Chater and Oaksford's\nreview: https://onlinelibrary.wiley.com/doi/10.1111/mila.12181\n\n**Ziliak and McCloskey, Woodward, Paul, Jullien**: the sources listed in\n[four-books-the-corpus-had-not-read.md](four-books-the-corpus-had-not-read.md).\n\n## Cross-references\n\n[twelve-books-in-plain-words.md](twelve-books-in-plain-words.md) (the plain-language version) ·\n[the-tractatus-and-deliberus.md](the-tractatus-and-deliberus.md) (ten places where the two readings\nmeet are in the plain-language version, Part 3) ·\n[four-books-the-corpus-had-not-read.md](four-books-the-corpus-had-not-read.md) ·\n[the-analogy-daemon.md](the-analogy-daemon.md) · [what-a-rule-rests-on.md](what-a-rule-rests-on.md) ·\n[philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md) ·\n[writing-out-the-unwritten.md](writing-out-the-unwritten.md) ·\n[store-the-uses-read-the-meaning.md](store-the-uses-read-the-meaning.md) ·\n[core-periphery-and-the-concept-layer.md](core-periphery-and-the-concept-layer.md) ·\n[the-motte-and-the-periphery.md](the-motte-and-the-periphery.md) ·\n[what-counts-as-evidence.md](what-counts-as-evidence.md) ·\n[decomposition-axes.md](decomposition-axes.md) ·\n[auto-connect-upgrade.md](auto-connect-upgrade.md) ·\n[weakest-link-arithmetic-and-the-merge-hunch.md](weakest-link-arithmetic-and-the-merge-hunch.md) ·\n[legibility-under-power-red-team.md](legibility-under-power-red-team.md) ·\n[the-group-premise-evidence.md](the-group-premise-evidence.md) ·\n[saturation-and-the-long-tail.md](saturation-and-the-long-tail.md) ·\n[corpus-graph/README.md](corpus-graph/README.md) ·\n[corpus-graph/what-a-move-is.md](corpus-graph/what-a-move-is.md) ·\n[what-deliberus-is-for.md](what-deliberus-is-for.md)\n"}