{"path":"research/what-a-rule-rests-on.md","content":"# What a rule rests on: cases, likeness, and who can say what matters\n\n**Date**: 2026-10-08 · **Type**: the founder's thought, read against the corpus, the live graph, the\nstrength code and outside sources · **Status**: Claude's reading. Nothing here is ruled.\n\n**How it was read.** Three helpers read for this note, and their reports are kept beside it as they\ncame back: [the corpus](what-a-rule-rests-on/corpus-as-read-2026-10-08.md),\n[philosophy](what-a-rule-rests-on/philosophy-sources-as-read-2026-10-08.md), and\n[psychology with machine learning](what-a-rule-rests-on/cognitive-science-sources-as-read-2026-10-08.md).\nClaude did not open the outside sources itself. Each report says how much of each source was read,\nand § 12 lists what rests on an abstract or a second-hand account. The counts in § 2 come from the\nlive graph through [a script](what-a-rule-rests-on/links_by_kind_of_argument.py)\n([what it printed](what-a-rule-rests-on/what-the-count-printed.txt)).\n\n**His words**, 2026-10-08. First: *\"Just one thought I had: aren't multiple cases that are\nextrapolated from the basis of induction? And isn't the basis analogy? Analogy as the core of\ncognition has been argued for in the past. Ground extremely deeply on what the corpus and academic\nresearch and philosophy has on this.\"* Then, to a first answer given from memory: *\"Yeah I think\nthat makes sense...? And the analogy daemon can probably supply answers to the question of \"alike\nin what matters\"?\"* And to the answer that the second half of that question is a claim about the\nworld: *\"Which is where relevance etc comes in right?\"*\n\n## 0. The short version\n\n1. **As an account of where rules come from, his thought holds.** Mill and Aristotle say it, and\n   experiments show it: people who compare two cases draw the principle out of them, and people who\n   read the same two cases apart mostly do not.\n2. **As an account of what makes a step good, likeness cannot carry it alone.** Any two things are\n   alike in some respect. *Alike in what matters* needs a claim about what bears on what, and that\n   claim is general.\n3. **So neither is the floor.** The general claim about what bears on what is learned from earlier\n   cases, one level up. Rules and cases are adjusted against each other. The authors who looked\n   hardest at this say so in nearly the same words, in philosophy, in law and in psychology.\n4. **No source read shows that induction reduces to analogy with nothing left over.** The strongest\n   claims of that kind were disputed by the authors who came next.\n5. **Hofstadter's claim is about categories. He does not argue the induction step.** His claim\n   bears on a different place in a written-out link: seeing this case as one of the rule's cases.\n6. **Likeness sits in two places under a link**: behind the rule, and in what ties the case to it.\n7. **The graph holds a fourth kind of step, which the three shapes of 2026-10-07 left out**: from\n   cases to a rule. It is more than twice as common as a step from a case to a like case.\n8. **Yes, this is where relevance comes in.** In the strength code a link's relevance is *how much\n   the first claim, granted, bears on the second*. On a link from an earlier case that is the\n   likeness question, and the code already files it there.\n9. **The analogy daemon can write a likeness out. It cannot settle one.** Over time the best guide\n   is the graph's own record of which likenesses held.\n\n## 1. What the corpus held before today\n\n- **Three shapes of a step** ([philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md)\n  § 5; ledger AD23, open): from a rule and a case; from a sign to what it indicates; from a case to\n  a like case. The same section quotes Peirce's induction, from a case and a result to a rule, and\n  does not list it as a shape.\n- **Where a rule comes from** is answered there only as Mill's register and Goodman's record (§ 6;\n  ledger AD25).\n- **No document said that induction rests on analogy, or the reverse.** Three documents list\n  analogy beside generalisation as kinds of inductive argument. One taxonomy the corpus cites\n  (Kienpointner) files the inductive example under arguments that establish a rule, and analogy\n  under neither using a rule nor establishing one.\n- **Analogy as the core of cognition appeared once**, in\n  [perceiving-possibility-and-the-analogy-organ.md](perceiving-possibility-and-the-analogy-organ.md)\n  (2026-09-02), cited from memory and used for the sameness of claims. No document drew on it for\n  how a link or a rule is formed.\n- **The machinery for analogies**: [the analogy daemon](the-analogy-daemon.md) (a design on paper),\n  the four checks of [convictions-across-scales.md](convictions-across-scales.md), the channel in\n  which a person supplies the comparison (endorsed 2026-09-02), and Amgoud's four considerations in\n  [the-irit-harvest.md](the-irit-harvest.md) § 3.\n- **His own words on record.** 2026-03-29: *\"But of course there's also inductive arguments,\n  right? And the whole thing about what constitutes a good bayesian inductive argument and so\n  on.\"* 2026-08-18: *\"there might be a strong expectation that I hold for it based on analogies to\n  similar patterns in the world\"*. 2026-08-30: *\"represented within Deliberus as well-evidenced\n  analogies/claims that express analogies, simply. Is there an analogy scheme??\"* Earlier sentences\n  of his on analogy and induction, from his private archive, are kept apart from the public\n  documents.\n- **Not in the corpus at all before today**: Gentner, Holyoak, Bartha, John Wisdom, Tenenbaum, and\n  the word *overhypothesis*.\n\n## 2. What is on the live graph\n\nMeasured 2026-10-08. A link's kind of argument is a label the pipeline's classifier stored on it.\nNo person has checked these labels.\n\n| Links between claims taken from texts or written by people | Count |\n|---|---|\n| that carry a kind of argument | 780 |\n| from a case to a general claim (stored as *example*) | 84 |\n| from a case to a like case (*analogy* 28, *precedent* 9) | 37 |\n| from a sign, or to what explains the evidence | 116 |\n| from a general statement to a case (cause to effect, an established rule, a classification) | 206 |\n\nTwo more links carry two labels of the case family at once and are in none of the rows. Counted\nover every link that carries a kind, the figures are 745 of 4,897; most of those links join the\nclaim of a standing question to its target, and they are left out above.\n\n**Two real links.** *Rural job loss from minimums* supports *Averages mask regional wage\ndifferences*: a case offered for a general claim. Of eight links of that kind drawn at random,\nseven run from a case to a general claim, judged by their titles. And in the map of the\nmachine-consciousness debate, Anil Seth's conclusion has five supporting links read as analogy,\namong them *weather simulation cannot produce real hurricane* and *Brains are embodied in bodies*.\n\n**What hangs on such a link, in code.**\n\n| Kind | Its standing questions | Where an answer lands |\n|---|---|---|\n| example | the example is representative of the pattern claimed | the link's relevance |\n| | the claimed outcome did occur in the example | the supporting claim |\n| | there is no counter-example in similar cases | the conclusion |\n| analogy | no relevant difference between the two cases undermines it | the link's relevance |\n| | the claimed property is true of the earlier case | the supporting claim |\n| | there is no stronger counter-analogy | the conclusion |\n| precedent | the past case is *\"sufficiently similar to this one in the respect that matters\"* | the link's relevance |\n| | circumstances have not changed since | the link's relevance |\n\nThe quoted wording is the code's own, from the phrasing it uses when a question's slots cannot be\nfilled. The last column is `CQ_BEARS_ON` in `deliberus/extraction/schemes.py`, read against the\nheader of `deliberus/relevance.py`.\n\n**Also true today.**\n\n- No pattern is named induction or generalisation. *Example* carries it.\n- The four checks exist in documents only. The analogy pattern in code carries the three questions\n  above, and a test pins that number (per the corpus read).\n- The arithmetic is the same for every kind of link, and several links into one claim are summed\n  with no discount for a shared origin.\n- A finder of structurally alike claims is built (`structural_kinship_candidates`). It proposes\n  pairs and writes nothing, and it was measured as the weakest of three finders on precision\n  ([the-reuse-flywheel-prior.md](the-reuse-flywheel-prior.md) § 2).\n- The thirty links of the written-out trial are support links drawn by the linking step between\n  texts, and none of them carries a stored kind (§ 6). The same holds for every such link: 871\n  live support and attack links drawn by that step carry no kind, so no standing question hangs\n  on them and the badge does not read them.\n- The analogy daemon is a design on paper.\n\n## 3. Is induction built on likeness? What the philosophy says\n\n**It says yes.**\n\n- **Hume**: *\"All our reasonings concerning matter of fact are founded on a species of Analogy,\n  which leads us to expect from any cause the same events, which we have observed to result from\n  similar causes.\"*\n- **Mill**: *\"All inference is from particulars to particulars: General propositions are merely\n  registers of such inferences already made, and short formulæ for making more\"*. And: *\"In the\n  strictest induction, equally with the faintest analogy, we conclude because A resembles B in one\n  or more properties, that it does so in a certain other property.\"*\n- **Aristotle**: the example is an induction that runs *\"of part to part, or like to like\"*, and\n  the universal comes out of many memories of the same thing.\n- **Keynes**, on Hume's eggs: *\"We argue from Analogy in so far as we depend upon the likeness of\n  the eggs, and from Pure Induction when we trust the number of the experiments.\"*\n- **Kant**'s printed logic keeps induction and analogy as twin inferences from particulars. Notes\n  from his lectures fold analogy into induction.\n- **Law and ethics.** Levi: legal reasoning goes from case to case, and the rule is made in the\n  act of comparing. Jonsen and Toulmin: a maxim is understood through the cases it covers.\n- **Indian logic**: the early five-part argument is analogical, and Dignāga's inference runs from\n  particulars to particulars.\n- **John Wisdom**, second-hand: every general principle a father cites ends in instances, as the\n  mother's reasons did, and the final appeal is to cases.\n- **Walliser, Zwirn and Zwirn (2021)** come nearest to his thought: every step of induction by\n  counting instances is analogical. They add that this does not reduce analogy to induction, and\n  that an analogy needs a further assumption that links the two sets of properties.\n\n**It says no, or turns the order round.**\n\n- **Goodman**: similarity is relative to a respect, and any two things share some. To say the\n  future will be like the past is empty, since it always is in some way. He then reverses the\n  dependence: *\"inductive practice may provide the basis for some canons of similarity\"*.\n- **Mill**, against his own formula: *\"Why is a single instance, in some cases, sufficient for a\n  complete induction, while in others, myriads of concurring instances, without a single exception\n  known or presumed, go such a very little way toward establishing a universal proposition?\"* He\n  poses the question and leaves it. Of analogy he says it is *\"a mere guide-post\"*.\n- **Peirce (1892)**: induction, hypothesis and analogy *\"are essentially inferences from\n  sampling\"*, and to say induction assumes that like circumstances give like events is *\"a\n  blunder\"*. In 1868 he treats analogy as a compound of induction and hypothesis, so as a product\n  and no foundation.\n- **Keynes**: both methods need an assumption about the world, that its independent variety is\n  limited.\n- **Norton**: an analogical inference is as good as the particular fact that backs it. Stated\n  similarities are mostly shorthand for that fact.\n- **Davies and Russell**: what licenses a step from one case is knowledge that one feature\n  *determines* another. Their example: make, model and year determine what a used car is worth.\n  That is weaker than a rule, so the earlier case still does work, and one case is then enough.\n- **Hesse and Bartha**, both second-hand: likeness counts only where the shared features are tied,\n  inside the earlier case, to the feature carried over, and no difference between the cases breaks\n  the tie. Bartha adds that an analogy can make a conclusion plausible and cannot give it a degree\n  of probability.\n- **The legal sceptics**, second-hand: similarities are without number, so a rule or a policy\n  picks the one that counts.\n- **Venn and Hájek**: a thing belongs to any number of classes, and how likely a conclusion is\n  depends on which class it is counted in.\n- **Quine** is on both lists. Every reasonable expectation rests on a sense of similarity, which\n  is inborn. And science replaces that sense with kinds fixed by theory.\n\n**It says each needs the other.**\n\n- **Goodman**: a rule of inference is justified by its fit with the particular inferences we\n  accept, and those inferences by their fit with the rules. Each is revised in the light of the\n  other. He calls the circle a virtuous one.\n- **Brewer**: an argument from an earlier case needs a rule that says how the shared features\n  relate to the inferred one. That rule is guessed from the examples and then tested against them.\n- **Levi, Sunstein, Waller, Schauer and Spellman, Jonsen and Toulmin**, each in their own terms:\n  principles are worked out with constant reference to cases, no rule interprets itself, and the\n  likeness a judge sees reflects the principles the judge brings.\n- **Mill** himself: a generalisation is *\"a security for good reasoning, not a condition of all\n  reasoning\"*.\n\n## 4. Is analogy the core of cognition? What the psychology says\n\n**Hofstadter's claim.** In the 2001 essay of that title he writes: *\"analogy is everything, or very\nnearly so, in my view.\"* His analogy is wider than a mapping between two situations. Any time\nsomething new calls up a stored category, that is analogy-making, since no instance equals an\nearlier one. The helper searched the essay for induction, generalisation and law and found none of\nthem. So his claim is that categories are bundles of analogies. The step from there to *induction\nrests on analogy* is not his.\n\n**Comparing cases is how people come to see a rule.** Three experiments, one result:\n\n- Given two stories with the same solution, 45% solved a new problem before any hint. Given one\n  story, 21% (Gick and Holyoak 1983).\n- Novices who compared two negotiation cases carried the principle to a new case 48% of the time.\n  Those who read the same two cases one at a time, 19% (Gentner, Loewenstein and Thompson 2003).\n- Preschoolers extended a new word by its pattern 60% of the time after comparing two examples,\n  and 14% after seeing one. Shown two examples apart, 13% (Christie and Gentner 2010).\n\nGentner and Medina (1998) draw the consequence. Repeated comparison turns what two cases share into\na schema with variables in place of objects, and such a schema can be applied as a rule. With the\nrule in hand, lining the new case up against it is still needed in order to apply it.\n\n**People are reminded by looks and convinced by structure.** In one study, stories that shared\nsurface features were recalled at .78 and stories that shared structure at .44. Asked how sound the\nmatch was, the same people rated the structural matches 4.40 of 5 and the surface matches 2.80\n(Gentner, Rattermann and Forbus 1993). Matches in appearance alone had, in Gentner and Medina's\nwords, virtually no predictive use.\n\n**What fixes the respect.** Three answers, and they are disputed among themselves.\n\n- *The comparison itself*: lining two structured cases up constrains which features count, and it\n  favours relations that are tied together over loose attributes (Gentner; Medin, Goldstone and\n  Gentner 1993).\n- *What one believes about causes*: similarity is at least as much a result of a person's theories\n  as a cause of them (Murphy and Medin 1985). How alike two categories are judged depends on which\n  property is being carried over (Heit's review, 2000). Models with causal knowledge make\n  inferences that run counter to similarity (Kemp and Tenenbaum 2009).\n- *Earlier generalisation*: likeness can be derived from how far a learner would generalise, and\n  not the other way round (Shepard 1987; Tenenbaum and Griffiths 2001).\n\n**One case can be enough.** Kemp, Perfors and Tenenbaum (2007): *\"Inductive learning is impossible\nwithout overhypotheses\"*, their word for constraints on which hypotheses a learner considers. If\nevery bag seen so far held marbles of one colour, a single black marble from a new bag says\nsomething about the whole bag. The constraint is itself learned from the earlier bags. Mitchell\n(1980) gives the general point for machines: a learner with no bias cannot go beyond the cases it\nhas seen.\n\n**Rules against likeness is not a clean split.** Hahn and Chater (1998): each can imitate the\nother, and neither in its pure form can do the work. The working models are mixed: a few simple\nrules with remembered exceptions, or stored cases with attention shifting between features.\nAgainst likeness as the whole story stand Marcus (infants extend an abstract pattern to new\nsounds) and Lake and Baroni (networks that fail where the parts must be recombined). In 2023 Lake\nand Baroni got that systematic behaviour from an ordinary network by training it on streams of\nmany small tasks: rule-like behaviour learned from many cases, one level up.\n\n**Machines.** A language model matched or beat students on several analogy tests (Webb, Holyoak and\nLu 2023). On versions with the alphabet altered, people held up and the models fell sharply (Lewis\nand Mitchell 2024), a result the first authors dispute (2025). Inside the models, learning from\nexamples in the prompt has been described as completing a pattern by its nearest earlier match\n(Olsson and others 2022), and one model's handling of word endings fits a model of stored examples\nbetter than a model of rules (Hofmann and others 2024). A 2025 study finds parts of a model that\nturn tokens into abstract variables.\n\nGentner and Medina's own summary of the whole dispute: *\"similarities without rules are empty,\nrules without similarities are blind.\"*\n\n## 5. The answer to his thought\n\n**Two questions were in it.** Where does a rule come from, and what makes a step from it good.\n\n**Where a rule comes from: compared cases.** § 3's first list and § 4's experiments agree, with one\nqualification: compared cases are one route to a rule and the sources name others (being told, a\ndefinition, a theory, an inborn standard).\n\n**What makes a step good: likeness, together with a claim about what bears on what.** To say two\ncases are alike says nothing until the respect is named, and which respect counts depends on what\nis being concluded. The sources call that further claim by many names: a determination, a fact\nthat warrants, a tie inside the earlier case, an overhypothesis, a rule that warrants the analogy.\nUnder every name it is general.\n\n**And that general claim comes from earlier cases.** This is the answer the later authors give to\nMill's question. One chemical experiment settles the matter because we already hold, from much\nwider experience, that samples of one substance behave alike. Any number of black crows settles\nlittle because we hold that colour varies inside a kind. The wider claim was learned from cases\ntoo, of a more general sort.\n\nSo the picture is a spiral. Cases give a rule when they are compared. Which likeness matters is\ngiven by a wider rule. The wider rule came from comparing cases one level up. No source read finds\na level at which this stops.\n\n**His squirrel, as an illustration** (Claude's, from no source). *Dead squirrels smell of\ndecomposition* comes from earlier dead animals that smelled. This squirrel falls under it because\nit is like them in what matters. One earlier case would nearly do, because we hold that bodies of\none kind decay alike. One grey squirrel would not show that squirrels are grey.\n\n**What stays open in the sources** is what fixes the likeness. § 4 gave three answers. None has\nwon.\n\n## 6. What it does to a written-out link\n\n**Likeness sits in two places.** The form tried on 2026-10-07 writes a link out as the first\nclaim, what ties this case to a rule, and the rule\n([writing-out-the-unwritten.md](writing-out-the-unwritten.md)).\n\n| Part of the written-out link | Where likeness sits |\n|---|---|\n| the rule | behind it: the earlier cases it was drawn from, and that they are alike |\n| what ties this case to the rule | in it: that this case is one of those cases |\n| the first claim | nowhere: it is a fact about the case |\n\nThe second row is Hofstadter's point, and Gentner and Medina's, and Wittgenstein's that no rule\nfixes its own use. It is also the judged step the corpus already meets as the sameness of two\nclaims ([the-reuse-flywheel-prior.md](the-reuse-flywheel-prior.md): identity that has to be judged\nis what sets the pace).\n\n**Four shapes.** Peirce's three arrangements of a rule, a case and a result give three shapes. A\nstep from a case to a like case he analysed as a compound. The corpus had three shapes on\n2026-10-07 and left out his induction, which the graph holds as the kind called *example*.\n\n| Shape | Stored as | Links (§ 2) | What is written out | The general claim underneath |\n|---|---|---|---|---|\n| from a general statement and a case | cause to effect, an established rule, a classification | 206 | the case, the tie, the rule | the rule, and that this case is one of its cases |\n| from a sign to what it indicates | sign, evidence to a hypothesis, best explanation | 116 | the law from the thing to its sign; no rival here | how much likelier the sign is with the thing than without it |\n| from a case to a like case | analogy, precedent | 37 | the earlier case; what is carried over; the shared feature | that the shared feature bears on what is carried over |\n| from cases to a rule | example | 84 | the case; that it is representative | that things of this kind are alike in this respect |\n\nIn every row but the first, what sits underneath is a claim about what bears on what. In the first\nrow the same claim sits behind the rule.\n\n**A written-out step from cases to a rule yields exactly such a claim.** Forced into the three-part\nform, *this raven is black, so ravens are black* needs the sentence *ravens are alike in colour*.\nThat is the question the code already hangs on these links: is the example representative.\n\n**The trial's thirty links carry no stored kind.** A prediction was written before looking\n([the file](what-a-rule-rests-on/shape_prediction.txt)): links of a rule kind would more often end\nwith a step in which every part is needed than links of the other kinds. It could not be scored.\nAll thirty are support links drawn by the linking step between texts, and that step stores no kind\non them. So the shape of such a link has to be named when the step is written, which\n[philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md) § 5 had asked for on\nother grounds.\n\n**A connection to the concept layer** (Claude's reading). The corpus describes a sense of a word as\na set of examples, a list of criteria, and a weighting that differs by position\n([core-periphery-and-the-concept-layer.md](core-periphery-and-the-concept-layer.md)). The models\nof categories in § 4 have the same three parts: stored cases, rules, and attention that weights\nfeatures by the task. *What matters* is that weighting. For a word it is the periphery of its\nsense. For a link it is the link's relevance. A worldview lens, on this reading, includes a weighting\nof respects. (⚠ Narrowed the same evening, § 13.2: weighting is one thing a lens does among several,\nand the weighing partly follows the conclusion wanted.)\n\n## 7. Where this meets the strength rulings\n\n**Yes: this is where relevance comes in.** By the ruling of 2026-09-28 (ledger Q33; his words:\n*\"About the decision: your proposal is blessed, make it so.\"*) a link moves its target by the\nsupporting claim's own strength times the link's relevance. The header of `deliberus/relevance.py`\ndefines relevance as a reading of *\"how much s, granted, bears on t\"*. For a link from an earlier\ncase, that is the second half of *alike in what matters*.\n\n| The question about a like case | Its place in the ruled arithmetic |\n|---|---|\n| is the earlier case as described | the supporting claim's own strength |\n| do the two cases share a feature, and does it bear on what is concluded | the link's relevance |\n| is there a similar case where the conclusion fails | an objection to the conclusion itself |\n\n**A likeness can already be argued.** Relevance is read from a claim attached to the link (*s\nbears on t, because …*) when a person has written one, and that claim is supported and attacked\nlike any other. So a dispute over which likeness matters has a home in the built system. What is\nmissing is anything that writes such a claim. On 2026-10-08 the live graph held no relevance claim\nat all ([what the script printed](what-a-rule-rests-on/what-the-count-printed.txt)), so on the 37\nlinks from a case to a like case the question stands unanswered.\n\n**What the likeness is worth, on one link.** From the run in\n[the-squirrel-and-the-strength-calculus.md](the-squirrel-and-the-strength-calculus.md) § 3, with\nthe earlier claim held at 0.9: the conclusion reads 0.519 while nobody has argued the likeness,\n0.569 when the likeness is fully established, and about 0.518 under a standing objection to it.\nThe likeness is the whole of what the link adds beyond what is presumed.\n\n**Several cases.** Today each is one more link into the conclusion and the links are summed.\nKeynes says something else: *\"an increase in the number of experiments is only valuable in so far\nas, by increasing, or possibly increasing, the variety found amongst the non-essential\ncharacteristics of the instances, it strengthens the Negative Analogy\"*. On his account a further\ncase earns its place by differing in what does not matter, which shows that those respects do not\nmatter. That gives the same cases two possible homes: as further links into the conclusion, or as\nsupport for the rule or for the likeness. Counted in both, they are counted twice. This is the\nquestion of how several reasons combine (ledger Q22; the September review's questions 4 to 6),\nmet from a new side. Nothing is proposed for the arithmetic here.\n\n**One account of variety.** The corpus treats the number and variety of cases in three places with\nno common account: Amgoud's Number and Diversity\n([the-irit-harvest.md](the-irit-harvest.md) § 3), variety of evidence\n([weakest-link-arithmetic-and-the-merge-hunch.md](weakest-link-arithmetic-and-the-merge-hunch.md)\n§ 9b), and the first of the four checks. Keynes's sentence is a candidate for the common account.\nThe warning in § 9b stays in force: a repeat of the same case can be worth more than a varied one.\n\n**A caution about numbers.** Bartha holds that an analogy can make a conclusion plausible and\ncannot give it a degree. On 2026-10-07 no model read the degree of a link the same way twice\n([pass-quality-and-the-choice-of-model.md](pass-quality-and-the-choice-of-model.md)). Both say the\nsame thing about a figure taken from a likeness.\n\n## 8. Can the analogy daemon say what matters?\n\n**The question has two halves, and a machine can answer the first.**\n\n1. *What do the two cases share?* This can be answered from the cases alone. Lining them up is\n   what picks out the candidates (§ 4).\n2. *Does what they share bear on the conclusion?* This cannot be read off the cases. It changes\n   with what is being concluded and with what one believes about causes. It is a claim about the\n   world, and it is where people disagree.\n\n**The design of 2026-08-30 already asks for most of this.** Its contract returns a claim that two\nthings share a mechanism, the cases, the strongest case against, and a verdict on four checks.\n\n| Our check | It asks | In the outside sources |\n|---|---|---|\n| Recurrence | does the pattern appear at two or more independent levels | the number and variety of cases (Keynes; Amgoud's Number and Diversity) |\n| Mechanism identity | the same mechanism, or only the same shape | the shared feature bears on what is carried over (Hesse, Bartha, Davies and Russell); structure over looks (Gentner) |\n| Adversary | does the new level bring an opponent the old one lacked | a difference between the cases that breaks the step |\n| Substrate condition | what the earlier case needed in order to work, and whether we have it | the same, for conditions |\n\n**What the sources add to the design.**\n\n- **A likeness names what it carries over.** The same earlier case supports opposite conclusions\n  depending on what is mapped onto what. The contract asks for the shared mechanism and does not\n  yet ask what is being concluded from it.\n- **A likeness earns its keep in what it predicts.** Hesse's term for the features of the earlier\n  case not yet known to hold in the new one is the neutral analogy (second-hand). For the graph:\n  an objection or a distinction found in one debate is a candidate question in its kin.\n- **Finding and judging are two stages.** People find by looks and judge by structure (§ 4). A\n  finder that runs on the likeness of wording will bring back look-alikes. The second stage lines\n  the two cases up.\n- **No figure.** An analogy gives plausibility (Bartha).\n\n**Five limits.**\n\n1. **The second half is contested.** Where a comparison is disputed, the daemon should write out\n   the rival likenesses and choose none.\n2. **Its reminding will favour the well-trodden.** § 4's findings on machines are mixed, and the\n   mixed part is exactly this. The channel in which a person supplies the comparison is the answer\n   he endorsed on 2026-09-02 ([the-analogy-daemon.md](the-analogy-daemon.md) § 7).\n3. **What counts as alike is partly made in the comparing.** It depends on how the two cases were\n   written down (Chalmers, French and Hofstadter 1992). The daemon compares claims as an earlier\n   pass wrote them. It cannot see a feature nobody wrote, and it cannot step outside the frame the\n   claims were written in. These are the two classes of error the canon keeps for people.\n4. **Kinship has no answer key**, so the daemon's error rate will not be proven soon, and its\n   unexamined error is what has to stay low (ledger T2).\n5. **A machine's grade of a link moved with the wording** on 2026-10-07. A likeness is written out\n   and checked as writing. It is not graded.\n\n**An example from the live graph.** Anil Seth's claim, as mapped: *the information processing\ninside the model is no more likely to result in genuine consciousness than a computational\nsimulation of a weather system is likely to generate an actual physical hurricane*. The link from\n*weather simulation cannot produce real hurricane* is stored as an analogy. What the two cases\nshare can be stated: each is a computation that stands for something. Whether that bears on\nconsciousness depends on a further claim. If being conscious is like being wet, a property of the\nstuff, the likeness holds. If it is like adding, a property of the pattern, it fails, since a\nsimulated sum is a sum. Which of the two it is like is the dispute itself. (The fork is put in\nClaude's words.) A daemon could write both likenesses out with their cases. It could not choose.\n\n**The long-run answer is the record.** Goodman's reversal says standards of likeness come from\ninductive practice. A graph that keeps every likeness that was used, what it carried over, and how\nit fared under attack is such a practice written down. Kemp and his co-authors show the same thing\nin a model: what makes one case enough is learned from how earlier cases went. The philosophy\nnote's caution applies in full: a record is conservative, and it favours the comparisons already\nin use. The way for a person to say *this whole comparison is the wrong one* has to stay open.\n\n## 9. Proposals\n\nAll unruled. They are filed as refinements inside ledger rows AD23, AD25 and S6, with no new rows.\n\n1. **Four shapes** (AD23). A written-out step records one of four: from a rule and a case; from a\n   sign; from a case to a like case; from cases to a rule.\n2. **A link from an earlier case names three things**: what is carried over, the shared feature,\n   and why that feature bears on it. It is shown with one case that shares the feature and where\n   the conclusion fails.\n3. **A likeness is argued as the link's relevance claim.** No change to the arithmetic.\n4. **Two standing jobs for the daemon**: for a written-out rule, fetch the cases behind it; for a\n   disputed comparison, write out the rivals and choose none.\n5. **The daemon gives no figure.** It names cases. Code counts how the conclusion fared among\n   them. The choice of cases is shown and can be argued, since that choice is the judgment.\n6. **Questions travel between kin.** A question raised in one debate is offered in a debate that\n   shares its mechanism, as a move beside a gap.\n7. **Three tests before any build**, each with its expectation written first:\n   - the write-out pass names the shape of each of the thirty links, and the prediction in the\n     dossier is then scored;\n   - the likeness question is put two ways on the 37 links from a case to a like case, and the\n     links that change side are counted;\n   - for ten rules, the earlier cases found by the wording finder, by the structural finder and by\n     a person are compared.\n\n### 9b. The eight proposals of 2026-10-07, after this reading\n\nSpoken in the thirteenth listening piece\n([2026-10-08-eight-proposals-pondered-again.md](../listening/2026-10-08-eight-proposals-pondered-again.md))\nand filed in the ledger rows named. All unruled.\n\n| Proposal, by its picture | What this reading adds |\n|---|---|\n| the toll bridge (AD22) | The right to cross was earned by earlier crossings: a rule has its cases behind it. The written-out sentences are parts of the link's relevance claim. Today a fully sound step earns its author almost nothing and one refuted part nearly removes the link, so whether sound parts should lift a link is his to rule ([the-squirrel-and-the-strength-calculus.md](the-squirrel-and-the-strength-calculus.md) § 5) |\n| four ways to walk (AD23; it was *three ways*) | A fourth shape, from cases to a rule, and the shape decides the weighing |\n| a place for the stranger (AD24) | What the stranger lacks is often a case and not a rule. And two readers can weigh likenesses differently, the weighing partly following the conclusion each wants (§ 13.2) |\n| no vehicles in the park (AD25) | The questions about what would break a step are how a rule gets drawn from cases, so the answered ones are kept under the rule as its cases |\n| why the house burned (AD26) | Inside one sufficient set the weakest part governs (a chain). Across sets the ways add (a cable; Peirce 1868). The code's third setting takes the strongest. How the strands combine is ledger Q22 |\n| the madman's weapon (AD27) | The daemon's design already demands the strongest case against. A case against lands either on the link's relevance or on the conclusion. Rival comparisons are shown side by side |\n| whose sentence is it (AD28) | A fourth answer to *by what right*: a case (*this reminds me of*), through the channel endorsed 2026-09-02. And a written rule may be a reason found afterwards |\n| agreeing without agreeing why (AD28) | ⚠ Narrowed the same evening (§ 13.1): no source measured an order, and large publics show the reverse. What stands: a body that must decide can agree on a case for different reasons, and that agreement is recorded as such |\n| a map the size of the country (AD29) | The writing-out has a second direction, into the cases behind a rule. A chain of whys can come to rest on a shared case |\n| the friend who says *this reminds me of* (S6) | § 8 |\n\n## 10. The view from here\n\n**Real today.** The question *alike in what matters* hangs on 121 links, in three wordings, and an\nanswer to it moves a link's relevance. A likeness can be argued as a claim, and nobody has argued\none yet: the graph holds no relevance claim. One finder of structural likeness exists and proposes\npairs. Nothing answers the questions, nothing writes a\nlikeness out, and links drawn between texts by the linking step carry no kind at all.\n\n**What it would open.** A rule under a link would come with the cases it was drawn from. A\ncomparison would say what it carries over, and its rivals would stand beside it. Questions would\ntravel between debates that are kin. And the graph would hold, in the open, a growing record of\nwhich likenesses held.\n\n## 11. Corrected with this note\n\n| Where | What it said | What holds |\n|---|---|---|\n| the first answer in the conversation, given from memory | Hume, Mill and Hofstadter support the thought | Hume and Mill do. Hofstadter's essay makes categories bundles of analogies and does not treat induction |\n| [philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md) § 5; ledger AD23 | a step has one of three shapes | four: a step from cases to a rule was left out, and the graph holds 84 |\n| [the-analogy-daemon.md](the-analogy-daemon.md) § 3 | the four checks were *already ruled* as an enrichment of the analogy pattern | their home calls the mapping unruled, *\"load-bearing-but-unratified\"*. Corrected there in a commit of its own. The same word stands in the project's always-loaded file and is listed in `TODO.md` |\n| an earlier answer in the conversation | *about one link in seven* rests on a case, with no word on which links | it holds for links between claims from texts and people (121 of 780). Over every labelled link the share is the same, and most of those links belong to standing questions |\n| the brief given to the philosophy helper | Peirce's 1878 paper treats analogy as a compound; Keynes names a *Principle of Limited Independent Variety*; Levi 1949; Waller's view covers all analogies | the compound is in Peirce 1868; Keynes writes *limitation of independent variety*; Levi's book reprints a 1948 article; Waller's view covers a priori analogies. None of the four had reached a document |\n\n**Reported by the corpus read, and not checked by Claude**: a docstring in `schemes.py` gives\ncounts of patterns and clusters that differ from the code; the note on dogfood run 6 calls\nstructural matching unbuilt, and the finder shipped on 2026-08-20; ledger row AD13 still says the\nsquirrel has the three-part form; a line in the always-loaded file says Hofstadter was never cited\nbefore; and the proposal of 2026-09-02 to add him to the list of academic foundations still waits.\nThese are listed in `TODO.md`.\n\n## 12. Read and not checked\n\n- **Claude opened none of the outside sources.** Every outside claim above is a helper's report.\n- **Not read first-hand by any helper**: Hesse 1966; Bartha's book; Wisdom 1991; Carnap; Watanabe;\n  Wittgenstein's text; the Nyāya-sūtra and Dignāga; Hofstadter and Sander's book; Marcus's book;\n  Rips 1975; Osherson and others 1990; Gelman and Markman 1986; Alexander and Sherwin; Posner.\n- **Abstract or excerpt only**: most of the 2022 to 2026 papers on language models; Nosofsky 1986;\n  the rule-and-exception models; Smith, Langston and Nisbett 1992; Sunstein 1995; several papers\n  in the dispute over whether an analogy hides a rule.\n- **Hofstadter's 2001 essay** was fetched whole, read in its main sections, and searched by word\n  for the rest.\n- **Works in copyright are paraphrased** in the philosophy report, so exact wording is at its\n  links. Quotations above from such works are a sentence or less.\n- **The labels on the graph are a classifier's.** The direction check in § 2 is eight links, read\n  by their titles.\n- **What is Claude's own**: the split between origin and ground, the two places of likeness in a\n  written-out link, the fourth shape, the reading of the strength code in § 7, the link to the\n  concept layer, and the fork in the hurricane example.\n\n## 13. Two readings, checked the same evening\n\nTwo of this note's readings went into overview documents that afternoon: that people meet at a case\nbefore a principle (`docs/convergence.md`), and that a worldview lens is a weighting of likenesses\n(`docs/worldview-lenses.md`). The founder kept them and asked for a check: *\"Keep them, I like them.\nThink a bit more on them and research whether they're true more deeply\"*. That keeps them in the\ndocuments. It does not rule on them.\n\nTwo read-only helpers looked for evidence for and against, one reading each. Their reports are in\nthe dossier unchanged\n([cases and principles](what-a-rule-rests-on/cases-and-principles-sources-as-read-2026-10-08.md),\n[likeness and worldview](what-a-rule-rests-on/likeness-and-worldview-sources-as-read-2026-10-08.md)).\nClaude opened two abstracts itself (Awad et al. 2020; Braman and Nelson 2007). Everything else below\nis as the helpers read it, at the depth stated in each row.\n\n**Both readings hold in part, and both were worded too strongly.**\n\n### 13.1 Cases and principles\n\n*As first written: people who disagree often agree on a case before any principle, so agreement on\ncases comes first.*\n\n| What was found | Source, and how far it was read |\n|---|---|\n| Sunstein names three kinds of agreement, and the first is the reverse of the reading: agreement on a general principle with sharp disagreement about cases. His own subject is the third: agreement on an outcome and a narrow reason, with the theory left open. | Sunstein 1995, through a page that quotes the passage; the article itself was blocked |\n| The commission's mandate from Congress was to identify principles, and its members read moral philosophy first. *\"The Commissioners believed, as principalists, they worked as casuists.\"* | Jonsen 1998, Toulmin's co-author and a fellow member; full transcript |\n| Each abstract democratic principle was endorsed by 94.7 to 98.0 percent of voters in two cities. None of ten applied statements reached 90 percent. | Prothro and Grigg 1960, full text |\n| White Americans' support for racial equality as a principle runs ahead of their support for the programmes that would carry it out. | a recent paper on survey data from 1994 to 2018, abstract and introduction |\n| A citizens' jury moved from no option being the first choice of more than four of twelve to seven of twelve for one option. The authors report a high degree of agreement, at the abstract level, on which basic values are legitimate. | Dryzek and Niemeyer 2006, selected passages |\n| Three dilemmas were ranked in the same order of acceptability in every one of 42 countries. The level of acceptance varied a great deal by country. | Awad et al. 2020; abstract, read by Claude |\n| Asked in the abstract, 86 percent said nobody can be responsible in a determined universe. Given one concrete wrongdoer in it, 72 percent held him fully responsible. | Nichols and Knobe 2007, from a page that reports the figures |\n| Reasoning from cases needs a norm that connects them, and it meets conflicting deep readings of the same case. | Beauchamp 2003, method section; Arras 1991, abstract |\n\n**What does not hold.** That agreement on cases comes *first*: no source measured an order, and\nSunstein writes of levels of generality, not of a sequence. That the pattern is general: in large\npublics the measured pattern is the reverse. And Toulmin's commission is weaker evidence than it\nlooked, since its mandate and its reading were about principles.\n\n**What holds.** A body that must decide, shares a tradition and can settle for narrow reasons can\nagree on a case without agreeing on a theory. That is a real and useful kind of agreement, and it\ndepends on norms the members already share. Beauchamp's point is this note's § 5 arriving from\nbioethics: a likeness between cases carries no conclusion without a general claim about what bears\non what.\n\n**The reading, as it stands now: agreement on a principle's words comes more easily than agreement\nat a case, and it is tested when the principle meets a case.** (First written as *is cheap*, which\n§ 14.3 finds too flat.) People part there in two ways: on whether the case is covered, which\nis the question of a rule's breadth (ledger AD25), and on what else weighs in the case, which is the\nweighing scheme's question.\n\nFor the convergence wager (Claude's reading, unruled):\n\n- **Agreement on an abstraction is not convergence until it has held at a case.** The corpus has its\n  own instance. Mapping two opposed legal texts on Israel and Palestine, the authors agreed on nearly\n  every value in play and disagreed on which legal framework governs\n  ([dogfood-run-6-israel-palestine-cross-domain.md](dogfood-run-6-israel-palestine-cross-domain.md)).\n  The abstractions in such surveys are also stated in words people want (*democracy*, *free speech*),\n  and the ruling of 2026-10-07 on wanted words applies: each reader is sure they understood, and\n  understood their own sense ([prime-real-estate-words.md](prime-real-estate-words.md)). That link\n  is Claude's.\n- **Agreement on a case inside one tradition says little across traditions.**\n- **Orderings may converge where verdicts do not.** Awad's finding is agreement across 42 countries\n  on which case is worse than which, with disagreement on where the line falls. It has the shape the\n  corpus calls shared structure and divergent weighting, measured at the level of cases. A question\n  of the form *which of these two is worse* is untested here.\n\n### 13.2 Likeness and the lens\n\n*As first written: a lens is a weighting of likenesses, which is why a comparison feels apt to one\nside and forced to the other.*\n\n| What was found | Source, and how far it was read |\n|---|---|\n| How alike two things are depends on which features are weighted, and *\"The relative weights assigned to these features, of course, may differ for different people.\"* | Tversky 1977, full text |\n| Weighting is one of five ways in which worlds differ. Two worlds can hold *\"just the same classes sorted differently into relevant and irrelevant kinds\"*. | Goodman 1978, full text; Goodman 1972 from a repost |\n| Knowledge decides which features there are, and not only their weights. Likeness is *\"at least as much a consequence as a cause\"* of how a person's concepts hang together. | Medin, Goldstone and Gentner 1993, read in part; the quoted words are Murphy and Medin's, cited there |\n| People judged a precedent more like the case at hand when its outcome suited their policy view. The effect was strongest when the cases were moderately alike, and legal training did not reduce it. | Braman and Nelson 2007, two experiments; abstract, read by Claude |\n| Ratings of arguments move toward the verdict a person is forming. | Simon and colleagues 2001, abstract and introduction |\n| Experts sort physics problems by principle, beginners by surface. | Chi, Feltovich and Glaser 1981, as reported in another paper |\n| The difference between American and Chinese sorting replicates. The authors offer a reading on which the notion of likeness is shared and the statistics of the two languages differ. | Le and colleagues 2023, abstract and introduction |\n| Opposed groups watching the same film disagreed on what happened in it. The mechanisms named concern which facts are noticed and who is believed. | Kahan and colleagues 2012, abstract and theory section |\n| Three well-known supports are weaker than their fame: one named effect of context on feature weights (small and not significant on a second look), the effect of a metaphor on policy views (absent in four experiments), and the five-part structure of the moral foundations questionnaire (three clusters found). | a later study of Tversky's principle, from a summary; Steen and colleagues 2014, full text; Harper and Rhodes 2021, abstract |\n\n**What holds.** Likeness depends on weighted respects, and people differ in the weights: by what\nthey know, and by what they want to conclude.\n\n**What does not.** A lens is more than a weighting. Goodman has four other ways. Knowledge changes\nwhich features exist. Kuhn's change of paradigm changes meanings and groupings, as the Stanford\nEncyclopedia summarises him. And much of what divides people is which facts they credit and whom\nthey trust. `docs/worldview-lenses.md` already lists four things a lens changes, so the weighting of\nlikenesses is at most a fifth beside them. And *that is why a comparison feels apt* had the arrow\npointing one way only: the weighing is partly produced by the conclusion.\n\n**The reading, as it stands now: a lens weighs which likenesses count, and the weighing partly\nfollows the conclusion wanted.** A likeness that feels apt is therefore no independent evidence for\nthe conclusion it serves.\n\nFor the design (Claude's reading, unruled):\n\n- **The likeness question is put before the conclusion is shown.** § 8 split the question in two:\n  what the cases share, and whether that bears on the conclusion. The first half can be asked with\n  the conclusion hidden, of a person and of a machine. Braman and Nelson's experiment is the template.\n- **The risk sits with comparisons that are moderately alike**, which are the contested ones.\n- **Training did not protect.** The corpus holds the same finding from another literature, as its\n  reason for having no expert gate.\n- **It is a measured case of a mind unable to audit its own frame**: the same person's verdict moved\n  the likeness they saw.\n- **The lens page's own line has an ancestor.** *The premises are shared; the weighting is where you\n  diverge* is close to Goodman's sentence above. His book had not been cited in the corpus.\n\n### 13.3 One finding, seen twice (Claude's reading)\n\nPeople agree on a principle's words and part when it meets a case. Whether a case falls under a\nprinciple is a judgment of likeness to the principle's clear cases. That judgment is weighted by\nwhere a person stands, and partly bent by what they want to conclude. So the two readings describe\none place: **the sentence that ties a case to a rule**, which is the second part of a written-out\nstep (the first claim; what ties this case to a rule; the rule).\n\n- **A prediction, written before any count.** When people challenge written-out steps, the challenges\n  will land on the tie more often than on the rule. Nothing on the graph can score this yet, since no\n  step is stored written out.\n- **It fits the spiral of § 5.** Tversky has grouping come before likeness, and Murphy and Medin have\n  likeness follow from the concepts a person holds. So likeness is no floor under a rule: it rests in\n  part on groupings and theories already in place.\n- **It gives the corpus's picture of disagreement living in the middle a measured form**: agreement\n  on a principle's words, often agreement on clear cases, and the parting between the two.\n\n### 13.4 Corrected with this section\n\n- `docs/convergence.md`: the section is retitled and rewritten, and its first title is kept in a\n  note under the new one.\n- `docs/worldview-lenses.md`: the paragraph is rewritten.\n- `docs/depth.md`: *often comes to rest on a case* is now *can come to rest*. Nothing measured how often.\n- § 6 and § 9b of this note, and ledger rows AD24, AD28 and S6.\n- [philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md), its short version and\n  § 11, where Toulmin's commission stood as evidence of where agreement sits.\n- The thirteenth listening piece says in its eighth part that people meet at the case and write the\n  principle up afterwards. The audio was delivered before this check. Its kept text carries a note.\n\n### 13.5 Read and not checked\n\nNot opened by anyone: Sunstein's article itself (blocked; his three kinds come from a page that\nquotes them), Jonsen and Toulmin's book, McClosky 1964, Schuman's book, Nichols and Knobe's paper\n(so whether the same people answered both questions is unknown), Kuhn's own texts, Khong, Lakoff,\nJi and colleagues 2004, Graham and colleagues 2009. Levi and Quine stand as the afternoon's helper\nreported them. No study was found that measures agreement at both levels in the same people and\nalso which came first. The helpers' own lists of what did not check out are in their reports.\n\n### 13.6 The founder's comment while listening, and where our own practice already does this\n\nListening to part eight that evening, the founder wrote: *\"Agreeing on rules comes late, yes. By\nreviewing cases you reflect on them and see/distill regularities/learnings. This was also mentioned\nin the evals walkthrough iirc.\"*\n\nHis sentence and § 13.1 are about two different things, and both stand.\n\n- **How a rule forms.** A person or a team reviews cases, compares them, and the regularity becomes\n  visible. That is what § 4's experiments measured (comparing two cases made the principle usable\n  for about half of the learners, against one in five who read the same cases one after the other),\n  and it is the half of the first reading that holds.\n- **Where people who disagree find agreement.** § 13.1 withdrew the claim that they find it at cases\n  before principles, in general. Nothing in his sentence needs that claim.\n\nHe recalled the evals walkthrough correctly. Its first step is one person reading real output and\nwriting a note on each case. Its second is grouping the notes into kinds of fault, and counting.\nAnd the reader stops when new cases stop teaching new kinds\n([the listening text](../listening/2026-10-05-the-evals-loop-and-our-machine-passes.md);\n[the-evals-loop-and-our-machine-passes.md](the-evals-loop-and-our-machine-passes.md) § 1). That is a\nrule drawn from compared cases, in this project's own practice, and it has been a blessed principle\nsince 2026-10-06: a person reads what a machine pass produced before anyone proves how often it is\nwrong. The stopping rule is Keynes's point, read for this note, seen from the other side: a further\ncase of a kind already seen adds little.\n\nNeither the walkthrough nor this note had made that connection. The founder made it.\n\n**How much of the older corpus these notes name at all.** A first count is in the dossier\n(`what-a-rule-rests-on/older_ideas_mentioned.py`, and what it printed beside it). Of 30 older ideas\nthat Claude thought might bear, each looked for by a word or two, 23 are named in none of the four\nnotes or the walkthrough. A zero there says an idea is not named. It does not say the idea is\nmissing under other words, and it says nothing yet about what each would change. That reading was\nunder way when this was written, and it is owed (`TODO.md`).\n\n## 14. The walkthrough read against the older corpus (2026-10-08, late)\n\nThe founder asked that night: *\"Is there any part of the corpus that bears on any aspect or even any\nsingle sentence of the latest audio walkthrough that you haven't fully grasped or related it to?\"*\n\n**Yes, a great deal.** The four notes and the walkthrough were written reading outward (philosophy,\npsychology) and downward (the code, the live graph), and almost never sideways into this corpus: of\n30 older ideas looked for by name, 23 stand in none of them (§ 13.6). Three read-only readers then\ntook one region of the older corpus each. Their reports are in the dossier unchanged:\n[strength and machine work](what-a-rule-rests-on/older-corpus-bearings-strength-and-machine-work-2026-10-08.md),\n[people and agreement](what-a-rule-rests-on/older-corpus-bearings-people-and-agreement-2026-10-08.md),\n[likeness, shapes and resting places](what-a-rule-rests-on/older-corpus-bearings-likeness-shapes-and-bottoms-2026-10-08.md).\nWhat follows is Claude's reading of them. Nothing in it is ruled, and § 14.7 says which passages\nClaude opened itself.\n\n### 14.1 Seven statements in the walkthrough that the code or the record contradicts\n\n| The walkthrough says | What stands |\n|---|---|\n| *Showing your work can only cost you* (part one, and the first thing to carry) | True of a link a text states or a person wrote, which is presumed half relevant. A link a machine drew between other people's claims presumes no relevance, so it counts for nothing today, and a written-out step that somebody examines is the only way it comes to count. The proposals of 7 October are about those links. The squirrel note kept the scope (*as an extracted link*); the walkthrough dropped it. Parts a machine wrote also weigh nothing until a person or a blessed pass stands behind them, so *one rotten plank removes the bridge* describes a step a person examined and tagged. Read in `deliberus/relevance.py` and `deliberus/support_semantics.py`, not run. |\n| *The three friends with the table… the code has a setting for that. It takes the strongest* (part five) | That setting is for parts of a claim where one is enough. *Any two of three* is no setting, and separate reasons for a claim simply add. |\n| *Our checker reads one way of the four* (part two) | True of the blind reader of the trial. The standing questions hang on every stored kind. |\n| *We wrote that rule twice, two months apart* (part six) | 38 days: 30 August to 7 October. |\n| *Something nobody has had: a public record of which comparisons held* (part ten) | Case law is one. The philosophy note says so of the law report. |\n| *He never argues that rules come from cases* (Hofstadter, part ten) | True of the 2001 essay a helper read. His 2013 book with Sander is cited in the corpus and unread here. |\n| *People meet at the case. The principle is what they write up afterwards* (part eight) | Withdrawn in § 13.1. *Firmer than agreement on any rule*, in the same part, has no support either. |\n\n### 14.2 Nine places where a proposal meets something already ruled or built\n\n1. **Who gives a link its second number.** On 2026-09-28 the founder ruled that a machine of a second\n   model family proposes a link's relevance, people argue it, and nobody votes it (ledger Q33; his\n   direction of 2026-09-19 was the same). That judge is built and waits on his word to write. On\n   2026-10-07 ledger row AD7 proposed that no machine reading of a link *may write, hold, retire,\n   weigh or rank* it, and AD17 that a machine's sentence *weighs nothing*. Neither row, nor the note\n   they come from, names the ruling. They cannot both stand as worded. The October measurements\n   (one question refused 4 of thirty links and another 20) are a reason to put the judge's question a\n   second way before it is allowed to write. Which governs is his to rule. Part ten says the second\n   number *is where people disagree* and does not say that a judge is already ruled to propose it.\n2. **A named kind of stop.** AD29 has each chain *end in a named kind of stop*, and part nine lists\n   five. The canon line ruled on 2026-09-15 is that a kind is read, never set, and nothing is\n   permanently terminal; the founder's own position is that asking why all the way down is the\n   untyped move. A resting place is then a dated reading, mapped to the horizon kinds that exist, and\n   conservative. The spoken list also dropped the one the corpus keeps, a practice where reasons run\n   out, and added *a case we all remember*, which is the hazard of a premise presumed shared\n   ([the-load-bearing-unsaid.md](the-load-bearing-unsaid.md)).\n3. **The map tells no stories.** *Deliberus does not tell stories, and must not*\n   ([storytelling-and-the-story-layer.md](storytelling-and-the-story-layer.md) § 4), since a machine\n   that restates people shifts what they decide. So *hand the stranger the story* means: quote and\n   locate a case that a person or a source gave. The measured effect of narrative was between people,\n   face to face.\n4. **The stranger is a person.** By the ruling on the stranger test, a model reading as a stranger is\n   an advisory pass and an actual stranger is the authority. AD24's first reader is a model, and it\n   uses a word that is taken.\n5. **Replacing a rule.** Contributions are added, never overwritten. A rule swapped under challenge\n   stays beside its replacement, owned and dated, or the swap is the retreat\n   [the-motte-and-the-periphery.md](the-motte-and-the-periphery.md) describes (as a reader reported\n   it).\n6. **A third standing.** The blessed end-state of 2026-08-31 is that the graph holds contributions\n   and telemetry and nothing else. *Committed by the step, not asserted* fits that as a displayed\n   reading and collides with it as a stored standing. An author's *own*, *replace* and *withdraw* are\n   contributions only they can make, which is all consent was.\n7. **Asking the author.** A person shown a machine's proposal has taken its frame before answering\n   ([what-human-judgment-is-for.md](what-human-judgment-is-for.md)). So the author is asked for their\n   own reason before being shown the machine's rule. Where the step is a value step, when to ask is\n   itself a matter of consent and timing (the ruling of 2026-08-25 on the sacred), and the four notes\n   never mention it.\n8. **One hand all the way down.** The spoken design is a machine-drawn link, a machine-written rule,\n   a machine-found case against it and a machine check that the step is complete, all of one model\n   family. Three rules of the corpus bear: a pass writes what it found and holds what it added (13 of\n   15 written rules stood in neither text, so they were added); machine work built on machine work is\n   capped by a measured rate; and a daemon acts when two detectors of different mechanism or family\n   agree. So a case against counts when it was found in a text, as Van den Haag's was, or by another\n   family; and AD27's *met none* proves little when the same hand looked.\n9. **Where a written step lives.** The proposal of two days earlier, also unruled, is that a use is\n   kept, a machine judgment is recorded once, and everything else is read\n   ([store-the-uses-read-the-meaning.md](store-the-uses-read-the-meaning.md)). Under it a written-out\n   step is a recorded judgment with a pointer on the link, and *on request* asks the record first.\n   AD2 and AD6 store it as text. The two proposals have to be reconciled before either is ruled.\n\n### 14.3 What the corpus already had\n\n- **The four ways.** [scheme-set-exhaustiveness.md](scheme-set-exhaustiveness.md) holds a two-layer\n  design: a small closed layer that includes *reasoning direction (deductive/abductive/inductive)*,\n  and an open layer of named kinds. Three of the four ways are that direction, and likeness is a\n  relation there and no direction, so the list mixes two axes. The four cover 443 of the 780 links\n  that carry a kind; expert opinion, consequences and practical reasoning fit none. The proposal\n  gives no way to say *none of these*, which the corpus requires of a closed list. This afternoon's\n  own corpus report carried the passage, and it was dropped. Going from three to four within a day\n  is the corpus's law that a taxonomy breaks on the next register.\n- **From cases to a rule** had a designed home: a *generalises* inference with its own questions\n  about whether the case carries over ([what-counts-as-evidence.md](what-counts-as-evidence.md);\n  ledger Q4, Q5, Q11, Q22). *Is this case typical of its kind* ships as a standing question. And a\n  rule with its cases is representable today: the 84 links stored as *example* already file a case\n  under a general claim.\n- **The variation questions.** Exceptions to a rule, the edge of a classification and relevant\n  differences ship as standing questions, and weighings have a family of *would it shift if*\n  (as a reader reported the code). New in AD25: asking term by term, and keeping the answers as\n  cases. Deutsch's test, in [depth.md](../depth.md), adds what AD25 lacks: a good rule is hard to\n  vary *while still doing its job*, and breadth alone rewards a rule that says nothing.\n- **Conjecture and test.** The corpus gives its lineage as Popper's and Deutsch's: rules are guessed\n  and cases test them. None of the six notes of this week names either. It fits § 5's split better\n  than it fits the walkthrough: comparing cases is what prompts the guess (§ 4), and it justifies\n  nothing by itself.\n- **The chain and the cable.** Whether separate reasons add is the contested choice, and Pollock,\n  cited in [weakest-link-arithmetic-and-the-merge-hunch.md](weakest-link-arithmetic-and-the-merge-hunch.md)\n  for the weakest-link principle, rejects adding outright. A written-out step is the marker that\n  Yu's test for linked against convergent support lacks. Fair-share routing for a strand counted\n  twice is built and unadopted. And two sufficient sets under one link can also be the sign of an\n  argument that changes when it is cut again, which the threat model watches for.\n- **Held for different reasons.** The ruling on the discursive dilemma already says to show both\n  readings and resolve neither. Nothing records who holds a link, since votes attach to claims, so\n  the state needs a holder first. Rawls's overlapping consensus is the precedent. And the corpus's\n  own proof that agreement without shared reasons can hold is the Universal Declaration, agreed on\n  *\"provided no one asks us why\"*\n  ([moral-convergence-metaethics-and-psychology.md](moral-convergence-metaethics-and-psychology.md)).\n  That makes § 13.1's first wording, *cheap*, too flat. What survives is narrower: **say at which\n  level an agreement was found, on words, on a case, or on reasons.**\n- **What goes without saying.** People overestimate how far their sense of a word is shared\n  ([core-periphery-and-the-concept-layer.md](core-periphery-and-the-concept-layer.md)). That is a\n  measured cause for what the stranger lacks, and a warning about *a case we all remember*.\n- **The lens.** [claim-sameness-philosophical-readings.md](claim-sameness-philosophical-readings.md)\n  put likeness relative to a respect, and a lens as a set of equivalences, as a hypothesis in April.\n  It also warned that the hypothesis *has the shape of an elegant reconciliation*, which is the\n  shape of the walkthrough's opening line, that most of the proposals are about one thing.\n- **What it costs to show a premise.** [incentives-analysis.md](incentives-analysis.md) names the\n  cost of exposure, and expects supports to be offered freely and attacks rarely. So planks will be\n  over-supplied and cases against under-supplied. *On request* also dropped *with a reason*, which\n  the philosophy note kept, and a bare demand is a way to stall.\n- **The evals loop**, which the founder found (§ 13.6).\n\n### 14.4 The three things the walkthrough asked him to carry\n\n1. **Should sound planks lift a bridge?** It is the old open question of whether established parts\n   lift a whole, with a lean on record (`TODO.md`, the up-transmission whisper). The variable chosen\n   there is how much of what had to be shown was shown. Whether the written parts are sound is a\n   different question, and a reader of the writer's own family cannot be its judge. And the cost in\n   the question holds for some links only (§ 14.1).\n2. **Does the fourth way join the other three?** The better question is whether a shape field is\n   needed beside the stored kind at all, or whether direction is read from the kind, with a way to\n   say *none of these* (§ 14.3).\n3. **What is at the bottom?** The three answers offered are levels the corpus already tells apart.\n   *A way of being moved* is what [decomposing-value.md](decomposing-value.md) has as the felt floor,\n   with a gap between knowing and being moved, and [depth.md](../depth.md) holds that the weighting\n   of what is felt *\"may be more universal across humans than any political conclusion\"*. *A sense\n   of what is like what* is measured in\n   [why-representations-converge-and-whether-minds-do.md](why-representations-converge-and-whether-minds-do.md)\n   (as a reader reported it): shared in perception and at the core of concepts, divergent at the\n   periphery, which is where judgments of likeness live. That fits § 13: clear cases and orderings\n   are shared, hard cases split. A sense shared by descent is also no evidence that it tracks\n   anything. And a machine's sense of likeness is the average of many people's, which can be a\n   comparison nobody makes, and a second model family shares it. Part ten's *will favour the\n   comparisons everyone already makes* understates that.\n\n### 14.5 The founder's second comment that night: not alone\n\nClaude had paraphrased the wager as people who *dig deep enough* finding shared ground. He wrote:\n*\"Not just 'dig deep enough' but also engage with each other deeply, ofc. If I just sit in my room\nand 'dig deep enough' I might find new insights through self-reflection but what's really fruitful is\nengaging with people of vastly different perspectives, because it's cross-pollinating,\ncross-referencing, cross-illuminating, you know?\"*\n\nThat is the wager as he holds it, and the paraphrase had lost it: the descent is made with people who\nstand elsewhere. It is the conviction that what is invisible from inside a position is visible from\noutside it. It meets two findings above. Agreement on a case inside one frame says little\n([the-group-premise-evidence.md](the-group-premise-evidence.md): readers from one community are one\nframe sampled many times), so an agreement is recorded with *between whom*, and how differently\nplaced. And the same note asks for independence before pooling: each person's sorting is taken\nbefore they see the other's, and the engaging comes after. Claude's proposal, unruled: an agreement\non the map carries its level (§ 14.3) and its parties.\n\n### 14.6 Why this was missed, and one change of method\n\nThe instruction file's index of the corpus was in view the whole time, and a corpus report that\nafternoon had carried one of the passages. The ideas were filed, correct and not in use: the failure\nthe corpus calls the uncollected idea. A walkthrough is written under a frame, here *it is all about\ncases*, and a frame does not go looking for what would complicate it. The change: a piece that walks\nproposals is read against the older corpus before it is rendered\n([how-to-write-for-listening.md](../listening/how-to-write-for-listening.md)).\n\n**Found on 2026-10-09, from the session's own record.** The note that holds the ruling on who weighs a link's\nrelevance ([relevance-and-the-two-factors.md](relevance-and-the-two-factors.md)) is cited by all four notes that\ncarry the proposals of 7 October, and the session had opened it 49 minutes before it first wrote the proposal\nthat meets it. So the worst of the nine was a failure to apply what had been read, from inside the draft's own\nframe: the session had just seen four ways of having a machine judge a link fail. A rule that the right notes\nbe opened before proposing would not have caught it. Three readers outside the draft did. This is the\nprinciple *never ask a mind to audit its own frame*, met in our own work\n([what-agent-memory-research-says-about-agent-surfaces.md](what-agent-memory-research-says-about-agent-surfaces.md)).\n\n**Checked further the same day.** That opening read the whole note. The note carrying the proposal cites § 5\nof it in the sentence before it proposes ([what-a-link-rests-on.md](what-a-link-rests-on.md) § 4.7, as\ncommitted that afternoon) and calls the machinery *already ruled*; in that passage it does not say that the\nruling has a machine propose the weight. So the note was in use and not forgotten. What was missing was the\nact of setting the proposal beside the ruling and saying which of the two governs. The outside check was\nthen tried once by hand on this case: a reader with no knowledge of the project, given the proposal\nunlabelled among four and the 84 ledger rows settled before it, each in full, named the conflict with Q33\nand quoted the words. The same reader given only each row's headline did not, and asked to see Q33 in full.\nNeither raised an alarm on the two proposals the founder blessed that day. One case, and readers of the same\nfamily of models as the session.\n\n### 14.7 Read and not checked\n\nOpened by Claude: the relevance and support-weight code, the squirrel note's scope line, ledger rows\nAD7, AD17 and AD29, the to-do entries on the whisper, the scheme note's two-layer passage, the\n*generalises* passage, the Maritain line, the storytelling boundary, the founder's position on asking\nwhy, the relevance judge's header, the passage in depth.md, the Pollock passage, the daemon's prior\npass, and April's warning. Everything else in this section is as the three readers reported it, with\nthe file and line in their reports: the bridging types, the four tests in the motte note, the\nratification record, the depth rule, the derivable requirement space, latent scaffolding, the red\nteam's attacks, the reader tiers, the representations notes, the Lean notes, and the sameness\ncheck's answers. The correction to *can only cost you* was read in the code and not run.\n"}