{"path":"research/what-a-comparison-produces.md","content":"# What a comparison produces, and where each product lives in the data model\n\n*2026-11-09. An all-angles pass on ledger row BL3 (RULED the same evening in the wording of § 5 item 1, with BL4: *\"Recommendations blessed. Build whatever is buildable.\"*) (the four products have homes), run on his word after Astra's review\nof the row and his own two misgivings. His words, verbatim: \"I feel like most of BL3 is still a bit up in the air for\nme, I don't feel a 'hell yeah' for everything yet.\"; \"I feel like maybe there's not enough connection to theory here\nfor me or that this is not clearly anchored in theory, why the 'four products' are what they are and how they should\nbe thought of best in terms of/translated to the core data model. Reflections?\"; \"Maybe I just need a super vivid\nexample or two to illustrate this better.\"; \"Why haven't we read this paper?! Still?!??! ... Gimme a post-mortem about\nthat separately later. I'm upset about this. /all-angles on BL3 please so we get this just right.\"; and, while the\npass ran: \"I mean maybe we read the paper at some point in some session but then it seems like we've been able to\nforget its contents/distinctions and our thinking has been muddled as a result. That paper is not all that's relevant\nfor structured analogies to be fruitfully compared in Deliberus however of course and so I want you to research this\nin depth and propose a path forward that is fully aligned with Arc BL otherwise.\" Nothing here is ruled; § 6 names\nwhat only he decides. The post-mortem he asked for is owed separately (`TODO.md`).*\n\n## 0. The answer first\n\n**The four products are Gentner's four kinds of learning from a structural alignment**, named in her own words in\n*Bootstrapping the Mind* (Cognitive Science 34, 2010, pp. 754–755, read from the PDF today): *\"Achieving a structural\nalignment sets the stage for four kinds of learning: abstraction, contrast, inference-projection, and\nrerepresentation.\"* Codex drew the four from that passage on 2026-11-07; the walkthrough of that evening spoke them as\na pattern, a difference that matters, a guess to check and a better description; the row of 2026-11-09 placed them in\nthe data model and dropped the theory's name. That is why the row read as unanchored: the anchor existed and was not\ncarried.\n\n**Why four.** A structural alignment maps two structured representations role to role (Gentner 1983; Falkenhainer,\nForbus and Gentner 1989). Once a mapping stands, every element of either structure bears one of five relations to it,\nand each relation is a product or its absence: **matched** (the common system: abstraction) · **present in the richer\nside, connected to the matched system, absent in the other** (a candidate inference, which the Structure-Mapping\nEngine defines as a base root that intersects the mapping and is not in it) · **in a matched role with a different\nfiller** (an alignable difference: *\"differences that occupy the same role in the two systems\"*) · **a predicate that\nshould match and does not** (rerepresentation, by minimal ascension or Yan, Forbus and Gentner's four repairs) · **in\nno role** (a nonalignable difference, which the Engine's own worked case says yields nothing). The partition into five relations\nis this note's synthesis of three published pieces, not a theorem any one source states (§ 2, the literature report\n§ 10); each piece is theirs, the join is ours, and it is said so.\n\n**Translated to the core data model.** Deliberus's structured representation is the argument: premises, a\ngeneralization, a conclusion, and the chain under a link (AG14). A comparison aligns two arguments role to role, and\n**the mapping is the one judgment** (the keeper's likeness: the two terms, the respect with `{x}`, what carries over,\nthe difference, the mechanism), recorded with its inputs and version like every model call. Each product is then an\nelement's relation to that recorded mapping, read off it into a kind that exists (BL1): the abstraction is the\ngeneralization both arguments instantiate, minted only where no claim already states it (AL16, match before mint),\nelse the existing one; a candidate inference is a claim nobody asserts (AG12) with its link marked proposing; an\nalignable difference is a claim at a role, and **the role names what it attacks**; a rerepresentation is a change to a\ndescription, never to a link: a concept distinction or shared sense, a sameness verdict, or a new claim replacing a\nwording (a claim's text is written once); a difference at no role is a stop in the record. Which product a comparison\nmade is a reading over what it wrote and the record, never a `product` stored on the link.\n\n**Where a difference lands, from the literature and seven real cases (§ 4).** The default is the carrying: an attack\non the transfer's relevance claim (AD27's slot). Every philosophical and argumentation source read holds the source's\nrule fixed and lets a critical difference block only the transfer (Bartha's *no critical disanalogy*, Juthe's\ndetermining relation, Walton's first critical question, Lamond's and Horty's reason model of precedent). The\nmembership claim is the target only where the difference is on a condition the generalization's own population\nstates (Farr v Butters; Read v Lyons; thalidomide, if its mechanism is ever settled). The generalization itself is\nattacked by a different move, never by the pair's difference: an eligible case that fails it (AT13's exception for a\nnormic one, a move of the count for a statistical one) or a rival case under the same respect (Walton's third\nquestion, the counter-analogy). The row of this morning, which sent the difference at the generalization, and the\nlater reading, which sent it at membership, each named one site and missed the default.\n\n## 1. The question, and what is decided\n\n- **Ruled and binding here**: BL1 (through what is already there) and BL2 (parsimony is counted, elegance examined,\n  and the core moves by ruling); S6 (kinship means a shared mechanism; a shared shape with no mechanism is a\n  metaphor); AD27 (each written-out rule shown with its strongest counterexample, as an attack on the link's relevance\n  claim); AG12 (a machine's sentence is a claim from the moment it is written); AT13 (an exception is a claim of its\n  own); BC8 (a consideration counts once); BI1, BI2, BI4 (a generalization and its cases calibrate each other; every\n  case in the graph); BD13 (two things are one only when every operation applies to both); AM8 and AM12 (a microtask\n  is grounded in the rulings and taught the method, never the graph's contents); Q33 and Q50 (relevance as the\n  difference a reason makes); the horizon ruling (read, never set); the 2026-08-31 ruling that bare judgments are\n  telemetry.\n- **Open and bearing**: BL3 itself; BH4 (a shared aim is no likeness) and BH5 (a judged kinship becomes a\n  generalization with its cases) in Arc BH; AM3 (the mind of Deliberus); W1107-what-a-comparison-produces-1, the\n  walkthrough's first question to carry (*should a judged comparison record which of the four it produced?*), which\n  this note answers.\n- **The open half, in one sentence**: how each of a comparison's products is to be thought of, with the theory it\n  comes from, and where each lands in the data model so that it is found, challenged and reused through the same\n  machinery as everything else.\n\n## 2. What the corpus already held, read whole\n\n| Angle | Read whole this session | What it holds |\n|---|---|---|\n| The four products' origin | `.private/collab/codex-gentner-and-deliberus-development-2026-11-07.md`; `docs/listening/2026-11-07-what-a-comparison-produces.md` | Codex's § 2 names Gentner 2010 § 2 (*\"abstraction, contrast, projected inference, and rerepresentation as products of comparison\"*) and the Engine's three criteria (structure, validity, relevance); its § 3.3: *\"A difference can defeat a particular transfer, define a narrower scope, expose a missing variable, or supply the informative contrast\"*, and treating it as an attack on the whole pattern *\"throws away information\"*; its § 3.2: membership alone does not say which roles correspond or which difference threatens a conjecture, so the mapping must be recoverable. The walkthrough spoke the four without the theory's name; Elliot's line, *\"Her later work adds that comparison can change how a situation is represented\"* |\n| Gentner's school, read before | `what-a-rule-rests-on/cognitive-science-sources-as-read-2026-10-08.md` (8 October: Gentner 1983 in passages, Gentner and Medina 1998 whole, Gentner 2003, Gentner, Rattermann and Forbus 1993, Christie and Gentner 2010, Gick and Holyoak 1983, Forbus et al. 1998 on *\"re-representation allows relational identicality to arise out of an analogical alignment\"*, Chalmers, French and Hofstadter 1992); `tools-and-analogies-as-engines-of-discovery/analogy-in-science-sources-as-read-2026-11-05.md` (Gentner et al. 1997, *J. Learning Sciences*: *\"analogy changes knowledge in four ways: highlighting, projection, rerepresentation, restructuring\"*, on Kepler) | The school was read twice, on the continuum from similarity to rule and on retrieval against soundness; Gentner's own four-way list was in the corpus since 5 November in its 1997 form (restructuring where 2010 has contrast) and was never joined to the products; Gentner 2010 § 2 and Gentner and Markman 1997 were not read before this pass |\n| Where analogy was placed | `overlaps-and-analogy.md` §§ 1–15b | An analogy is a relation between two cases carried by a pattern, never a key (§ 1); Carnap's intensional isomorphism as the structure axis (§ 4); Hesse's positive, neutral and negative analogy as the keeper's `respect`, `carried_over`, `difference` (§ 11, the terms table); **in a full graph the analogy is read, never asserted: two cases under one generalization, a membership claim whose tree decides how much carries, every disanalogy an attack in its place** (§ 12); the keeper's likeness built 2026-11-06 and the eleven candidates (§ 15, § 15a); Rutherford's atom: *\"the error lived in the warrant, which is where an attack belongs\"* (§ 4) |\n| The daemon and the kinship structure | `the-analogy-daemon.md` §§ 1–10 | Every analogy ships with its strongest counter-analogy (§ 5); the human-seeded channel (§ 7); Amgoud's four considerations, Relevance, Number, Diversity, Disanalogy (the IRIT harvest § 3); Woodward's test, *a kinship is a generalization that survives the change of field* (§ 9); **§ 10: the respect minted as a generalization that is a reference class, the two cases filed as members, the analogy read off the structure, the difference as the first attack on the relevance claim** (BH5, open) |\n| The both-ends loop | `cases-and-generalizations-from-both-ends.md` §§ 2–7 | Her findings mapped to the organs on 2026-11-06 (§ 2b: alignable differences as *\"the keeper's `difference` field and the first attack on the generalization (AD27)\"*, which this note narrows); the resilience pair's count withdrawn (§ 4b); the upward move over-produces, the downward under-admits (§ 4c); Codex's condition, *a difference becomes an attack only through an inference saying what it defeats* (§ 7) |\n| The sort | `claim-or-system.md` § 3 | Claim · record · reading · form, in that order of asking; *which shape the step has* is form recorded per step, never a fixed property |\n| Composition | `a-tree-per-claim.md` §§ 2d, 3–6 | A microtask never calls another; it writes a state and the other's condition owes it; the handoff table: a likeness with a shared mechanism writes *a draft analogy: the generalization, the two cases, the membership claim, the disanalogy as an attack* |\n| Likeness and the strength rulings | `what-a-rule-rests-on.md` §§ 5–8 | *Alike in what matters* is the link's relevance (Q33); the daemon can say what two cases share and not whether it bears; Keynes on the negative analogy as what further cases strengthen |\n| Kinds of likeness | `corpus-graph/what-a-move-is.md` § 4 | Six likenesses from which of a pattern's four parts two cases share; a judge asked only *is it the same move* refuses four of them |\n| The row and its reviews | `session26-open-rulings.md` rows BL1–BL3, S6, AD27, AG12, AM8, AM12, AT13, BC8, BD13, BH1–BH5, BI1, BI2, BI4; `.private/collab/codex-arc-bl-catch-up-review-2026-11-09.md` (Astra, cut off at its § 4) | Astra: the four are achievements, not exclusive kinds; a difference is not already an attack; the crowds case does not establish its attack; *read, never store* keeps BL1's history qualification; `analogy_candidates.py --judge` overwrites a verdict and records nothing (verified) |\n| His own words | `mannaminne --said` (2026-10-08, 2026-10-27, 2026-11-06) | *\"aren't multiple cases that are extrapolated from the base…\"* (8 October, the seed of Arc AY); *\"Did you read everything the corpus has on analogy and the connection to cases…\"* (27 October); *\"Can we also look back at what was said in the structure-mapping podcast\"* (6 November) |\n| Scanned, not read whole | `catching-a-proposal-made-unnecessary.md`, `codex-review-19-october-read-back.md`, `what-deliberus-is-for.md`, `four-books-the-corpus-had-not-read.md` | returned by the corpus graph's search on the helper report's subject; their hits on *analogy* and *comparison* are passing, none on the products |\n\n## 3. What the literature adds (two helpers, reports kept verbatim beside this note)\n\n`what-a-comparison-produces/literature-report-2026-11-09.md` (ten angles; Gentner and Markman 1997 and the Structure-Mapping Engine paper read whole by the helper, Yan, Forbus and Gentner 2003 whole, Horty 2011 and Lamond 2005 whole, Bartha through his 2020 restatement, Walton's scheme through five consistent secondary quotations, Weitzenfeld abstract only) and `what-a-comparison-produces/real-analogies-report-2026-11-09.md` (seven historical analogies, three well sourced). What the corpus lacked:\n\n1. **Alignable against nonalignable is a salience cut, not a relevance cut.** Markman and Gentner 1993: differences *\"connected to the common structure (alignable differences)\"* against *\"unrelated to the common structure (nonalignable differences)\"*; the 1997 review: alignable differences are more salient and weigh more in choice, and *\"it does not detract from the analogy that the planet does not look like a boat\"*. Nothing in either paper says an alignable difference refutes; structure-mapping is descriptive and leaves validity to other modules (the Engine: *\"Each candidate inference is only a surmise, which must be tested by other means\"*). So the keeper's `difference` is Hesse's negative analogy and Gentner's alignable difference, and whether it *matters* is a further claim.\n2. **Which differences matter is decided by the vertical relation, and the default landing is the carrying.** Bartha's critical factors (*\"those features that play an essential role in the prior association\"*) and his *\"no critical disanalogy: no critical factor belongs to the negative analogy\"*; Juthe's determining relation, which he identifies with relevance (*\"In this way, we define relevance\"*); Weitzenfeld's determining structure (abstract and secondary only); Walton's first critical question, *\"Are there differences between C1 and C2 that would tend to undermine the force of the similarity cited?\"*, which attacks the similarity premise and *weakens*. In every one the source's rule stands and only the transfer is blocked. This is AD27's slot, the relevance claim of the carrying link.\n3. **Membership is a separate use of comparison, and the law shows both sites.** Gentner and Markman's Figure 5 lists *categorizing* and *discovering a new category* as uses apart from inference. In precedent, Lamond's reason model: a later case *\"may clearly fall within the scope of the earlier ratio … but the later court may decline to reach the result C\"* (membership holds, the rule stands, a new reason defeats the carrying: R v Sharp on duress); Horty: distinguishing adds a trumping rule and refutes nothing; the SEP's Farr v Butters: outside the principle's own condition (*no possibility of intermediate examination*), a membership finding; Mullins 2026: narrowing a precedent so the new case falls outside it is *\"a form of constrained overruling\"*, a change to the rule, which the reason model should not call distinguishing. Three landing sites, one body of law, told apart by where the difference sits relative to the rule's stated conditions.\n4. **The rule is attacked by a counterexample case, not by the pair's difference.** Walton's third question (*\"Is there some other case C3 that is also similar to C1, but in which A is false?\"*), the counter-analogy; Shelley's taxonomy (via Juthe, secondary): a *false analogy* rejects the comparison, a *misanalogy* re-construes it to a new conclusion, a *disanalogy* and a *counter-analogy* accept the comparison and override it with further data or a rival case. Only the first two are about the pair's own difference.\n5. **Rerepresentation is a repair licensed by a good overall match, and its methods are open while its occasions are closed.** Yan, Forbus and Gentner 2003: *\"Rerepresentation re-construes parts of compared situations in order to improve a match\"*, licensed *\"only under special circumstances\"* such as *\"a very good overall structural match\"*; the occasions (holes, gulches, rivals, leftovers) *\"exhaust the constraints of structure-mapping\"*, the methods (transformation, decomposition, entity splitting, entity collecting, and the older minimal ascension) are *\"relatively open\"*. Gentner 2010 lists it as a product; Yan et al. treat it as a step outside the matcher. Both are right: the occasion is read off the mapping, the repair is a new description.\n6. **The real cases.** Carnot's heat as a waterfall carried two inferences from one abstraction: the work depends on the drop (held) and the quantity is conserved (failed), the difference landing on the carrying while Carnot's principle stood (Clausius, 1850, secondary). Darwin's selection: no difference mattered, and the one pointed out (no chooser) was in no relevant role; the product that did the work was a rerepresentation, *selection* widened past intention (Wallace's letter of 2 July 1866, the Darwin Correspondence Project). Thalidomide: rodents outside the sensitive population, a membership failure, its mechanism *\"enigmatic\"* still. Read v Lyons: no escape from the premises, outside Rylands's own condition, membership. The Wright brothers: the warping of the wings carried, coordinated turning needed a condition the bicycle lacked, the carrying. Card and Krueger: contested fate, the monopsony difference unverified in their own framing. Black and Scholes: a formal analogy with no difference recorded.\n\n## 4. Theory first, then the tests\n\n### 4a. The theory, in the project's terms\n\nA comparison is a structural alignment of two arguments. The **mapping** is the one judgment a model makes: which roles line up (the two terms), the common system stated as one general sentence (the respect, Hesse's positive analogy, the abstraction), what the richer side has that the poorer lacks (carried over, the neutral analogy, the candidate inference), the strongest difference in a matched role (the negative analogy, the alignable difference), and whether any predicate had to ascend to match (the mechanism label today; the rerepresentation occasion). It is recorded with its inputs and version as every model call is. **Products are readings over that record and over what was written**: nothing about them is stored on the link, and a person's verdict is a take on the reading (`reading_takes`), never a field overwritten.\n\nEach product lands in a kind that exists, by BL1:\n\n| Gentner | What it is in the graph | Where it lands | Why there (the sort, `claim-or-system.md` § 3) |\n|---|---|---|---|\n| abstraction (the common system) | the generalization both arguments instantiate | the claim that states it, if one exists (AL16); else a generalization minted asserted by nobody, a reference class, each case filed by a membership claim (BH5, BI4) | two people can disagree about it; it recurs under other links; cases hang under it |\n| inference-projection | a premise or conclusion the richer argument has and the other lacks | a claim nobody asserts (AG12), its link marked proposing | Gentner's school always marks it *to be checked*; the status *proposing* is that mark |\n| contrast (the alignable difference) | a different filler in a matched role | a claim at that role; its inference written out (AD27, AG14), and **the role names the target**: on the carrying, the transfer link's relevance claim (the default); on a condition the generalization's population states, the case's membership claim; on neither, nothing | the difference is a claim (disputable); what it defeats follows from where it sits in the mapping, which the record holds |\n| rerepresentation | two wordings or terms that should match and do not | a concept distinction or shared sense (the concept layer), a sameness verdict (two wordings, one claim), or a new claim replacing a wording; never a link | a description is changed, and a claim's text is written once |\n| (the fifth relation) nonalignable difference | a difference at no role | nothing: a stop answer in the record | the Engine's worked case: a base item outside the mapping yields no inference |\n\n**The rule is reached by a different move.** An eligible case that fails the generalization is a counterexample (AT13: an exception claim for a normic one; a move of the count for a statistical one, AY1); a rival case under the same respect with the opposite outcome is Walton's counter-analogy, filed under the generalization and counted (BC8). Neither is the pair's own difference, and a design that let the pair's difference attack the generalization would let a generalization escape criticism by excluding its failures (Astra § 3.2) and would count one consideration twice (BC8).\n\n**Several or none.** Carnot's one abstraction carried two projections with different fates; Darwin's comparison produced an abstraction, a rerepresentation and a projection, and no difference that mattered; the expected value cluster on the live graph produced memberships under a generalization already written and no new abstraction. So `products_of` is a reading over several objects, not a label.\n\n### 4b. Test one, the eleven live analogy candidates (`scripts/analogy_candidates.py --json`, 2026-11-09)\n\nEach read under three candidate readings: the row's morning wording (R1: the difference attacks the generalization or the relevance claim), the later reading (R2: the difference attacks the membership claim), and this note's (R3: the role names the target; the rule only by a counterexample case).\n\n| Pair (mechanism; verdict) | Products present | The difference, and where it lands | R1 | R2 | R3 |\n|---|---|---|---|---|---|\n| distributed resilience ~ group diversity (shared problem; **his** kinship) | abstraction (the respect, to be minted); projection (*guards against correlated failure* carried to the other system) | *structural robustness against takeover* against *epistemic accuracy of group judgments*: a different filler in the role *what the whole does*, which touches the carrying; and Codex's exact model shows a condition the respect lacks (the aggregation rule): a rerepresentation of the respect, a narrower generalization (AY4) | misfiles it at the generalization | misfiles it at membership (both are members) | carrying, plus a narrower generalization proposed |\n| small reductions in existential risk ~ a low-probability poverty solution (strict generalization; kinship) | **no new abstraction**: the generalization is already written (`claim_aig_fanaticism`); two memberships; no projection worth minting | *survival of all future value* against *a bounded benefit*: a filler difference in the role *size of payoff*, inside the generalization's condition, so it bears on neither carrying nor membership; a contrast kept | nothing to attack | nothing | nothing (recorded) |\n| the same ~ a positive Singularity (labelled partial pattern; kinship) | memberships under the written generalization | *reducing a risk* against *achieving a good*: in no role the rule names | nothing | nothing | nothing |\n| the fanaticism rule itself ~ the poverty case (neither) | **not a comparison of two cases**: a generalization and one of its cases; the product is a membership claim | none | no role for R1 | the membership claim is the whole product | membership filed, no attack |\n| existential risk reasoning licenses restrictions ~ the poverty case (kinship) | memberships; a projection (*what the rule licenses here it licenses there*) | *imposing restrictions* against *choosing between funding options*: a role difference (*what is done to present people*) that touches the carrying | misfiles | misfiles | carrying |\n| the same ~ a positive Singularity (kinship) | memberships | *avoiding catastrophe* against *pursuing a good*, *does not itself assert restrictions*: carrying | misfiles | misfiles | carrying |\n| the same ~ thirty years of poverty (kinship) | memberships; projection | *uncertain risk* against *a stipulated necessity*: in the probability premise, a premise an attacker denies, so an attack on a premise of the transfer, carrying | misfiles | misfiles | carrying |\n| AI governance should preserve control ~ research should prioritise Friendliness (neither) | none: a shared aim (BH4) | the two prescriptions can conflict | | | nothing; BH4 refuses the respect in code |\n| Madison's tyranny ~ its application to a machine (neither) | a generalization and its case | *an agent unlike the human hands Madison meant*: a difference on a condition of the definition's population, so a membership question, the one place R2 is right | misfiles | membership | membership |\n| early AI safety meetups funded DeepMind ~ founded OpenAI (neither) | none: one subject, two claims | none | | | code refuses |\n| technological advancement ~ advanced AI (partial pattern; unjudged) | a membership (AI is a technology, under A's own generalization); a projection (*AI expands the work of administering it*); a rerepresentation (*sector size* and *job content* ascend to *the share of human work spent controlling the technology*) | *a past correlation* against *a forecast*: a difference of kind, not of role; nothing to attack, a contrast | | | nothing |\n\nWhat the eleven show: in nine of eleven the abstraction was already a claim in the graph or one side *was* the generalization, so minting would have made a duplicate (AL16); the difference landed on the carrying in four, on membership in one, nowhere in five; R1 misfiled every carrying case and R2 every one but Madison's. The keeper writes a `difference` for every pair because the field is required, so most differences recorded are contrasts that defeat nothing, which is what Gentner's school predicts of alignable differences.\n\n### 4c. Test two, seven historical analogies with known fates (the helper's report, § 3 item 6)\n\n| Case | Products | The difference that mattered | Landing | R1 | R2 | R3 |\n|---|---|---|---|---|---|---|\n| Carnot: heat falling as water falls | abstraction (motive power from a conserved quantity falling through a level); two projections | heat is converted to work, not conserved | carrying (inference 2); Carnot's principle stood | misfiles | misfiles | right |\n| Darwin: artificial → natural selection | abstraction; projection (held); rerepresentation (*selection* without a chooser) | none; *no chooser* was in no relevant role | nothing | would have attacked | would have attacked | right (nothing) |\n| thalidomide: rodents → humans | abstraction (one placental mammal predicts another); projection (failed) | species differ in bioactivation; mechanism unsettled | membership, for rodents on this compound | misfiles | right | right |\n| Read v Lyons after Rylands v Fletcher | abstraction (strict liability for a dangerous thing brought onto land); projection (failed) | nothing escaped the premises: a condition the rule states | membership | misfiles | right | right |\n| R v Sharp after Graham (Lamond) | the duress test; projection (defence available: failed) | voluntary exposure to the threats: a new fact outside the test's stated conditions | carrying (a trumping reason, the rule untouched) | misfiles | misfiles | right |\n| Wright brothers: bicycle → aircraft | abstraction (stability by the operator); projection (roll control: held) | coupled roll and yaw needs a rudder | carrying (one inference needed a further condition) | misfiles | misfiles | right |\n| Rutherford: solar system → atom | abstraction; rerepresentation (gravity and Coulomb attraction ascend); projection (stable orbits: failed) | an accelerating charge radiates | carrying; the planets' rule stood; the electron stayed a member of *bodies under a central attraction* | misfiles | misfiles | right |\n\nR3 is right on all seven; R1 on none; R2 on two. The literature's default (carrying) holds in four of the five that had a difference that mattered; the two membership cases are both differences on a condition the rule itself states, which is the test R3 gives.\n\n### 4d. Test three, the row's own case\n\n*In crowds diversity is an intention, in governance an effect* as an attack on *what holds of crowds holds of institutions*. Under R3 it is not an attack at all: intending diversity and obtaining its effect are compatible (Astra § 3.3), so the sentence is a contrast in the role *how diversity got there*, which the respect does not mention; and *what holds of crowds holds of institutions* is not the generalization (it is an unrestricted transfer); the respect is. What Codex's exact model found (§ 4 of its note) is the real product: a condition the respect lacks (how local failures combine), which is a rerepresentation of the generalization into a narrower one (AY4), and until it is written the projection from crowds to institutions is held by that missing condition, an attack on the carrying. The row's case was the wrong case for the clause it illustrated.\n\n### 4e. What the tests do not settle\n\nA reading's correctness on seven sourced second-hand histories and eleven machine-written likenesses is evidence about the reading's shape, not a measured error rate for any microtask; no person but him has judged a difference's landing, and he has judged one candidate (the resilience pair, kinship). The membership test (*a condition the generalization's population states*) needs the generalization written in full; where only a respect with `{x}` exists, code cannot read the condition off it, and the keeper must say which role the difference touches (§ 6, BL4).\n\n## 4b. Set against what is settled\n\nExtends S6 (kinship as a shared mechanism: the mechanism is the generalization, now the abstraction product), AD27 (the counterexample as an attack on the relevance claim: the default landing, now with its theory), AG12 (a candidate inference as a claim nobody asserts), AT13 (the rule reached only by an exception or a count), BC8 (one consideration, one place), BH5 (the structure of a judged kinship, now with *minted only where no claim states it*), BI2 (the mapping read against its cases), AM8 and AM12 (Gentner's four and Bartha's test of critical factors are *method* to hand the keeper, never cases from the graph), BD13 (the products are readings, not objects, since no operation applies to a *product*). Narrows the row's first wording (the difference at the generalization) and the later reading (at membership) to the role rule. Reverses nothing settled. **Assumptions, weakest first**: that the partition into five relations is exhaustive for arguments as Deliberus stores them (the literature gives it piecewise; a test on more than eleven likenesses is owed); that the role of a difference can be read off a written mapping by a microtask with a measured error rate (unmeasured; BL4); that Shelley's four counterarguments map cleanly onto the graph's four attack sites (secondary source only).\n\n## 5. Recommendations, in dependency order\n\n1. **BL3 recut, the row's sentence** (the ledger row carries it; his call): *A comparison is a structural alignment of two arguments (Gentner), its mapping the one recorded judgment; each product is an element's relation to that mapping, read off it, and a comparison may yield several or none. The abstraction is the generalization both instantiate, minted only where no claim states it, each case filed by a membership claim; a candidate inference is a claim nobody asserts, its link marked proposing; an alignable difference is a claim at a role, its inference written out, and the role names what it attacks: the transfer's relevance claim by default, the case's membership claim where the difference is a condition the generalization's population states, nothing where the role is none; the generalization is reached only by a counterexample case; a rerepresentation is a change to a description, never to a link. Which product a comparison made is a reading; the mapping is recorded with its inputs and version, never a `product` stored on the link.*\n2. **BL4, new, proposed: the keeper's likeness names the role its difference touches** (*a condition of the respect* · *the carrying* · *none*), as one more slot of the likeness checked by code (AM14, AM21), so that routing the attack is code and never a second reading. A slot of form in the record, no graph property (BL1). Unbuilt; three real cases above would test it (Read v Lyons, Sharp, Darwin).\n3. **The grounding block for the keeper's likeness** carries Gentner's four kinds, Bartha's test of critical factors and Walton's three questions as method (AM8, AM12), rendered from this note once BL3 is ruled; BH3 asked for exactly this (*\"the result sets what the likeness field's three slots must carry\"*).\n4. **The census of the fourteen properties**, with the overwrite fix first, stays in `TODO.md` as work (filed earlier today); the acceptance test is Astra's round trip of one real comparison.\n5. **The post-mortem he asked for**, separately: why Gentner 2010's four and Gentner and Markman 1997 were read in part by Codex on 7 November, mapped to the organs on 6 November in the 1997 form, and still absent from the row of 9 November (`TODO.md`).\n\n**Alternatives set aside**: a fifth product (*a sharper question*, the walkthrough's best gift): a question is a potential rendered on read, so it is the candidate inference's or the contrast's question, not a product of its own; a stored `product` property on the resemblance link: a reading over the mapping, stale the moment a person reclassifies (BL1, Astra § 4.1); letting the pair's difference attack the generalization: BC8 and Astra § 3.2.\n\n**What only he decided, the same evening**: all five, in his words *\"Recommendations blessed. Build whatever is buildable.\"* (BL3 in the wording of item 1, BL4 as item 2; the grounding block, the census as work and the post-mortem as owed). What was then built: § 7.\n\n## 7. What was built the same evening, on his *\"Build whatever is buildable\"*\n\n- **BL4, the role slot**: `keeper.Likeness.difference_role` (`condition_of_respect` · `carrying` · `none`, default `none` so older likenesses still load), with its method in the field's own description (a condition the respect's population states; a critical factor the carrying depends on; a contrast); `keeper.likeness_faults` refuses a role with no difference and a carrying role with nothing carried over; `keeper.difference_target` routes to `membership`, `relevance` or nothing, and the generalization is unreachable from it by construction. The prompt's likeness paragraph asks for the role, so the keeper is a new version (`docs/research/microtask-versions.json`); nothing is re-run for it (R8). `scripts/analogy_candidates.py` prints, per candidate, which role the difference touches and what it attacks. Tests: `tests/test_keeper_likeness.py` (13).\n- **The overwrite fixed**: `analogy_candidates.py --judge` writes a person's verdict into the record first, as a take on the keeper's newest reading of the pair (`reading_takes`: kinship is `agree` with the likeness, metaphor and neither `disagree`, the verdict word in the note), then the link's four fields as a view of the newest verdict; a second verdict keeps the first; a link from before readings were named (2026-10-19) says so and gets the view only. `--takes ID_A ID_B` lists every verdict ever given. The eleven candidates' verdicts of 2026-11-06 were written before this and exist only as the view.\n- **The grounding**: BL3 is in the keeper's grounded rows (`deliberus/grounding.py`), rendered into `docs/research/grounding/keeper.pair.md`, so the ruled wording reaches the pass; Gentner's four and Bartha's test enter as the method in the likeness fields' descriptions (AM12), not as cases.\n- **Not built**: minting a judged kinship as a generalization with its members and the difference as the attack the role names, which is BH5 (open); the census of the fourteen properties (work, `TODO.md`); the daemon's second microtask.\n\n## 8. The unit of analogy, and what remains of the arc (2026-11-21, a session's synthesis for his word; unruled, no ledger row yet)\n\n*His ask, 2026-11-21: \"Could we try to build what remains of the whole analogies / unit of analogy / cases / induction arc? … Gimme your best synthesis after that.\" Read whole for it: this note, `the-analogy-daemon.md`, `cases-and-generalizations-from-both-ends.md`, Codex's review of 13 November and its three-comparisons study (`.private/collab/`), Gemini's Antigravity conversation of 12 to 14 November (transcribed in `~/dotfiles/docs/research/life-simulation/antigravity-conversation.md`), his messages in the three analogy sessions since 6 November, and 43 ledger rows of Arcs S, AY, BA, BC, BD, BH, BI, BK and BL (seven whole). Read in the passages named, not whole: Codex's review of 7 November (§§ 4.4, 4.5, 9, 10) and its catch-up of 11 November (its answer on the note about generalizations and cases read from both ends, the judgment that goes unused, the two candidate inferences). Not read this session: `overlaps-and-analogy.md`, `a-tree-per-claim.md`, `what-a-case-counts-for.md`, `reasons-and-cases.md`, Codex's Gentner note, Gemini's design note (its source conversation was read).*\n\n**Three takes on the unit, and where they meet.** Gemini (12 November): the unit is an isomorphic derivation subgraph, attached through its conclusion by `IN_CLASS` with the membership claim as grounds, and a compared argument is a second tier of *case*. Codex (7, 11 and 13 November): isomorphism asks too much and too little. One diagram can hide two mechanisms (in its three-comparisons study a reader matched how two records had been split and missed that the economics differed); a partial match is where a candidate inference comes from; one pair can carry two transfers that meet different objections (the pair on expected value: the arithmetic carries, the priority judgment may not); and two memberships under one generalization determine neither roles, direction nor the transfer, so *read, never stored* holds for the analogy's status and is too broad for the comparison. This note (BL3, ruled): the mapping is the one recorded judgment and its products are readings. His own two sentences set the test: *\"The conclusion merely is the point of connection but it brings in all the other structure around it\"* (to Gemini, 12 November) and *\"we need to consider all possible ways in which the MEANING of all involved claims describe/encode analogy case/s\"* (to Codex, 13 November).\n\n**Proposed: the unit is the transfer.** A comparison has no fixed size. What is addressable is each thing it proposes to carry from one side to the other, and each transfer is one link of a kind that exists, with its own objections. The resilience pair (his verdict: kinship) as it would sit in the graph:\n\n- A generalization, minted, asserted by nobody, also a reference class (BH5, ruled; unbuilt): *A system that spreads function or perspective across many independent parts resists a single shared failure or dominant pull; where it does not, something explains it.*\n- *Where distributed resilience exists, it reduces capture risk, monoculture risk, and the risk of systemic takeover* (the AI governance text) lies within it (`WITHIN`): a narrower generalization, which adds no count.\n- *Ensuring group diversity is intended to mitigate the negative impact that consensus-seeking behavior has on the accuracy of a crowd's collective judgment* (the crowds page) lies within it too.\n- One transfer: a claim nobody asserts (AG12), in the keeper's own words for what carries over (the keeper recorded *guards against correlated failure*), its link marked proposing, the link's relevance claim grounded in the generalization (AD22).\n- The keeper's difference (*structural robustness against takeover* against *epistemic accuracy of group judgments*) as a claim attacking that transfer's relevance claim (BL4, role *the carrying*; the routing is built). ⚠ Corrected below: as recorded the contrast attacks nothing, and the role was this session's reading.\n- In the record, not the graph: which claims and links on each side played which role, what was aligned (the relation in the world the claims describe, the argument for a claim, or a decision rule; Codex's three), and which keeper reading he judged. The minted generalization names it by `minted_from`.\n- Observed cases with outcomes: none filed yet. Only these count (BH5 as amended, BI4).\n\nSo no object called *analogy* and no subgraph type. A selection larger than one claim is named in the record by the ids it used and grows when an objection finds a missing condition (Codex, 13 November: *start with the smallest selection that explains the mapping*); where a selection is exactly one argument, BI3's stored argument (open) would give it an id. Where nothing is carried over, BL4 already answers: the difference has role *none* and nothing is written.\n\n**One word.** Gemini's *Tier 2 case* gives *case* a second sense, against BD9 (a case is one instance under a generalization, counted once). The words of structure-mapping (Gentner's theory of analogy) for the two things compared are the *base* and the *target*, together the *analogs*. Proposed: say *analog*; an analog is a case only when it is an observed instance with an outcome. The resilience pair shows why it matters: both analogs are generalizations, so they add no count.\n\n**What remains, in dependency order.**\n\n*Ruled and unbuilt, no decision owed:*\n1. The writer BH5 rules (**built in part the same night**: `deliberus/kinship.py` and `scripts/analogy_candidates.py --mint`, seven tests; it mints, files, keeps `minted_from`, `minted_at`, `minted_by`, and refuses a verdict that sits on an earlier reading; reuse of an existing generalization is not done; run on the live graph for the resilience pair the same night, `claim_kin_0386d1e4fb96`; the sentence, population and outcome are worded by whoever mints): `scripts/analogy_candidates.py --mint` on a pair judged *kinship* mints the generalization, or reuses a claim that already states it at the same scope (class sameness, AL16); files generalizations `WITHIN` and observed cases `IN_CLASS` with role *formulation*; writes `minted_from`, `minted_at`, `by`; and takes the keeper reading he judged, never the newest (Codex's acceptance case of 11 November: a verdict on one respect shown beside a later, different respect). Run twice, it writes once. First on the resilience pair (his blessing names it), then on the six candidates about expected value, where the generalization already exists and its wording (*astronomical*, *every other consideration*) is checked against the poverty claim before reuse.\n2. BL3's products on the same write: each transfer a claim nobody asserts with its link proposing; an objection through an ordinary attack link, its inference the grounds of that link's relevance claim, routed by BL4's role (as corrected below). **Built the same night** (`kinship.write_transfer`, `write_objection`, `write_membership_objection`, `undo`; `scripts/analogy_candidates.py --transfer`, `--object`, `--undo-proposals`): every sentence is worded by whoever calls, and the microtask that words them from the keeper's likeness is not built. The resilience pair carries one transfer and Codex's objection on the live graph, both a session's wording, unread by him.\n3. BI4: `class_count` reads filed cases first, bundles for what no case covers. Not built: nothing in the data model records a filed member's outcome, or which bundle counted a filed case, and both are his to decide (`TODO.md`, the analogy loop).\n4. The case channel for a rule under a link: AY2, AY4, AY5, AY6, BA1, BA2, BD20, AT13.\n5. The keeper's inspection of a candidate whose endpoint was cut: fetch the conditional, its antecedent, its consequent and the source passage (Codex, 13 November).\n\n*For his ruling, one at a time:* (1) each proposed transfer has its own inferential address, the comparison's selection kept in the record (this section, as narrowed below); (2) BI3, an argument has an identity of its own (open); (3) *analog* as the word; (4) BH6, the hunt owes cases for every written generalization (open; after the writer exists); (5) BH4, a shared aim is no likeness (open; after his reading of the ten).\n\n*For a reading, one case at a time (Y9); today his alone, shareable with a second reader once there is one:* analogy candidates two to ten (the keeper re-run first, since its version changed), Kelvin's vortex atom for BH3, pairs two to ten of the nine kept out of the keeper's examples.\n\n*Designed, further off:* the daemon's second microtask (the strongest rival case, owed with every analogy), the sweep that proposes a prior, the roaming signal (BH1), the channel where a person seeds an analogy, and Gemini's retrieval by a card that states problem and mechanism with the field's nouns removed, which BH6's use of a generalization's wording as the hunt's probe already half does.\n\n**His question the same night, on the difference as a claim** (verbatim: *\"Hmm, the Difference as a Claim, does it pass the Stranger Test? Does it need to? Did Gemini and Codex reflect on this? Can you ask Codex to reflect on this with you, and on the parsimony doc/rulings? I feel like this is core...\"*, then *\"(Maybe I misunderstand...)\"*). A session's answer, unruled. The keeper's `difference` is a slot in the record, two noun phrases (*structural robustness against takeover* against *epistemic accuracy of group judgments*); as recorded it names neither system, so it fails the stranger test, and once it is written as a claim it must pass it like any stored claim (AP3: a claim a microtask writes names every referent). One claim is also not enough: a sentence saying how the two analogs differ is true and attacks nothing; the attack is a second claim, the inference that this difference defeats this transfer (Codex, 7 November, § 4.5; AD27; AG14), and that inference is where a reader disagrees. So a difference that matters is two claims at the floor: the difference itself, naming both analogs, and what it defeats (⚠ withdrawn the same night, see the next paragraph); where it defeats nothing, BL4's role *none* writes nothing and it stays a slot in the record. In what this session read, Gemini put the stranger test to an answer under a question (his *\"The property\"? \"The lawn\"? :)*) and Codex named the further inference; neither put the stranger test to the difference. Put to Codex the same night: `.private/collab/brief-for-codex-difference-as-claim-2026-11-21.md`, its reflection to be saved beside it.\n\n**Codex's reflection the same night, and what it corrects here** (`.private/collab/codex-difference-as-claim-reflection-2026-11-21.md`, read whole; Codex changed no file, and its worked example is its own reconstruction, no decision of his). Unruled, like the rest of this section.\n\n- *Where the stranger test applies.* Three representations owe three things. A slot in the comparison record must be recoverable from the inputs the record names, its other slots and its version. A stored claim must state its proposition, its referents, its scope and its qualifications for a stranger (AP3). A shortened label on display must leave the whole proposition open to inspection. So the keeper's `difference` passes as a slot and fails as a claim, and so does `carried_over`: *guards against correlated failure*, or any sentence still carrying `{x}`, is no stored proposition. A claim resolves the referents its own proposition uses; nothing requires every objection to name both analogs.\n- *\"Two claims at the floor\" is withdrawn.* What BL3, AD27 and AG14 require is one explicit, addressable inference, and the graph already has its home. The claim that does the work attacks the transfer link's relevance claim by an ordinary attack link, and why it bears on that relevance claim is written as the grounds of **the attack link's own relevance claim** (`machine_sentences.write_under_link`). The second claim asked for above was that relevance claim under another name. How many claims an objection takes follows from the separability and smallest-inference tests on the objection itself, and a claim the graph already holds is used before one is written.\n- *A difference is not yet an objection.* The resilience contrast (takeover against accuracy) supplies no bridge, since different outcomes can instance one mechanism; it stays in the record as a contrast until an argument uses it. The list above drew it as already attacking and gave it the role *the carrying*. The stored likeness predates BL4 and names no role (read on the live graph this night: none of the eleven candidates carries one), so that role was this session's reading and no judgment of the keeper's or his.\n- *The unit, narrowed.* Three identities are kept apart: the comparison (the selection and mapping somebody judged, in the record), the projected proposition (what is proposed of the target), and the inference (the grounds offered for it; two inferences can reach one proposition, and one comparison can propose several). **Each proposed transfer has its own inferential address** stands. *The unit of analogy is the transfer* says too much: a comparison can yield an abstraction, a contrast or a redescription and propose nothing (BL3). No `Transfer` kind is added, and BI3 stays open.\n- *Five more places where this section said more than its sources.* Role *none* does not mean nothing is written wherever nothing is carried over: a difference can touch membership with no projection, and `keeper.likeness_faults` allows it. The writer's guard refuses a verdict that sits on an earlier reading; it does not recover the judged mapping from the record and mint from that. A selection that grows on an objection must leave the selection that was judged recoverable. BD9 alone does not forbid calling a compared argument a case of an argument pattern; the limit on counting comes from BH5 as amended, BI4 and AY6. And BH5's *carries the generalization as its relevance claim* beside this section's *relevance claim grounded in the generalization* is a reconciliation (a reusable rule kept apart from one application's warrant), not one wording.\n- *Parsimony, counted (BL2).* Parsimony counts kinds of object, link types, stored properties a decision reads and unrecorded settings, never claim instances. A claim attacking a relevance claim with ordinary grounds, or premises the graph already holds: no kind, no link type, no property. A stored flag saying the difference defeats the transfer: one property, and a defeat nobody can inspect. A `Transfer` object: one kind and at least three fields. Always writing a difference and a sentence that it defeats: no kind, and a fixed decomposition nobody chose.\n- *The resilience pair, worked by Codex.* A transfer needs a direction and a conclusion. Codex's reconstruction: *In collective-judgment systems, reducing dependence among component errors reduces the probability of an incorrect collective judgment.* Whether that follows depends on how the parts' failures combine. For three parts each failing with probability 0.2 (a calculation under stated assumptions, no observation of institutions or crowds): all three fail with probability 0.008 when the failures are independent against 0.2 when perfectly shared; a majority fails with 0.104 against 0.2; any one fails with 0.488 against 0.2. So independence helps a whole that fails only when most of its parts fail and hurts a whole that fails when any part does. The objection this justifies: *the comparison's stated grounds do not establish the aggregation conditions needed to infer improved collective accuracy from reduced dependence among component errors.* It attacks the inference; it shows neither the conclusion false nor the kinship empty.\n- *The same objection reaches the generalization's wording.* Its last clause (*where it does not, something explains it*) states no condition that tells cases apart, and its first leaves the combination of failures unsaid. **State:** the generalization was minted on the live graph with that wording before this reflection was read: `claim_kin_0386d1e4fb96`, both claims within it, no count, embedded for the membership sweep, from his verdict of 2026-11-06 on the pair's one keeper reading (in the record: made the night before the verdict, none since). A claim's text is written once (AL10), so a sharper wording is a new claim and his to decide; until then the objection stands as an objection to that claim.\n\n## How this was done\n\nThe six parts of the all-angles pass: question pinned from the row and the rulings read whole (§ 1); the corpus inventoried by the corpus graph's search and read whole (§ 2); two helpers for what the corpus lacked, their reports kept verbatim; theory before tests; tests on the eleven live candidates (one model-free query against the live graph), seven histories and the row's own case; set against the settled rows; recommendations by dependency. **Read whole this session**: Gentner 2010 (pp. 754–755 of the PDF, the four kinds); `.private/collab/codex-arc-bl-catch-up-review-2026-11-09.md`; `.private/collab/codex-gentner-and-deliberus-development-2026-11-07.md`; `docs/listening/2026-11-07-what-a-comparison-produces.md`; `overlaps-and-analogy.md`; `the-analogy-daemon.md`; `cases-and-generalizations-from-both-ends.md` §§ 2, 4, 4b, 5, 6, 7; `claim-or-system.md` § 3; `a-tree-per-claim.md` §§ 2d–6; `what-a-rule-rests-on.md` §§ 5–8; `corpus-graph/what-a-move-is.md` § 4; `the-irit-harvest.md` § 3; `what-a-rule-rests-on/cognitive-science-sources-as-read-2026-10-08.md`; `tools-and-analogies-as-engines-of-discovery/analogy-in-science-sources-as-read-2026-11-05.md` (the Gentner 1997 lines); the two helper reports; ledger rows BL1–BL3, S6, AD27, AG12, AM8, AM12, AT13, BC8, BD13, BH1–BH5, BI1, BI2, BI4; `deliberus/keeper.py` (`Likeness`, `likeness_faults`); `scripts/analogy_candidates.py`; `docs/graph-properties.md` lines 110–135. **Scanned, not read whole**: the four notes named in § 2's last row, and three more the corpus graph returned for this note's subject (`session49-what-goes-under-a-link.md`, `checks-on-machines-and-people.md`, `the-feed-reconsidered.md`), whose hits on *analogy* and *comparison* are a floor's analogy to the chain and a part test, none on the products. **Not done**: the outside check on this draft waits for his word on spend; no model call was made by this session beyond the two helpers.\n\n## Cross-references\n\n[parsimony-and-the-core-data-model.md](parsimony-and-the-core-data-model.md) § 6 (BL3 as first proposed; Astra's review) · [overlaps-and-analogy.md](overlaps-and-analogy.md) §§ 11, 12, 15 · [the-analogy-daemon.md](the-analogy-daemon.md) §§ 7, 10 · [cases-and-generalizations-from-both-ends.md](cases-and-generalizations-from-both-ends.md) §§ 2b, 7 · [saying-what-kind-a-claim-is.md](saying-what-kind-a-claim-is.md) § 10 · [what-a-rule-rests-on.md](what-a-rule-rests-on.md) §§ 7, 8 · [claim-or-system.md](claim-or-system.md) § 3 · the ledger, Arc BL.\n"}