{"path":"research/overlaps-and-analogy.md","content":"# Overlaps and analogy: two relations on two axes\n\n*2026-10-27. His question, verbatim: \"Think hard about \"overlaps\" vs \"analogous\" from a formal logic / analytic\nphilosophy perspective. Help me see this clearly: what's the load-bearing idea/s? One bridging analogy. Competing\nperspectives if relevant. And after you have answered that: what doesn't fit neatly into that view?\" Asked after the\nkeeper's first unseen pair (`one-inference-per-link.md` § Held-out pair 1), where he said \"there's certainly an\nANALOGY here, and \"unrelated\" is thus not really true\", and after Claude proposed filing an analogy under `overlaps`.\nClaude's reading; nothing here is ruled, and the pair-1 thread is paused until 2 November on his word.*\n\n## 1. The answer\n\n`overlaps` and *analogous* are not two answers to one question. **`overlaps` is a relation between what two claims\nsay; an analogy is a relation between two cases, carried by a pattern they share.** So analogy is not an eighth key\non the keeper's menu and does not belong under `overlaps`, and the keeper stays the keeper of seven keys. Two rulings\nalready say what an analogy must be: **kinship means a shared problem and a shared family of solutions, and a shared\nshape alone is a metaphor** (ledger S6, ruled 2026-10-25), and **a written-out inference records its shape, one of\nwhich is from a like case** (ledger AD23, ruled and built as `machine_sentences.STEP_SHAPES`). And the corpus already\nsays where the likeness is argued: as the link's relevance claim, *alike in what matters*\n([what-a-rule-rests-on.md](what-a-rule-rests-on.md) § 7, under ruling Q33).\n\n*Changed the same evening, after reading the corpus whole on his question \"Did you read everything the corpus has on\nanalogy and the connection to cases and all that?\": the first version of this note made analogy a sameness of build,\nwhich S6 calls a metaphor, and proposed adding analogy as a kind that the claim graph's schemes and AD23 already have.\nClaude's earlier proposal to file an analogy under `overlaps` stays withdrawn.*\n\n## 2. What is decided, and what is open\n\n- **Decided.** The seven answers (same, A contains B, B contains A, opposes, overlaps, unrelated, can't tell), blessed\n  2026-08-31, a coarsening of MacCartney and Manning's seven relations (`soft-canonical-clustering-and-reversible-\n  merge-semantics.md`). The keeper's `overlaps` is evidential: establishing one partly establishes the other, through\n  one shared inference both claims assert (`deliberus/keeper.py`). An analogy owes its respect as a rule\n  (`one-inference-per-link.md`, the kinds table).\n- **Open.** Whether and where a pair's analogy is recorded, and whether analogy joins the keeper's kinds.\n\n## 3. What the corpus holds, read whole for this note\n\n| Note or row | What it holds that bears here |\n|---|---|\n| [what-a-rule-rests-on.md](what-a-rule-rests-on.md) | Likeness sits in two places under a link: behind the rule (the cases it came from) and in the tie of this case to it. Cases give a rule when compared; which likeness matters is a wider rule, learned one level up (§ 5). *Alike in what matters* is the link's relevance (§ 7). The claim graph stores *analogy* (28 links) and *precedent* (9) as kinds with three standing questions (§ 2). Likeness is judged partly by the conclusion wanted (Braman and Nelson; § 13.2), so the likeness question is asked before the conclusion is shown. People are reminded by looks and convinced by structure (Gentner; § 4). Seth's weather-simulation analogy, from the same debate as pair 1, turns on whether consciousness is like wetness or like adding (§ 8) |\n| [the-analogy-daemon.md](the-analogy-daemon.md) | Every analogy ships with its strongest counter-analogy; the daemon writes a likeness out and cannot settle it; a kinship is a generalisation that survives a change of field (§ 9) |\n| [corpus-graph/what-a-move-is.md](corpus-graph/what-a-move-is.md) | A pattern has four parts (what is wanted, what stands in the way, what is done, what comes of it), belongs to the pair, and six kinds of likeness follow from which parts two cases share |\n| [careful-and-broad-minds.md](careful-and-broad-minds.md) § 3–4 | Broad matching proposes and a structural check accepts; in the one measurement, 22 of 30 top machine pairs were not genuine |\n| [what-a-case-counts-for.md](what-a-case-counts-for.md) | Cases are filed under the rule whose condition they meet (AY2, ruled) |\n| Ledger S6, AD23, AD27, AY2 | Ruled: kinship versus metaphor; the shape of a written-out inference; the strongest counterexample shown; a case counts once, under its rule |\n| [soft-canonical-…](soft-canonical-clustering-and-reversible-merge-semantics.md), [sameness-merge-…](sameness-merge-and-split-across-fields.md) | The seven keys and the set relations behind them (read in part) |\n| Carnap's intensional isomorphism | not in the corpus; read in *Meaning and Necessity* § 14 |\n\n## 4. The theory\n\n**Content, in logic.** MacCartney and Manning sort two sentences by their truth-sets into seven relations:\nequivalence, forward and reverse entailment, negation, alternation (exclusive but not exhaustive), cover, and\nindependence, which \"covers all other cases\" and is the least informative. The keeper's `same`, the two containments\nand `opposes` are the first five (opposes taking negation and alternation together). **Both `overlaps` and `unrelated`\nsit inside independence**: logic cannot tell them apart, and the keeper splits them by probability (does establishing\none change the other's?). So the content axis is already two layers, logic for the first four keys and confirmation\nfor the next two.\n\n**Structure, in logic.** Carnap (*Meaning and Necessity*, 1947, § 14): two sentences are *intensionally isomorphic* if\nthey are \"built in the same way out of designators … such that any two corresponding designators are L-equivalent\";\nhe offers this as the explication of *synonymous*. Drop the last clause, keep the same build with a mapping between\nthe parts, and what remains is analogy as Gentner's structure-mapping describes it: the same relations among\ndifferent objects. So **sameness is analogy whose every mapped pair is equivalent in meaning**, and analogy is\nisomorphism of build without that.\n\n**Bearing.** Structure alone carries no evidence: by Goodman any two things are alike in some respect. An analogy\nmoves a probability only through a further claim, its warrant: things alike in the mapped respect are alike in what\nis carried over (Hesse, Bartha; Norton adds that the warrant is a particular fact, never a general schema). So\nanalogy enters the graph as an inference of kind *analogy*, whose relevance claim is its warrant, which is how the\nkinds table already has it.\n\n**Content and pattern**, with cases not used before in the corpus (a shared build counts as an analogy only where, by S6, the two share a problem and a family of solutions; otherwise it is a metaphor and bears nothing):\n\n| | Shared content | No shared content |\n|---|---|---|\n| **Shared build** | same: *remote work raises productivity* / *working from home makes people more productive* | analogous: *electrons orbit the nucleus, held by electric attraction* / *planets orbit the sun, held by gravity* |\n| **No shared build** | overlaps: *Card and Krueger found no job loss in New Jersey* / *modest minimum-wage rises do not reduce employment* | unrelated |\n\nThe axes are independent, which a third case shows: *raising the minimum wage reduces employment* and *raising the\nminimum wage increases employment* have one build and opposite content (mapping *reduces* to *increases*). An analogy\ncan sit beside any of the seven keys, so it cannot be one of them.\n\n**A failure on record.** Rutherford's atom borrowed the solar system's build; the warrant (an orbit held by a central\nforce behaves the same at both scales) fails, since a classical orbiting charge radiates and spirals in, which Bohr's\nmodel of 1913 answered. The build matched and the carried-over conclusion did not: the error lived in the warrant,\nwhich is where an attack belongs.\n\n## 5. The bridging analogy\n\nTwo maps. Maps of overlapping ground share territory: some streets are on both, and what one shows about them the\nother confirms. Maps of two different cities drawn with the same legend and the same pattern of roads share a build\nand no ground: one teaches you to read the other, and tells you something about the other city only if cities like\nthese are laid out alike for a reason. Two copies of one map share both.\n\n## 6. Competing perspectives\n\n- **Extensional (Carnap's semantics, MacCartney):** relations hold between truth-sets; analogy is a heuristic for\n  finding claims, not a relation the logic knows. Keeps the keeper to content and leaves analogy to the analogy\n  daemon.\n- **Confirmational (Carnap's relevance, the keeper's prompt):** only bearing matters, so an analogy counts exactly as\n  far as its warrant moves a probability. Supports recording the warrant as a claim and nothing else.\n- **Structure-first (Gentner, Hofstadter):** analogy is the core of cognition and sameness its limit case; the build\n  should be recorded whenever it exists, whether or not it bears.\n- **Material (Norton):** there is no general logic of analogy; each is as good as its local backing fact, so a generic\n  *analogous* label says nothing until the fact is written.\n\nThe two-axis reading takes something from each: content keys from the first, bearing only through the warrant from\nthe second and fourth, the build recorded as an object from the third.\n\n## 7. What does not fit neatly\n\n1. **Sameness is the limit of analogy**, so on the structure axis the cells are not disjoint: a partition menu cannot\n   hold it, which is one more reason analogy is not a key.\n2. **Build depends on how a sentence is parsed.** Carnap's notion was attacked as too strict (Mates's puzzle about\n   belief and synonymy, and Church's reply), and two parses of one sentence give two mappings. Choosing the logical\n   form is itself an argument.\n3. **Structure is graded.** Gentner's systematicity ranks matches by how many tied relations they carry over; a key is\n   binary.\n4. **Pair 1 is neither clean cell.** A and B are two descriptions of one theory, which is Frege's case of two senses\n   with one referent, as much as an analogy between brain and model. Three relations stack in it: co-reference (both\n   describe global workspace theory), a vocabulary mapping between cognitive and neural terms, and, in the source\n   context, an analogy between a human brain and a language model. The two axes do not name the first.\n5. **The content axis is two layers**, logic and confirmation, and they can disagree: logically independent claims can\n   bear on each other, which is the whole of `overlaps`.\n6. **Without a respect, *analogous* is never false** (Goodman), so only the written mapping and warrant can be\n   assessed, never the label.\n\n## 8. Recommendations (unruled; his word on pair 1 waits until 2 November)\n\n1. Keep the seven keys as the content relation; an analogy is never a key and never filed under `overlaps`.\n2. Where the keeper sees a likeness, it writes it as the rulings already shape it: the two cases filed under the wider\n   rule they share (AY2), or the pattern with its four parts, and it says whether the two share a problem and a family\n   of solutions (S6); without that it records a metaphor, which bears nothing.\n3. The likeness bears only through a like-case link (AD23's shape; the claim graph's *analogy* kind) whose relevance\n   claim says what matters, with the strongest counter-analogy attached (AD27).\n4. The gate asks about bearing of any kind, analogy included, and the likeness question is asked before the\n   conclusion is shown (Braman and Nelson).\n5. Finding analogies across the graph stays the analogy daemon's job; the keeper records only the pair it is given.\n\n## 9. Ruling S6 read again (2026-10-27, on his question \"Ruling S6 - is it sensible? Reflect.\")\n\nS6, as ruled 2026-10-25 (*\"All ruled, please build them\"*): *\"In the daemon's second pass, kinship means a shared\nproblem and a shared family of solutions; a shared shape alone is a metaphor.\"* Home:\n[major-transitions-heredity-and-evolvability.md](major-transitions-heredity-and-evolvability.md) § 8. Claude's\nreading, unruled.\n\n**What holds.** As a gate on what an analogy may carry as evidence, it is the test the literature gives: Hesse\nrequires that the shared features be causally tied to what is carried over and that no essential difference break\nthe tie (Bartha, *Stanford Encyclopedia*, § 3.3); Gentner's systematicity prefers tied relations to loose\nattributes. The major transitions are its model case: each faces the lower level's selfishness against the\nwhole's function and solves it from one short list. And its guard is right: name the problem at the level where\nthe mechanism works, or any two things share one.\n\n**What does not fit.**\n\n1. **Kinship without a problem.** A pendulum and an oscillating electric circuit solve nothing; the same equation\n   governs both, which is Hesse's *formal analogy*, and analogue computers ran on it (a capacitor's voltage standing\n   for a mass's velocity). That is not a shared shape in S6's sense, since a known law licenses what is carried\n   over, and it is not a shared problem either. S6's wording fits functional systems (organisms, institutions,\n   practices) and leaves out lawful ones (physics, much of chemistry).\n2. **Partial kinship.** [corpus-graph/what-a-move-is.md](corpus-graph/what-a-move-is.md) § 4 sorts likenesses by\n   which of a pattern's four parts two cases share: the same means to another end, the same difficulty noticed\n   twice, one mechanism judged two ways. Each licenses something smaller than full kinship. S6 calls them all\n   metaphors.\n3. **What a metaphor is worth.** In S6 *metaphor* is the losing side. In the corpus a metaphor carries no evidence\n   and still does discovery work: Hesse's neutral analogy raises questions, and the human-seeded channel exists to\n   develop exactly such seeds ([the-analogy-daemon.md](the-analogy-daemon.md) § 7).\n\n**The wording, RULED 2026-11-05 (*\"Do it\"*, then *\"Ruled\"*; first proposed for his ruling with pair 1 on 2 November, ruled after the walkthrough *Judging the connections*):** *In the daemon's second microtask, kinship means a shared\nmechanism, named at the level where it works: a shared problem with a shared family of solutions, or one strict generalization\nthat both cases instance (Hesse's formal analogy: the pendulum's equation is the circuit's). A likeness that shares some of a pattern's parts carries what those parts license. A shared shape\nwith no mechanism is a metaphor: it carries no evidence and may still seed a question.*\n\n*Law* replaced 2026-11-05, on his question after hearing the walkthrough (*\"what is 'a law shared'? Is that a rule/generalization? ... I'm very annoyed at muddling ambiguous words like 'law' being used if there's already an established technical term that we've begun using\"*): the project's term is *generalization*, its kind *strict* (AT12), and Hesse's formal analogy is one strict generalization that both cases instance, which is also the shape § 12 gives an analogy (two cases under one generalization, AY2). The walkthrough's spoken lines stand as rendered; its header carries the correction.\n\n## 10. What goes into the keeper (2026-10-27, his word: \"So tie this back to where we started. Weave this all into what should go in the keeper.\")\n\nClaude's proposal, unruled; it waits with pair 1 for 2 November and builds nothing before his word (the hold on\nkeeper runs stands). Read for it: `deliberus/keeper.py` (the answer form, `resolve`, `ROUTING`, `CLASSIFY_AFTER`),\n`machine_sentences.STEP_SHAPES`, the claim graph's *analogy* scheme in `deliberus/extraction/schemes.py`.\n\n**Where we started.** Pair 1 (the Anthropic text's *information becomes consciously accessible when it enters a\nsmall shared channel* against Seth's *conscious experiences in humans occur when information is made widely\navailable to other parts of the brain*). The keeper said `same`; his verdict was *too different*; clingo counted\nthree premises one way and five the other, so not `same`; played by its prompt the keeper reads `unrelated`; his\ncorrection: *\"there's certainly an ANALOGY here, and \"unrelated\" is thus not really true\"*.\n\n**Why the keeper loses it today, in its code.** Three places. The gate's description asks for bearing *\"through a\nshort chain whose steps the claims themselves assert\"*, so a likeness, which neither claim asserts, reads `no`.\n`resolve()` turns a refused `overlaps` into `unrelated`. And the chain's kinds are strict, normic, statistical and\ninduction, with no like case, though AD23 built `like_case` as a step shape and the claim graph has an *analogy*\nscheme with its standing questions. Because a stop answer keeps the pair from the relationship classifier\n(`CLASSIFY_AFTER`), no analogy link between two texts can ever be written through the keeper.\n\n**What to change, in dependency order.**\n\n1. **The gate asks about relevance of any kind** (Carnap's probabilistic relevance, § 11): does granting one claim\n   change the other's probability, by applying a generalization, by an induction, or by an analogy. A shared subject is\n   still not shared evidence.\n2. **The seven keys answer one question only: how the two contents stand** (same, one contains the other, opposed,\n   overlapping in a step both assert). They stay seven. A likeness is never a key.\n3. **A likeness is written as its own field, before the chain** (field order is asking order, so what the two share\n   is written before anything is said about bearing): the term each claim speaks of (here *the human brain* and,\n   from the source passage AM9 hands over, *the language model*); the respect, as a rule with `{x}` (*a system in\n   which information in a shared channel becomes available to the whole system is conscious of that information*);\n   what is carried over; the strongest relevant difference (AD27; here *accessible* against *experienced*); and the\n   mechanism, one of four: a shared problem with a shared family of solutions, one strict generalization both cases instance, a partial\n   pattern (which of its four parts are shared, by `what-a-move-is.md` § 4), or none.\n4. **Code checks the likeness as it checks a chain** (AM14, AM21): the respect is general (it does not name either\n   term) and, said of each term, gives a sentence about that case; a difference is named; the mechanism is one of the\n   four. Nothing is written twice.\n5. **Routing follows the mechanism.** A shared mechanism, full or partial, with the gate at `yes` and no content key:\n   the pair is distinct and related, so the relationship classifier may run, with the *analogy* scheme offered, and\n   the link it writes carries the respect as its relevance claim (Q33) and the difference as its first attack. `none`:\n   a metaphor, which writes no link and is kept as a seed for the analogy daemon's human-seeded channel. A refused\n   `overlaps` with a likeness goes to the likeness route instead of `unrelated`.\n6. **The grounding teaches the method** (AM8, AM12): S6 as reworded (§ 9), the four parts of a pattern, Hesse's\n   formal analogy (one strict generalization both cases instance), and one contrast pair of a kinship and a metaphor; never cases from the graph.\n7. **Tested before it runs, by the review he asked the credit for**: pair 1 first, then known cases (the pendulum and\n   the circuit, the major transitions, Seth's weather simulation, the general and the doctor), each read one at a\n   time with him, his reasons becoming the rules.\n\n**What pair 1 would read** (corrected 2026-10-28, § 14). Neither claim is a case: both state the theory's\ngeneralization, A for any system and in terms of access, B for humans and in terms of experience. Gate `yes`; no\ncontent key, since A is wider in scope and B stronger in outcome, so neither entails the other; both render one\ntheory, which is reported speech with the theory as the party (ledger AL14, open). The analogy he saw runs from the\nhuman evidence to the language model the Anthropic text is about, through that theory; the keeper records it only\nwhere a claim about the model is in the pair.\n\n**Boundary with the analogy daemon.** The keeper writes a likeness only for the pair it is handed; finding pairs\nacross the graph stays the daemon's job, and its contract and the keeper's likeness field should be one form.\n\n## 11. The exact terms (2026-10-28, his question: \"'bearing' - do you think all those philosophers would advise to use that word? Make sure all wordings and prompting in the microtasks uses as technically correct/exact terms and phrasings as possible.\")\n\nNo. *Bearing* is an everyday word; the discipline's term is **relevance**: Carnap's *positive* and *negative\nrelevance* (e is relevant to h when granting e changes h's probability, *Logical Foundations of Probability*, 1950,\nch. VI), which is already this project's word for a link's degree (ruled Q50: relevance is the difference a reason\nmakes, Jeffrey's rule). So the gate and the degree are one concept at two grains: the gate asks whether the relevance\nis zero, the relevance reading says how large it is. By the ruling that a concept takes its discipline's name (BD11),\neach term below is proposed to him with where our use departs; the code's field names change only on his word, and\na changed prompt outdates the keeper's readings by design.\n\n| In the keeper's prompt today | The discipline's term | Source; where our use departs |\n|---|---|---|\n| *evidential bearing*, the gate `bearing` | **relevance**: relevant (positively or negatively) or irrelevant | Carnap 1950; identical concept |\n| *establish A* | **granting A** (conditioning on A) | Jeffrey; the relevance judge already says *granted* |\n| *say the same thing in other words* | **equivalent** (mutual entailment), judged by the consent test | MacCartney and Manning's *equivalence*; Carnap's *synonymy* is stricter (§ 4) |\n| *A contains B* | **A entails B and not conversely** (forward entailment) | MacCartney and Manning |\n| *opposes: cannot both be true* | **contraries** (cannot both be true, may both be false) and **contradictories** (exactly one true) | the traditional square; MacCartney's *alternation* and *negation*; rival answers (AM22) are contraries |\n| *overlaps* | **logically independent and relevant through a premise both assert** | MacCartney's *independence* names only the first half; the second is ours |\n| *chain*, *link* | **derivation**, **inference** | standard logic |\n| *rule* | **generalization** (strict, normic, statistical) | already ruled for the corpus (BD11); the prompt still says *rule* |\n| *induction from one case to the general claim* | **enumerative induction** | standard |\n| *wider rule that says cases of this kind go alike* | **projectibility premise** (Goodman) or **overhypothesis** (Kemp, Perfors and Tenenbaum) | ours keeps it as a written claim |\n| *likeness*, *the respect* | **analogy**, the **positive analogy** (the relevant similarities) | Keynes, Hesse |\n| *what is carried over* | the **projected property** (the hypothetical analogy) | Bartha |\n| *the strongest relevant difference* | the **negative analogy**, a **disanalogy** | Keynes, Hesse, Bartha |\n| *shared mechanism* | **material analogy with a causal tie** (Hesse's causal condition), or **formal analogy** (one strict generalization, one calculus, that both cases instance) | Hesse 1966; § 9 |\n\nThe same audit is owed for every microtask prompt (`standing_work.MICROTASKS`): *bearing* alone appears 48 times in\n`deliberus/` code and prompts, *rule* 483. Filed in `TODO.md`.\n\n## 12. What would lie between the two claims in a full graph (his question the same night)\n\nHis picture: each claim carries its whole tree, every reason and every piece of evidence cut to the floor and\nscrutinised, every argument anyone could muster in a hundred years written in, and a mind at a high level of\nabstraction noticing that features of the two trees are shared.\n\n**What would be there, for pair 1** (corrected 2026-10-28, § 14). One node for the theory's generalization, G:\n*In any system, information that enters the system's global workspace (a limited-capacity stage whose contents are broadcast to the system's specialised processes) is consciously accessible in that system.* A renders G almost whole (any system, *accessible*, entry to a small channel). B renders G restricted to humans,\nwith a stronger outcome (*experienced*, wide availability): a narrower generalization whose evidence is the\nneuroscience of ignition, masking and the attentional blink, and which supports G by induction across systems only\nthrough the premise that what holds of humans holds of the kind. Block's argument that experience overflows access\nattacks the step from B's *experienced* to G's *accessible*. The cases are elsewhere: individual experiments under\nB, and the language model the Anthropic text is about, under G, resting on a membership claim (the model is in\nG's reference class) with its own tree: functionalism and substrate independence, and Seth's weather simulation\nagainst the view that consciousness is like adding, not like being wet.\n\n**So in a full graph the analogy is not asserted; it is read.** It is the shape of the graph: two cases under one\ngeneralization (ledger AY2), a membership claim whose tree decides how much the human evidence carries to the model\n(the reference-class rulings, BC1 and BC3), and every disanalogy an attack in its place. *Alike in what matters* is\nthen computed, not judged: the relevance of the human case to the model case is the generalization's strength times\nthe membership's. A mind reading at a high level would find it the way structure-mapping finds it, as a matching of\ntwo subgraphs whose relations line up, and S6's test becomes checkable: do the matched relations include the causal\nones (a shared problem and family of solutions, or one strict generalization both instance), or only the shape. Sameness, containment and overlap\nbecome code too: one node, a node inside a tree, a shared premise node.\n\n**Against that, what we have is two sentences and one model call.** The keeper sees no tree, no evidence, no\nmembership claim; it is asked to say in one model call what the full graph would hold in thousands of nodes. Two\nconsequences for the design:\n\n1. **The keeper's output is a first, coarse draft of that structure, in the same shapes**, so the graph refines it\n   rather than replacing it with a different kind of object: its likeness field writes a generalization claim with\n   the two cases filed under it, a membership claim for the target, and its disanalogy as an attack on that\n   membership. Not a free-standing *analogous* label.\n2. **Until the structure is built under it, the drafted analogy moves no strength**: the generalization has no\n   counted cases and the membership no reasons, so by the rulings already made it reads as unexamined. The keeper\n   proposes where structure should go; the graph, filled, decides. As the graph fills, the keeper's judgment shrinks\n   to proposing what is not yet written, and what it once judged becomes something code reads.\n\n## 13. The microtask thinks like the mind of Deliberus (2026-10-28, his word: \"It should. It should basically \"think like the mind of Deliberus\" within the microtask. As far as we can take it.\")\n\nSaid of § 12's second consequence, that the keeper's likeness drafts the full graph's structure. Claude's reading,\nunruled. What the corpus already holds: AM1 has a microtask write *the piece of the subgraph its verdict rests on, which\nthe mind of Deliberus would then keep* ([three-minds-and-the-leap.md](three-minds-and-the-leap.md) § 4); AM12 teaches\na microtask the method and never the graph's contents; the floor-first order cuts first, finds sameness among the pieces,\nand links last ([one-challenge-one-home.md](one-challenge-one-home.md)); AG2 fills each level the cheapest sound way,\nfound before made; AG12 makes every sentence a machine writes a claim nobody asserts, moving no strength.\n\n**Taken as far as it goes, the pair stops being the unit.** The full graph around a claim belongs to the claim, not\nto a pair, so the mind of Deliberus would build it once per claim and read every pair off two such trees. In\nmicrotasks:\n\n1. **A tree per claim, once.** One microtask drafts, for one claim, what the full graph would hold under it: its\n   parts at the floor; the generalization it applies or is a case of, with its reference class; the kinds of evidence\n   that would settle it; the premises the literature contests. Each line is a draft claim (AG12), matched against the\n   graph before it is minted (AL16), so what already exists is found, not made again. Written once, it serves every\n   pair the claim is in (AM21), and every such draft is the graph filling.\n2. **The relation read off two trees, by code where it can.** One node is sameness; a part of one inside the other is\n   entailment; contraries between nodes are opposition; a shared premise both assert is overlap; two cases under one\n   generalization, with a membership claim for the second, is an analogy. The model answers only where the trees do\n   not settle it, and its answer points at nodes.\n3. **The broad stroke last.** The mind he described is careful and also sees *\"the broad strokes\"*; the vibe-check\n   comes after the trees, as a proposal pointing at nodes (*these two share a pattern at this level*), held for the\n   analogy daemon's second microtask and S6's test.\n\n**Pair 1, so drafted** (corrected 2026-10-28, § 14). A's tree: G's statement, its parts at the floor (entry to a\nlimited stage; broadcast; access), its kind a strict generalization as the theory states it. B's tree: G restricted\nto humans with *experience* for *access*, its evidence the neuroscience above, Block's attack on its outcome. Code\nthen reads: not one node; neither entails the other, since A is wider in scope and B stronger in outcome; no\ncontraries; one theory rendered twice. The analogy to the model appears only when the model's claim is in the pair.\n\n**Its limits, as the canon names them.** The drafted tree is the model's frame: what no source it read contains, it\ncannot draft (coherent absence), and it cannot stand outside the frame it drafts in (frame lock-in); both stay with\npeople. And a drafted premise is a draft: by AM12 the model is taught how to draft, never shown the graph as an\nexample, and by AX2 what is already written enters only where code acts on it.\n\n*The design carried further, as microtasks composed around one claim: [a-tree-per-claim.md](a-tree-per-claim.md).*\n\n## 14. Three questions on the design (2026-10-28)\n\nHis words: *\"I want to go towards being as stringent as possible in this so without making double model calls or\nany such copouts, how could we ensure the LLM does what it should each time? What do you mean \"once\"? As opposed\nto...?! What would the generalization here be, exactly, the \"rule\" connecting the claims?\"*\n\n**The generalization, exactly.** G: *In any system, information that enters the system's global workspace (a limited-capacity stage whose contents are broadcast to the system's specialised processes) is consciously accessible in that system.* Neither claim of pair 1 is a case under it, which corrects §§ 10, 12 and 13\nas first written: A states G almost whole; B states G restricted to humans, with *experienced* where G says\n*accessible*. So A is wider in scope and B stronger in outcome, and neither entails the other. What connects them is\nG itself, rendered twice; the analogy from human to model runs through G and needs a claim about the model.\n\n**\"Once\" means once per claim, as against once per pair.** The linking step reads each claim against about a dozen\npartners, and today every pair's call reads both claims from scratch, so each claim is analysed again in every pair\nit is in. A tree drafted for a claim is drafted one time and read by all its pairs, until the claim's text or the\nmicrotask's version changes.\n\n**Stringency in one call.** No call can be guaranteed right in substance. What one call can be made to guarantee,\nand how the rest is caught:\n\n1. **Code computes everything that can be computed.** With two trees written, sameness, entailment, opposition,\n   overlap and analogy are read by code; the model writes only the trees.\n2. **Wrong form is unrepresentable.** Structured output with closed enums, required slots, placeholders, and\n   sentences built by code from the model's parts, so nothing is copied twice (AM14, AM21).\n3. **The order of fields is the order of thinking**: what each claim asserts, then its parts, then its\n   generalization and reference class, then its evidence kinds and contested premises, the relation last.\n4. **Every rule is checked by code after the call**: each part stands alone, the generalization names neither case\n   and carries its kind's form, the membership claim exists for every case. A failure is stored refused with the\n   fault named, never retried, and the share refused is the microtask's measured error rate.\n5. **The prompt teaches the method in the discipline's terms** (§ 11, AM8, AM12), and changes only when the\n   right-first-time count rises on cases its examples do not show, read one at a time with him (the credit's evals).\n6. **Reuse instead of a second asking**: the same claim text under the same microtask version is never asked again.\n\n## How this was done\n\nRead whole for this note: `what-a-rule-rests-on.md`, `the-analogy-daemon.md`, `corpus-graph/what-a-move-is.md`\n(§ 1–5), `careful-and-broad-minds.md` (§ 3–4), `what-a-case-counts-for.md` (opening), ledger rows S5, S6, AD23,\nAD27, AY2, BA1, BC3; Carnap § 14 and MacCartney and Manning online. Not read: `reasons-and-cases.md`,\n`perceiving-possibility-and-the-analogy-organ.md`, `twelve-books-read-against-deliberus.md` § 13.4. No research\nhelpers were sent and no outside check was run; the tests are illustrations, and the one real case is pair 1.\n\n## Cross-references\n\n[one-inference-per-link.md](one-inference-per-link.md) § Held-out pair 1 · [what-a-rule-rests-on.md](what-a-rule-rests-on.md)\n(analogy's respect, Goodman to Gentner) · [the-analogy-daemon.md](the-analogy-daemon.md) ·\n[soft-canonical-clustering-and-reversible-merge-semantics.md](soft-canonical-clustering-and-reversible-merge-semantics.md)\n(the seven keys) · [sameness-merge-and-split-across-fields.md](sameness-merge-and-split-across-fields.md)\n\n## To listen (his question, 2026-10-27: a podcast on this, the logic core, LLMs and human minds)\n\nFound by search; not listened to by Claude, chosen from show notes and transcript excerpts.\n\n- **The Cognition Project, \"Mapping Relations: Dedre Gentner\"** (Tom Griffiths, 14 August 2026). Structure-mapping\n  in her own words; alignable differences; and her finding that memory retrieves by surface likeness while reasoning\n  maps by structure (the MAC/FAC model), which is the line between the embedder that finds candidate pairs and the\n  keeper that reads them.\n- **The Information Bottleneck, \"Reasoning Models and Planning\" with Subbarao Kambhampati** (29 April 2026). LLM-Modulo:\n  the model proposes, an outside verifier checks and sends criticism back; the frame of ledger AM16 and AM17.\n- **Night Science 83, \"How science is secretly driven by analogy\" with Melanie Mitchell** (16 February 2026).\n  Analogy in discovery, how it misleads, and whether language models grasp abstraction. **Heard and read against the\n  corpus 2026-11-05** ([tools-and-analogies-as-engines-of-discovery.md](tools-and-analogies-as-engines-of-discovery.md)): judging a connection, not making one, is the skill she names, and\n  six real analogies with known outcomes were run through S6's test as evidence for the rewording in § 9 (ledger Arc BH).- **Night Science 83, \"How science is secretly driven by analogy\" with Melanie Mitchell** (16 February 2026).\n  Analogy in discovery, how it misleads, and whether language models grasp abstraction.\n\nWhere the three differ (Claude's reading from their published work, not from the episodes): Gentner treats analogy as\naligning relational structures already represented, and holds that comparing cases teaches the abstraction; Mitchell,\nin Hofstadter's line, holds that the representation forms during the mapping (Chalmers, French and Hofstadter's 1992\ncritique of the Structure-Mapping Engine) and finds language models weak on analogies outside their training (Lewis\nand Mitchell 2024); Kambhampati treats a language model as approximate retrieval whose correctness only an outside\nverifier can supply. For the keeper: whether the term map of an analogy can be written before reading is the\nGentner–Mitchell question, and who checks it is Kambhampati's.\n\n## 15. Built 2026-11-06, and the gate question read again (his: *\"Is that the right root/gate question to tell, considering everything we've learned since it was put in there? Just curious. Read up.\"*)\n\n**Built, under his word of that night**, from § 10 in its dependency order: (1) the gate asks about relevance of any kind, by a generalization, an induction or a likeness with a mechanism, in Carnap's word; (2) the seven keys stay seven; (3) the likeness is written before the chain as its own object (`keeper.Likeness`: the two terms copied from the claims, the respect as a general sentence with `{x}`, what carries over, the strongest relevant difference, the mechanism as one of four: `shared_problem`, `strict_generalization`, `partial_pattern`, `none`); (4) code checks it as it checks a chain (`keeper.likeness_faults`) and sends it back, up to two rounds now (`MAX_ROUNDS_BACK`, ledger AM14 noted); (5) routing: a likeness with a mechanism and no key that fits is an **analogy candidate**, recorded on the pair's resemblance link (`analogy_candidate`, `likeness`) and never a link, listed and judged by `scripts/analogy_candidates.py`; a refused `overlaps` beside a likeness goes there too; `none` writes nothing and is a metaphor seed for later; (6) the grounding block carries S6 as reworded (`deliberus/grounding.py`); (7) the keeper's pairs are read with the hold lifted locally for the night, the candidates judged by the session and kept for his review. Step 5's second half, the relationship classifier offered the *analogy* scheme and the respect as the link's relevance claim, is not built: a candidate stays a candidate until a person says kinship.\n\n**The gate question, read again.** The gate still asks the root question the keys rest on: is either claim relevant to the other (Carnap's relevance, the concept the relevance reading later grades), which the keys then sort into same, contained, opposed or overlapping. What changed since 2026-10-18 is what counts as relevance: the first gate admitted only a chain the claims themselves assert, which read every likeness as *no* (§ 10, *why the keeper loses it today*); the rewording of S6 says an analogy through a shared mechanism is evidence, so the gate admits it and the keys, which are about content, say *no key fits*, which is the analogy candidate. Two alternatives were weighed and set aside: a root question about a shared subject (topic is not evidence, the keeper's own first lesson) and a root question about the chain's length (a count, which the refusal in `resolve` already handles). What the gate still cannot see is coherent absence, a bearing through a claim neither text holds; the workup (§ 2d of `a-tree-per-claim.md`) narrows it by handing the keeper what earlier passes wrote.\n\n### 15a. The first ten candidates, judged by the session (2026-11-06, 01:30, under his exemption of that night; his to overturn)\n\nThe keeper's sweeps over the night's new sources and the older graph left ten analogy candidates\n(`uv run python scripts/analogy_candidates.py`). Each was read against S6 as reworded (kinship is a\nshared mechanism named at the level where it works: a shared problem with a shared family of\nsolutions, or one strict generalization both cases instance; a shared shape with no mechanism is a\nmetaphor) and against the two rules the first two cases taught the code (one subject is no\nlikeness; a generalization and its case is no analogy). The verdicts and their reasons are on the\nresemblance links (`analogy_verdict`, `analogy_words`), never on the keeper's likeness, which stays\nas the machine wrote it.\n\n| Pair | Verdict | Why, in short |\n|---|---|---|\n| Madison's definition of tyranny ~ its application to a machine (hoel-flight-93) | neither | a rule and its case |\n| early AI-safety meetups funded DeepMind ~ founded OpenAI (hoel-flight-93) | neither | one subject, two claims about it |\n| small reductions in existential risk carry large expected value ~ a low-probability high-payoff poverty solution outweighs direct aid | kinship | one strict generalization, expected value as probability times payoff, instanced twice; the payoff's kind differs, the arithmetic does not |\n| the same ~ a positive Singularity has enormous expected value | kinship | the same rule; the keeper's `partial_pattern` is the whole rule |\n| the fanaticism rule itself ~ the poverty case | neither | a generalization and one of its cases |\n| distributed resilience reduces capture and monoculture risk ~ group diversity guards crowd accuracy against consensus-seeking | **kinship** | a shared problem (correlated failure under one dominant pull) with a shared family of solutions (independence and diversity of the parts): institutions and crowds, one mechanism |\n| existential-risk reasoning can make restrictions on present people look justified ~ the poverty case | kinship | both instance the fanaticism rule; one diagnoses what it licenses, the other applies it |\n| the same ~ the positive Singularity | kinship | the same |\n| the same ~ accept thirty years of poverty to prevent trillions of years of it | kinship | both trade present welfare for far-future value by expected value; the named difference (uncertain risk against a stipulated necessity) sits in the probability term, which is the premise an attacker denies |\n| AI governance should preserve human control through plural institutions ~ AI research should prioritise Friendliness | neither | a shared aim, not a mechanism: two prescriptions toward one goal, which the keeper's own difference shows can conflict |\n\n**What the ten teach.** Six of the eight read tonight share one generalization, that a tiny\nprobability times an astronomical payoff can dominate a certain modest benefit, which the graph\nalready holds as a claim (`claim_aig_fanaticism`). The right structure for such a cluster is not\nsix analogy links but one shared premise with each case filed under it, so an attack on the rule\nreaches every case once (the folded-routes note, ledger AX4; the case channel, Arc AY). How, and for the resilience kinship too, with the scrutiny as standing work: [the-analogy-daemon.md](the-analogy-daemon.md) § 10 (ledger BH5, BH6, proposed 2026-11-06). The one\ncross-area kinship (resilience against capture, diversity against consensus-seeking) is the kind\nthe daemon was built to find: a mechanism that one source states about institutions and another\nabout crowds, with the corpus's own *Correlated Errors* reading as the third instance. A candidate\nrule for the code, from the last pair: a respect of the form *{x} should aim at Y* names a goal,\nnot a mechanism, and is refused like one subject (unbuilt, unruled). Two of the mechanism labels\nthe keeper chose (`partial_pattern`) were the whole rule; the label is worth measuring once he has\nread these ten.\n\n\n*Read again 2026-11-10, before his first verdict (Y9).* The pair put to him first is the small-reductions claim against Pogge's low-probability poverty solution (strict generalization). Two rulings since it was put change what the case asks: under *a judged kinship becomes a generalization with its cases* (BH5, as amended) and *every case in the graph* (BI4, ruled 2026-11-08), both claims are themselves generalizations, so a kinship verdict mints the respect as a reference class with the two lying within it as narrower reference classes and no count until an observed intervention with an outcome is filed; and under *the difference names its role* (BL4, ruled and built 2026-11-09) the keeper's difference is routed by the role it touches, which this candidate carries unread, so his verdict is asked together with the role: a condition of the respect (attacks a membership), the carrying (attacks the transfer's relevance claim), or none.\n\n### 15b. The analogy hunt (2026-11-06, later the same night)\n\nThe keeper finds an analogy only where a resemblance link exists, and resemblance links come from cosine\nsimilarity, which measures vocabulary: two texts about different things, kin by mechanism, share few\nwords, so the pairs the daemon exists to find never reach the keeper's queue. The hunt\n(`deliberus/analogy_hunt.py`, `scripts/analogy_hunt.py A B --apply`) is the daemon's first pass in its\ncheapest form: one call reads up to eighty claims of each of two named sources and names at most six\npairs that instance one mechanism, with the respect as a general sentence with `{x}` and the strongest\ndifference; the instructions refuse the three faults the first ten candidates taught (a shared subject,\na shared aim, a rule with its own case). Each proposal becomes a resemblance link marked `hunt_proposed`\nwith no similarity figure, so the keeper's standing sweep reads it like any other pair and writes its\nown likeness; an unread hunt link is removed by `--undo`. Propose-only on purpose: models are poor\njudges of which connection is good (22 of 30 cross-field proposals failing,\n`tools-and-analogies-as-engines-of-discovery.md`), so the hunt's reading is the first of three, the\nkeeper's form check the second, a person's the third. Run that night over fourteen far pairs of\nsources; what each proposed, and what the keeper then made of it, is on the links and in the record\n(`scripts/standing_work.py report`, pass `analogy.hunt`).\n\n**The fourteen pairs' proposals, read by the session before the keeper read them** (the hunt's records are\non the links, `hunt_respect` and `hunt_difference`). Fifteen proposals from eleven of the fourteen\npairs; three pairs (fermentation against AI as normal technology, Epictetus against the shutdown\nessay, Stoicism against the dark forest) proposed nothing, which is the right answer where nothing\nshares a mechanism. Read against S6 as reworded, eight look like kinship: precaution's *benefit of the\ndoubt* for animals of uncertain sentience against *if you can't be sure you're creating a self-aware\nAI, stop*; the quasi-option value of irreversible decisions against *you do not get to learn from your\nmistakes, because you are dead* (option value under irreversibility); Bostrom's *risk trumps\nopportunity cost* against Hardin's *injustice is preferable to total ruin* (a recoverable loss against\nan irreversible one); Mercier's confirmation biases balancing out in discussion against a crowd's\ndiversity lowering its error (independence cancels correlated error, the same mechanism as the first\nten's one cross-area kinship); the introspection illusion behind the feeling of free will against a\nlanguage model's learned claims of consciousness (an appearance produced by a process that does not\ntrack the property); basic income's payments to the wealthy against the minimum wage's leakage to\nteenagers in affluent households (an untargeted instrument's waste); and Graeber's administrative\nexpansion against Narayanan's AI-control jobs (automation moving labour into control roles). Four look\nlike a shared shape with no mechanism: the Coase invariance failing against *neither containment nor\nresilience is always safer*; *free will needs ultimate causation* against *an AI needs grounding*;\nGoodhart's car salesman against social media's market failure; and firms avoiding wages for unneeded\nlabour against the same market failure. One (the argument from marginal cases against disunified\npeople still being conscious) shares an argument form, a counterexample to a proposed criterion,\nwhich is a scheme the graph already names rather than a likeness between contents. The keeper's own\nlikeness on each, and a person's verdict, are what count; this reading is one more witness.\n\n**The keeper's reading of the hunt's pairs (02:10).** The hunt had listed wholes, and 22 of the 29 claims\nin its fifteen pairs were wholes the splitter then cut, so the keeper, which reads pairs at the floor\n(AL9), could read none of them as proposed. Each proposal was moved to the nearest pair of the two\nwholes' live parts (by the embedder, one keeper call per hunt pair), and eight of the fifteen could be\nread that way; seven could not, since one side's parts could not be found by the query used, or a side\nwas held for a person. Of the eight, the keeper made one an analogy candidate (Graeber's administrative\nexpansion against Narayanan's AI-control jobs, `partial_pattern`) and found no mechanism or no bearing\nin the other seven, among them three the session had read as kinship (the introspection illusion against\nLaMDA's imitation, Mercier against the crowd, Hardin's coercion against bounded machine authority). Two\nreadings of the hunt's two strongest proposals (precaution's benefit of the doubt, the quasi-option\nvalue) are still owed, their pairs among the seven. The code change the night leaves: the hunt lists\nonly claims at the floor or unread by the splitter (`analogy_hunt._claims_of`), so its proposals land\nwhere the keeper reads. What the disagreement between the session's reading and the keeper's means is\nhis to judge, one pair at a time (Y9); the likeness the keeper wrote on each is on the link.\n\n**2026-11-09: the likeness read as Gentner's mapping, and the products read off it.** The four things a comparison produces (abstraction, contrast, inference-projection, rerepresentation; Gentner 2010) are each an element's relation to the keeper's likeness, and the eleven candidates above were read for them: nine shared a generalization the graph already held or had one side as the generalization, five differences touched nothing, four the carrying, one membership. Where a difference lands, with the literature and seven histories: [what-a-comparison-produces.md](what-a-comparison-produces.md) (ledger BL3 recut, BL4 proposed). ⚠ § 2b of `cases-and-generalizations-from-both-ends.md` had the alignable difference as *the first attack on the generalization*; the default landing is the transfer's relevance claim.\n\n**2026-11-12: the unit of analogy, the whole-vs-part filter review, and the two tiers of cases.** The filter `placed.get(cid) != CUT` admitted `UNSEEN` composite claims from freshly extracted sources before `cut_source_to_floor.py` ran. The invariant under AL9 is `placed.get(cid) == AT_FLOOR`, with cutting guaranteed before hunting. More deeply, comparing atomic floor leaves misses relational mechanisms (disjoint vocabularies, zero cosine similarity), while comparing uncut wholes produces unstable nodes. The unit of analogy is an argument/derivation constellation whose focal claim connects via `[:IN_CLASS {grounds: membership_claim}]` to the Reference Class, unifying Tier 1 observation bundles (`[:FILED_UNDER]`) and Tier 2 structural argument cases (`[:IN_CLASS]`) in bidirectional reflective equilibrium without new graph primitives (Arc BL, Arc BI, Arc BH; review in `.private/collab/analogy-discovery-relational-mapping-and-core-ontology-review-2026-11-12.md`).\n\n## From the Topic Index\n\n*The `CLAUDE.md` Topic Index entry that pointed here, word for word as it stood on 2026-11-15, before the index was trimmed to the warnings only it carries. The index keeps a short entry; the full text lives here.*\n\n- **\"Overlaps or analogy?\" / the keeper's likeness / a tree per claim / exact terms in microtask prompts** → `docs/research/overlaps-and-analogy.md` (`overlaps` relates what two claims say, an analogy relates two cases through a shared pattern; ruling S6 read again; what goes into the keeper; *relevance*, not *bearing*, and the discipline's term for each word in a microtask prompt) and `docs/research/a-tree-per-claim.md` (a workup drafted once per claim, relations read off two trees by code, handoffs written as standing-work conditions). Unruled; item 6 of the order of work in `TODO.md`, from 2 November with the keeper's held-out pairs. **Built 2026-11-06 under his exemption of that night:** the keeper's likeness beside the keys (terms, respect with `{x}`, carried over, difference, mechanism), checked by code; an analogy candidate on the pair's resemblance link, the first ten judged by the session, six kinship and four neither (`overlaps-and-analogy.md` § 15a); the analogy hunt as the daemon's first sweep, one call per two sources, propose-only, read by the keeper at the floor (§ 15b); each claim's workup handed to the keeper (`deliberus/workup.py`, a tree per claim in its first form). His reading of the ten, one at a time, is what settles any of it (BH4, a shared aim is no likeness, open).\n"}