{"path":"research/cases-and-generalizations-from-both-ends.md","content":"# A generalization and its cases, calibrated from both ends: reflective equilibrium in the graph and in how we work\n\n**Date**: 2026-11-06 · **Type**: an all-angles pass on a founder's question, written the morning after the\ncredit night, while his first analogy verdict (the resilience pair, kinship) was being recorded.\n**Status**: Claude's reading, with one build proposal (BI1, open) and two candidate resident lines, of which the project line is BLESSED (BI2, 2026-11-06, *\"Bless the line\"*) and waits for the sitting's write; the global sharpening is unruled.\nLedger: Arc BI (a generalization and its cases, calibrated from both ends), BI1 open; BH5 amended with a\nprovenance clause.\n\n**His words, verbatim** (2026-11-06, in the review of the first analogy candidate):\n\n> Idk if it matters I feel like all this about analogies and cases, generalizations/regularities and\n> exact/mechanistic understanding/explanation... it should be bidirectional, which seems to me sort of the\n> same reason what I have our workflow to go back and forth between cases/examples and\n> rulings/theory/decisions/regularities, back to the next case/example which brings forth some other part of\n> the theory/regularities/understanding, right, and so on, rinse and repeat.\n>\n> The mind of Deliberus should be dynamic and constantly evaluating (by way of all its standing work,\n> machine sweeps, as well as human scrutiny and contribution, and most certainly as a result of our\n> architectural and data/object model/ontological choices and designs) its contents - both from the\n> direction of cases and edge cases and from the direction of generalizing about the world and then zooming\n> in on the cases that define/argue for/against generalities/generalizations/rules/natural/physical \"laws\"\n> even, etc etc. You see what I mean? Just like analysis and attunement recur fractally throughout the many\n> layers and meta-layers in/throughout/about Deliberus, I think this is another fractal recursion and\n> eternal-ish asymptotic oscillatory calibration that we need to pay attention to continually. Read up on\n> everything the corpus has to offer relating to all aspects of this question and advise. Perhaps there are\n> some candidates, in the end, for principles/directives/convictions/theorems to have in our project (or\n> even in my global) CLAUDE.md?\n>\n> Can we also look back at what was said in the structure-mapping podcast I listened to a while ago, I\n> feel like it talked about some of this (or at least about what the most valuable\n> analogy-hunting/finding looks like and why), and maybe the recent walkthrough had a bit about it too\n\nAnd, a few minutes later: *\"Also read all tidbits/ideas from older walkthroughs that talk about\nanalogies/cases/trials etc, think there was at least one (keep in mind their ideas might partly be\noutdated/superseded, but still might be valuable to look back to, we'll see)\"*.\n\n**The discipline's name for what he describes** (BD11, a concept takes its discipline's name): **reflective\nequilibrium**, Goodman's for induction and Rawls's for ethics, named in the corpus since 26 September for one\ncase of it (what decomposing a value claim does, `philosophy-of-evidence.md` § 8.1), the mutual adjustment of generalizations and\njudgments about cases until they fit; its formal twins are hierarchical Bayesian inference (cases inform\nthe overhypothesis, the overhypothesis informs each case) and alternating optimization. Where our use\ndeparts: the equilibrium is never reached and is not meant to be (his *asymptotic*), the object is a\ngeneralization claim that is also a reference class with its members, and the same loop runs at two\nlevels, the graph's and our own rulings'. Cognitive science's name for the upward half is Gentner's\n*progressive alignment*; for the whole, the humanities say *the hermeneutic circle*.\n\n## 0. The answer first\n\n1. **He is right, and the corpus already holds most of it, in pieces.** Goodman's circle, Gentner and\n   Medina's *similarities without rules are empty, rules without similarities are blind*, his own ruling that\n   a case is read before a rate is proven and each reason becomes a candidate rule (Y9), the rulings on what\n   a case does to a generalization (Arc AY), and the canon pass's own flow (asserted, ruled, compressed,\n   resident) are the two directions, written down separately. What no note said is that they are **one\n   shape at two levels** (⚠ 2026-11-07: one shape, an analogy between processes whose correctness conditions\n   differ, never an identity; Codex's review, § 7) and that **a reading in one direction only is where the\n   longest-lived mistakes lived** (§ 4e).\n2. **The literature agrees on three things, two of which the evidence notes of 26 September already held**\n   (§ 2, § 3): the loop converges only where both moves raise one shared score on a bounded space, and\n   otherwise cycles or settles in many places (new here); what keeps it from being merely circular is always\n   an anchor outside it (Haack's crossword, in `philosophy-of-evidence.md` § 7.3); and convergence inside\n   one head gives no convergence between heads (the standard criticism, in\n   `testimony-experience-and-normative-evidence.md` § 6.1). The one test that works on real, unbounded cases is Lakatos's: a\n   generalization is progressing when a case it predicted is found, degenerating when it only accommodates\n   cases already known, **which needs the dated order** of the wording and the case.\n3. **The graph already has the anchors and the dates.** The record of machine judgments timestamps every\n   reading; a reference class's members carry who filed them and on what grounds; published counts enter\n   from outside the graph; his verdicts enter from outside the machine. What is missing is small and\n   buildable (§ 5, BI1): a reading at either end outdated by a change at the other, and a count that keeps\n   the cases that proposed a generalization apart from the cases found after it, so Lakatos's test is a\n   number anyone can read.\n4. **The two directions were not symmetric on one night** (§ 3.8, § 4c): the upward move over-produced (six\n   of ten analogy candidates were one generalization), the downward under-admitted (the membership gate said\n   no to 29 of 30), and minds err both ways from one cause, a salient case standing in for its class. ⚠ A\n   hypothesis from one night and two passes, not a rule (Codex, § 7): six candidates collapsing shows\n   duplication in that candidate set, and 29 refusals with one defensible miss shows a sample with few valid\n   members. The brake on minting and the generosity on a proposal queue (never on acceptance into a count)\n   stand on their own reasons.\n5. **Candidates for the always-loaded file**: one project line (§ 5.2, graded *demonstrated*), and for the\n   global file a sharpening of the case-review directive rather than a new one (§ 5.3, graded *predicted*).\n   Put to him one at a time.\n\n## How this was done\n\nThe six parts of the all-angles pass: (1) the question pinned above, his words whole; (2) the corpus\ninventoried by the corpus graph's search and by reading (§ 2 lists each file read whole this session);\n(3) one research helper (Claude Sonnet) on the literatures the corpus lacked, with a skip list of what it\nholds; its report kept verbatim in `cases-and-generalizations-from-both-ends/literature-report-2026-11-06.md`\nand cited with its read-status tags; (4) theory first, then five tests on cases from the live graph and from\nour own record, with arithmetic recomputed here; (4b) the recommendations set against the ruled rows they\ntouch, and the new ledger rows read by the outside reader; (5) recommendations in dependency order; (6) this\nnote, the ledger, the pointers, the chat. **Read whole this session**: `what-a-rule-rests-on.md` §§ 0, 3,\n5, 8, 13, 14.5, 14.6; `what-a-rule-rests-on/cognitive-science-sources-as-read-2026-10-08.md` §§ 2, 3 and\nits closing; `what-a-case-counts-for.md` §§ 0, 5, 7, 8, 9; `the-evals-loop-and-our-machine-passes.md` §§ 1,\n3b; `three-minds-and-the-leap.md` §§ 1, 5, 5a; `the-canon-distillation-pass.md` § 4f; `a-tree-per-claim.md`\n§§ 6, 7; `flywheels-in-building-deliberus.md` §§ 2, 3; `the-kind-of-mistake-we-make.md` § 3;\n`ranking-reasoning-quality-without-bedrock.md` § 4; `overlaps-and-analogy.md` §§ 9, 10, 15, To listen;\n`the-analogy-daemon.md` §§ 7, 10; `priors-as-standing-work.md` § 11; `deliberus/reference_classes.py`'s\nheader; the listening pieces of 8 October (both), 23 October (what a case counts for), 24 October (checking\nevery inference, the trial passages), 5 and 6 November; the Mitchell transcript's passages on good\nanalogies; and, after the corpus graph named them unread, `what-counts-as-evidence/philosophy-of-evidence.md`\n§§ 3.5, 7.3, 7.4, 8.1–8.3, `what-counts-as-evidence/testimony-experience-and-normative-evidence.md` §§ 6.1,\n6.8, `what-a-rule-rests-on/philosophy-sources-as-read-2026-10-08.md` §§ 7–10, 17–20 and its lists,\n`what-the-group-held-2011-2017.md` §§ 0, 3, 4, `claim-sameness-philosophical-readings.md` (its traditions and\nhypothesis sections), `what-the-walkthroughs-and-the-reviews-are-for.md` whole. **Not read**: the Gentner\nepisode (§ 2b).\n\n## 1. What is decided, and what is open\n\n**Ruled, and this note builds on it**: a person reads what a microtask produced, one case at a time, before anyone proves\na rate, each reason becoming a candidate rule tried on the cases not yet read, until a rule he blesses is\napplied by code or a microtask (Y9, 2026-10-19; global as *Case review starts one case at a time and ends in\na rule he can bless*); a case counts once, under the rule whose condition it meets, pooled by count in a\nchannel of its own (AY2), a rule's number is read from its count (AY3), every case moves a statistical\ngeneralization a little (AY1), what an explanation of a failed case does depends on the kind (AY4), a study\nis a bundle of cases and a moderator found in the same trials is a fit, not a test (AY5), a single case is\nread through the narrowest counted reference class (AY6); whoever writes a generalization asks whether its\nfailure calls for an explanation (AT2); a claim's prior comes from its reference class, pooled by code\n(BC1); a consideration counts once (BC8); standing work owes what a graph condition says, with the brake\n(R8); machine work built on machine work is capped by measurement (T1); kinship is a shared mechanism, named\nat the level where it works (S6 as reworded 2026-11-05).\n\n**Open, and this note proposes**: how a judged kinship becomes a generalization that is also a reference\nclass, with the analogy read from its members (BH5); the hunt as standing work for a generalization's cases\n(BH6); and, new here, that each end of the structure keys the other end's readings, and that the count tells\nproposing cases from cases found after (BI1, § 5.1).\n\n**The fresh question**: whether the two directions, and the two levels, are one thing worth a line of their\nown, and whether they are symmetric.\n\n## 2. What the corpus already holds\n\n| Angle | Where, read whole | What it holds |\n|---|---|---|\n| The circle itself | `what-a-rule-rests-on.md` § 3 (Goodman, Brewer, Levi, Mill) and § 5 (*the picture is a spiral*) | Rules are tested against the cases we accept and cases against the rules; each adjusted to the other, a virtuous circle (Goodman); the wider rule that fixes which likeness counts came from cases one level up, and no source finds a level where this stops |\n| The upward half, measured | `what-a-rule-rests-on/cognitive-science-sources-as-read-2026-10-08.md` §§ 2–3 | Comparing two cases makes the principle usable for about half of learners against one in five who read them apart (Gentner, Loewenstein and Thompson 2003); two analogs yield a schema where one does not (Gick and Holyoak 1983); SEQL forms rule-like generalizations by progressive alignment; retrieval runs on surface, soundness on structure (recall .78 against .44; soundness 4.40 against 2.80, Gentner, Rattermann and Forbus 1993) |\n| Our own practice is the loop | `what-a-rule-rests-on.md` § 13.6; `the-evals-loop-and-our-machine-passes.md` §§ 1, 3b; `three-minds-and-the-leap.md` § 5a | His connection of 8 October: *\"By reviewing cases you reflect on them and see/distill regularities/learnings. This was also mentioned in the evals walkthrough iirc\"*; the evals loop's first two steps are a rule drawn from compared cases; finding stops at saturation, proving at a sample fixed in advance, and proving before finding is paid twice; his gist of 19 October, one case at a time, each reason a candidate rule |\n| The downward half, ruled | `what-a-case-counts-for.md` §§ 5, 7, 9; `the-canon-distillation-pass.md` § 4f | What a single case does to each kind of generalization; the count's arithmetic; a moderator found in the same trials is a fit, not a test (Thompson and Higgins, in AY5); the proposed resident line *what a case does depends on the rule's kind*, put to him and not blessed |\n| The recursion and its bounds | `a-tree-per-claim.md` § 7 | Each node owes a workup, each relation may write a generalization that owes one; the recursion stops at the separability floor, the measured error rate (T1) and the brake (R8), and the bound moves down as passes are proven |\n| The loop at the project level | `flywheels-in-building-deliberus.md` §§ 2, 3 | Flywheel 1, ruling-to-canon (a position asserted, ruled, compressed, resident); flywheel 5, the answer key (his labelled cases become rules that decide the next case); flywheel 2, mistake-to-check |\n| What a reading in one direction only costs | `the-kind-of-mistake-we-make.md` § 3 | Every long-lived mistake was a formula or a word that met no case for months: attacks counted as supports (March to August), part-of links backwards (April to October), parts for reasons in the canonical example, `confirmed` read as *a person checked this* |\n| What a microtask is taught | ledger AM12 (ruled 2026-10-24) | The method, never the graph's contents: a pass shown twenty linked pairs copies how they look; told what *overlaps* means it can judge a pair it has never seen |\n| The analogy half of the circle | `overlaps-and-analogy.md` §§ 9, 15; `the-analogy-daemon.md` §§ 7, 10 | A kinship is two cases under one generalization; the keeper writes the respect from two cases (upward); the membership pass files further cases under it (downward); the daemon can write a likeness out and cannot settle one |\n| Where the circle has no bottom | `ranking-reasoning-quality-without-bedrock.md` § 4 | No non-circular ground for any reasoning norm (Alston, Boghossian); the live question is whether the circle is wide, self-correcting and discriminating; a graph can hold its own norms as attackable claims |\n| The circle already named, and its anchor | `what-counts-as-evidence/philosophy-of-evidence.md` §§ 3.5, 7.3, 7.4, 8.1; `what-counts-as-evidence/testimony-experience-and-normative-evidence.md` §§ 6.1, 6.8 | Reflective equilibrium named on 26 September as what decomposing a value claim does (mother claim and children adjusted until they cohere), with its two standing criticisms (it starts from existing judgments; it guarantees no convergence across people with different starting points); Haack's crossword: mutual support is warrant only with *independent security* and an observation anchor, and a coherent cluster from one source family is not thereby true (Bovens and Hartmann, Olsson); a considered judgment from inside a position is a datum for the equilibrium outsiders lack, which is where impact testimony enters the loop |\n| The sources behind the circle | `what-a-rule-rests-on/philosophy-sources-as-read-2026-10-08.md` §§ 7, 17 and its list (c) | Goodman's virtuous circle, and his reversal in the *Seven Strictures*: inductive practice may provide the canons of similarity (the upward half, stated in 1972); Brewer's analogy-warranting rule, abduced from examples and then tested against them by reflective adjustment; Sunstein's refusal to call analogy a crude search for reflective equilibrium; eleven sources under *rules and cases depend on each other* |\n| The same loop in the concept layer | `claim-sameness-philosophical-readings.md` (Wittgenstein, James); ledger AM7 | A sense is a generalization over uses and each use is a case: two claims are the same iff used the same way; a difference that makes no difference is no difference; a use is a vote on the meaning (AM7, ruled); a sense splits when its uses diverge downstream (proposed, unruled), which is the downward reading in that layer |\n| Read whole, bearing on this only in passing | `what-the-group-held-2011-2017.md`; `what-the-walkthroughs-and-the-reviews-are-for.md` § 7 | The group of 2011 to 2013 proposed no case-and-rule loop (its validator ladder and rationality rating are rejected paths); the twelfth piece's thought that *the machine writes the rule; a life supplies the case*, a counterexample from somebody's life being testimony and objection at once, stands, while its *agreement may sit in the middle* was withdrawn in `what-a-rule-rests-on.md` § 13.1 |\n\n### 2a. The older walkthroughs, read for their tidbits (his second ask)\n\nNine pieces speak of analogies, cases or trials. What each held, and whether it stands:\n\n- **8 October, *Eight proposals, pondered again*** (the thirteenth piece). Part two, *four ways to walk*: a\n  written-out step says which of four ways it walks, and the fourth, from cases to a rule, was missing from\n  the night before and is more than twice as common on the graph as the step from a case to a like case;\n  Mill's question (why one grape is enough and one rude taxi driver is not) answered by the wider rule\n  about the kind. **Stands** (AD23 ruled with the fourth shape). Part four, *no vehicles in the park*: the\n  cards (*would it still hold, if it were a rat?*) are not only how a rule's width is found but **how a rule\n  gets made: each card is a case, kept under the rule**; Keynes on what a new case is for (a different one,\n  in ways that turn out not to matter); Goodman's circle. **Stands and is this note's seed**; the cards as\n  stored cases are what Arc AY built. Part ten, *the friend who says this reminds me of*: the daemon can\n  answer half of *alike in what matters* (what the cases share), never the other half (whether it bears on\n  the conclusion), and hands cases, never a number; *we are reminded by looks, convinced by structure*; the\n  machine favours the comparisons everyone already makes. **Stands**, and the keeper's likeness field is\n  its build.\n- **8 October, *What our own notes already knew*** (the fourteenth). Part six, *likeness, and nobody's\n  face*: a machine's first sense of what resembles what is everyone's average and no one's judgment; a\n  second machine is a weak second opinion on likeness; people judge a precedent more alike when its outcome\n  suits them (Braman and Nelson), law students too; and the August note already had the shapes list and\n  knew it needed *none of these*. **Stands** (R7, a second model is not a second observation; the\n  conclusion-hidden rule for the likeness question).\n- **13 October, *Three days to the floor***: his question in it, *\"isn't analogy, likeness seen from the\n  outside, quite valuable still, shouldn't we try to optimize for detecting this based on a stringent, deep\n  reflection, in order to fill up the corpus graph as intelligently as possible?\"*, and his observation that\n  the separability floor for splitting is *horizontally analogous* to the floor of each inference in the\n  chain. **Stands**: the first became the hunt and the kinship reading; the second is the two floors AG14\n  and B2.\n- **23 October, *What a case counts for***: every case moves a statistical generalization; cases get a place\n  of their own, counted once; a study is a bundle of cases with its spread kept; the critic on moderators\n  (*give me enough trials and I'll find a moderator for every misfit*); **and part four's line that a like\n  case filed under a shared rule is analogy made countable**, which is BH5 said three weeks earlier.\n  **Stands**; all ruled that day (Arc AY).\n- **24 October, *Checking every inference***: the floor trial's thirteen gaps, eight refused by the form\n  check; his *\"method, not repetition, without clarification, the answer is not liable to change much\"*; a\n  pass is taught the method (AM12); the correction that two trial scripts were not *code checks every\n  kind*. **Stands**; the correction is this note's § 4e in miniature (a claim about a rule made before the\n  case was looked at).\n- **5 November, *Judging the connections***: judging a connection, not making one, is the skill (Mitchell);\n  far analogies explain more often than they generate (Dunbar, two of ninety-nine); the six outcome-known\n  analogies through two judges, with the hits that cost something (Kelvin, the ratchet). **Stands**; the\n  *law* correction is in its header.\n- **Superseded in them**: the thirteenth piece's part eight said people meet at the case and write the\n  principle afterwards (withdrawn in `what-a-rule-rests-on.md` § 13.1: no source measured an order, and in\n  large publics agreement on a principle's words comes more easily than at a case); the 24 October piece's\n  advice to say *reference class* for a claim's kind, against the other session's *kind* (resolved 2026-10-25\n  for *reference class*, BD11); the 22 October piece's *premises that count only together* (reworded, Q49).\n\n### 2b. The structure-mapping podcast\n\nTwo podcasts are in view. **Night Science 83 with Melanie Mitchell** (16 February 2026, transcribed and read\nwhole for the tools note) is the one he brought on 5 November; its answer to *what the most valuable\nanalogy-finding looks like*, in her words: *\"what makes an analogy compelling is that it's not that it's an\nobvious similarity, but that it makes you realize something new about a system. Or it helps you explain\nsomething that you didn't understand\"*; *\"there's no sort of logical reason why any of those answers are\nworse than any others. But we humans have a very strong kind of aesthetic sense\"*; and *\"one of the skills\nthat really creative and successful scientists have is the ability to not only make such connections, but to\njudge that they are going to be good and useful ones\"*. **The Cognition Project's episode with Dedre\nGentner** (*Mapping Relations*, Tom Griffiths, listed 14 August 2026 in `overlaps-and-analogy.md` § To\nlisten from show notes) is the one whose title is structure mapping; no transcript of it is indexed on the\nweb, and Claude has not heard it. **Its description, from the show's public feed** (read 2026-11-06 evening at\n`feeds.acast.com/public/shows/6974202f7396c6aa98140159`; published 14 August 2026, 41 minutes 33 seconds): an\noral-history interview, Griffiths with Gentner, on how *\"her research has helped us understand how people form\nanalogies, and the broader importance of relations in human cognition\"*, her path into cognitive science (with\nDonald Norman) and an eyewitness account of the first cognitive science conference. So the episode is mostly\nbiography; the substance he remembers is most likely her published work, which the corpus has read.\nHer published answer to the same question, read on 8 October (Gentner and Medina 1998 whole but its experiments, the 2003 chapter searched, the 1993 retrieval paper and the 2003 negotiation study in part; the 1983 structure-mapping paper itself only in passages of an OCR scan, the 1989 Structure-Mapping Engine paper and Gentner and Markman 1997 not read beyond a critique and an abstract, per `what-a-rule-rests-on/cognitive-science-sources-as-read-2026-10-08.md` § 2 and its Part C): the valuable likeness is **relational\nand systematic** (higher-order relations tying lower ones), **found by comparison** (two cases side by side,\nnot one), and **judged by structure while retrieved by surface**, so a finder that runs on wording brings\nback look-alikes and the judging has to be a second stage. The audio is public in the same feed (one mp3, 41 minutes); downloading and transcribing it with the\n`transkribera` skill is about an hour on the Mac, on his word.\n\n**Her account applied to the mind of Deliberus** (his question, 2026-11-06 evening: *\"But if we apply that\nmind-framing to the mind of Deliberus? (Which I think is a habit that we should come back to...)\"*; the habit is\nledger AM3's open line on the mind of Deliberus, `three-minds-and-the-leap.md`). Read as one mind, the departure\nnamed above dissolves: the graph is its long-term memory, and each of her findings names one of its organs. Retrieval\nby surface, judgment by structure (MAC/FAC) is the embedder's shortlist and the hunt against the keeper's reading. A\nrule from two cases, almost none from one, is the hunt owing a second case before a kinship is minted (BH5, a judged\nkinship becomes a generalization with its cases). Progressive alignment is the membership pass filing members and the\nrespect reworded as they arrive (BI1, both ends key each other's readings). Alignable differences, the differences one\nsees only once two cases are aligned, are the keeper's `difference` field and the first attack on the generalization\n(AD27) (⚠ 2026-11-09: on the transfer's relevance claim by default, never the generalization, which is reached only by a counterexample case or a rival case; `what-a-comparison-produces.md` § 0). A relational label unlocking inert knowledge (Jamrozik and Gentner 2020) is the generalization's wording used\nas the hunt's probe, which is how the hunt already reads a source against one sentence. Where this mind differs from\nthe one she measured: it need not retrieve by surface (it can read by structure at a cost), it does not forget a\ncomparison, and every comparison it makes is public and attackable; and its measured fault, the upward move\nover-producing (§ 4c), is her retrieval stage running without her judgment stage. Not filed as a ledger row; it belongs\nto AM3 when that line is taken up.\n\n## 3. What the literature adds (the helper's report, read whole; tags are its read status)\n\n**No source treats the oscillation as one object.** Five literatures treat parts of it and agree on three\nthings; the second and third were in the corpus already, in the evidence notes of 26 September (§ 2), and\nthe first, with Lakatos's dated order, is what the report adds.\n\n1. **Convergence is a property of the score, not of the oscillation.** The formal model of reflective\n   equilibrium (Beisbart, Betz and Brun 2021, *Ergo*, read in part) alternates two rules, revise the theory\n   and revise the commitments, with a score that never decreases, so on a finite space it always ends at a\n   fixed point; but a fixed point is a global optimum only for extreme weights, which is ignoring one side.\n   The same holds of expectation maximization (Wu 1983, read in part): the likelihood never decreases, so\n   it converges to something, and nothing guarantees that something is the best thing; the stopping rule in\n   practice is a tolerance. Rawls himself (1971: 49–50, via the Stanford Encyclopedia) calls the full\n   equilibrium an ideal that comparison moves us toward without reaching, and concedes there may be no\n   unique equilibrium even for one person.\n2. **What keeps the loop from being merely circular is always something outside it**: Daniels's\n   independence constraint (the background theories may not recycle the judgments that support the\n   principles, or wide equilibrium gains nothing over narrow), Elgin's initially tenable commitments that\n   are not given up without reason, the top-level prior in a hierarchical Bayesian model (Kemp, Perfors and\n   Tenenbaum 2007: *\"we do not claim that overhypotheses can be generated out of thin air\"*), Lakatos's\n   protected hard core, and in the history of science **independent detection**: Pauli's neutrino (posited\n   4 December 1930 to save energy conservation, detected 14 June 1956) and Le Verrier's Vulcan (posited for\n   Mercury's residual, never found, the residual explained by Einstein's paper of 25 November 1915) were the\n   same move, posit a new entity, and nothing known at the time told them apart; detection did.\n3. **Convergence inside one head gives no convergence between heads.** Baumgaertner (*Ergo*, read in\n   part): even when every agent reaches individual equilibrium it is *\"surprisingly difficult\"* for a\n   community to converge on a few equilibria, and more intuitions mean more choice points; Kelly and\n   McGrath 2010 (abstract and quoted passages): a good method should bring impeccable inquirers to converge\n   over time, and reflective equilibrium does not guarantee it; Lewis's passage, which they quote, says two\n   people may both reach equilibrium impeccably while one is wrong about the facts. Singer (2005: 349)\n   calls the method *\"close to vacuous\"* because it absorbs criticism as internal revision; Schwitzgebel and\n   Cushman 2015 found order and framing effects in professional philosophers, expertise not reducing them.\n\n**The one test that works on real cases is Lakatos's**, and it needs dates. A research programme is\nprogressive when a new theory has excess content and some of it is corroborated, degenerating when it only\naccommodates what everyone already knew (*Falsification and the Methodology of Scientific Research\nProgrammes*: 34, via the Stanford Encyclopedia). In *Proofs and Refutations* a counterexample is global or\nlocal or both, a proof splits the conjecture into lemmas so a refutation targets a lemma, and incorporating\nthe refuted lemma yields proof-generated definitions. The real Euler record: the formula offered in 1750 as\n*\"founded on analogy only\"* from several kinds of solids (Lhuilier's gloss, read); Lhuilier in 1812–13\n(*Annales de mathématiques* 3: 169–189, read) listing three kinds of exception (a cavity, a through-opening,\nannular faces); Kepler's star polyhedra (recognised about 1619, rediscovered by Poinsot 1809, Cauchy proving\nthe list complete 1813, Cayley's density-corrected formula 1859), two of which Schläfli refused to call\npolyhedra at all, which is real monster-barring. About a century of work by many hands, no final\nequilibrium, and the object itself changed (the birth of topology).\n\n**The formal twin says the directions are one inference and differ in speed.** In a hierarchical Bayesian\nmodel the upward and downward readings are conditionals of one joint distribution. Goodman, Ullman and\nTenenbaum 2011 (results read in part) found the abstract theory learned before most specific structures, and\n**breadth beats depth**: with five systems learning was very slow, with five hundred it caught up after about\nthree samples per system, the *blessing of abstraction*. Kuhn's Postscript (read in part): exemplars, the\nworked problems, come before and teach the symbolic generalizations, a developmental asymmetry in the other\ndirection.\n\n**Psychology: errors run both ways, from one cause.** Tversky and Kahneman 1971 (abstract and first page):\n84 psychologists gave a median 0.85 for the chance that a result at z = 2.23 replicates in ten more subjects,\nwhere 0.48 would be reasonable. Hamill, Wilson and Nisbett 1980 (abstract): a single vivid case shifted\nattitudes toward a whole population even when subjects were told it was atypical. Nisbett and colleagues\n1987 (abstract): brief training in inferential rules improves everyday reasoning. The helper's own reading,\nnot a finding: a salient case stands in for its class and the class for the case, so the two errors share a\ncause, and the corrective differs by direction (abstract rules can be trained; correcting the upward error\nneeds the rule plus the full distribution).\n\n**Mendeleev's record is the pattern at full length** (Stewart 2018, read in part; Scerri 2019 and Scerri and\nWorrall 2001, abstracts): three complete predictions found (gallium 1875, scandium 1879, germanium 1886),\nabout half of his predictions right at best, accommodation (argon, first resisted as triatomic nitrogen, then\ngiven a new group in 1900) mattering as much as prediction, and tellurium placed before iodine against the\nweights, right for a reason (atomic number, Moseley 1913) nobody had yet.\n\n**What the report says would most change his picture**, in its words: convergence is a property of the\nscore, not of the oscillation, and Lakatos's test needs a dated order, so the graph would need to record\nwhen a case was filed relative to the generalization's wording; the shared version is harder than the\nprivate one, so a shared graph should record the choice points themselves (which case was barred, which\nexception adopted, which entity posited), not promise convergence; and every loop that works has an outside\nanchor, and breadth helps more than depth, so the graph should prefer cases from different sources over more\ncases from the same one.\n\n## 4. Theory first, then five tests on real cases\n\n### 4a. The theory, in the project's terms\n\n**Two levels, one shape.** At the level of the graph, the object is a generalization claim that is also a\nreference class (a `:Claim` with `claim_kind: 'reference_class'`, its population and outcome in\n`rc_population` and `rc_outcome`; its members reach it by `IN_CLASS {by, grounds}`, its counted bundles by\n`FILED_UNDER`, a narrower one by `WITHIN`). The **upward reading** writes the generalization from cases: the\nkeeper's respect compared out of two claims (AM11's likeness, BH5), the hunt's proposal, the scope reading\nof what kind it is (AT11, AT12), a count entered from a published study (AY5). The **downward reading**\nreads each case against it: the membership pass (BD2), the member's prior (BC1), the exception question\nwhen a member fails (AT2, AY4), the narrower reference class when a statistical one shows spread (AY4). At\nthe level of our work, the object is a ruling and the cases it meets: a case he judges ends in a candidate\nrule tried on the cases not yet read (Y9); a rule, once written, meets its next case in a walkthrough's\nquestion to carry, in the outside reader's comparison, in the wrong-by-default grading of a resident line\n(a line earns residency only when a session lacking it would assert its opposite, which is a rule tried on a\ncase). The flywheel note already names the project-level loop twice (ruling-to-canon; the answer key).\n\n**The anchors, which the literature says every working loop needs**, are in place at both levels: the\nrecord of machine judgments (every reading timestamped, versioned, its inputs named), published counts from\noutside the graph, his verdicts from outside the machine, the separability floor and the two floors of an\ninference (B2, AG14), the measured error rate that caps machine work on machine work (T1), the brake (R8),\nand for a resident line the stranger (the headline tests). Daniels's independence constraint has a project\nform: a pass is taught the method and never the graph's contents (AM12), so the generalizations the graph\nholds cannot recycle themselves into the instructions that read them; and Haack's *independent security*\n(a reason's warrant apart from the belief it supports) is what the provenance kind for observation-anchored\nclaims would compute (`philosophy-of-evidence.md` § 7.4, proposed there). **The attunement pole enters the\nloop at the case end**: a considered judgment from inside a position is a datum for the equilibrium that\noutsiders lack (the testimony note § 6.8), and a counterexample from somebody's life is testimony and\nobjection at once (the twelfth piece), so impact testimony is a case filed against a generalization, never a\nweight on it.\n\n**The stopping and pacing rules** are the two already named and one the literature adds: finding stops\nwhen new cases teach no new kind (saturation, Y9b, open as a row); proving stops at a sample fixed in\nadvance; and **a count is never read from the cases that proposed the generalization** (AY5's fit-not-test,\nLakatos's dated order), which is the clause § 5.1 makes a readable number.\n\n**What is new in the picture**: not the halves, which are ruled, but that they share one shape (⚠ an\narchitectural analogy, not an identity, § 7); that a reading at either end is keyed to the other end's version; and that the project's own method is the same\nloop one level up, with the same anchors (his verdicts, the stranger) and the same failure (a rule that met\nno case).\n\n### 4b. Test one, from the live graph: the resilience pair, and what two cases settle\n\nHis first verdict this morning (kinship) on the pair *where distributed resilience exists, it reduces capture\nrisk, monoculture risk, and the risk of systemic takeover* (`claim_aig_dist_reduces`) against *ensuring group\ndiversity is intended to mitigate the negative impact that consensus-seeking behavior has on the accuracy of\na crowd's collective judgment* (`claim_fe274dce8a28`). The upward reading is done: the keeper wrote the\nrespect (*{x} spreads function or perspective across many independent parts so that a single shared failure\nor dominant pull cannot corrupt the whole*). Under BH5 it is minted as a normic generalization and the two\nclaims filed as members. **What the count then says, by the arithmetic of `what-a-case-counts-for.md` § 5**\n(a flat prior, the normic band read as a rate of 0.95 or more, so the mass above 0.95 of a Beta(1 + k, 1)\nis 1 − 0.95^(k+1)): two conforming members of two give the generalization 0.143; four of four (model panels\nand discussion joining) 0.226; ten of ten 0.431, which is the note's own figure for ten open Tuesdays. So\nthe structure carries the doubt his verdict leaves: a kinship confirmed on two cases is a generalization\nthat two cases barely move, and the posterior says so instead of a label saying *analogous*. The test passed\non the arithmetic; the build (BH5) is what makes it true on the graph.\n\n⚠ **Withdrawn 2026-11-07 as a strength for this generalization** (Codex's review, § 7): the two claims are\ngeneral claims, one asserting a reduced risk and one an *intended* benefit, and neither supplies an eligible\nsystem, an observed outcome and the conditions it was sampled under. Under BH5 as amended they lie *within* the\nminted generalization as narrower reference classes (`WITHIN`), and the count comes from bundles of observed\ncases, of which there are none yet. The arithmetic holds for the model it uses (two successes, one shared rate,\na flat prior: 0.143 is the mass above 0.95; the next-case predictive probability in the same model is 0.75);\nthat model has not been justified for these sentences, and a system that reduces a risk and still fails is no\ncounterexample to a risk-reduction claim. What the pair is: a proposed relational mapping, with the transfer,\nthe role correspondences and the decisive difference still to be written (§ 7).\n\n### 4c. Test two, from last night's record: the two directions err differently\n\nUpward, the keeper's analogy candidates: of the first ten, six instanced one generalization already in the\ngraph (`claim_aig_fanaticism`, a tiny probability times an astronomical payoff can dominate), two were a rule\nwith its own case or one subject, one was a shared aim; one was the cross-field kinship. The upward move\n**over-produced**: it offered as six likenesses what is one reference class with five members, and twice\noffered a generalization and its own case as kin. Downward, the membership pass over the Dunning-Kruger\ntext: the gate said *no* to 29 of 30 claims and refused the one it admitted, rightly; read one by one, at\nmost one of the 29 was a defensible miss. The downward move **under-admitted**, which is the cheaper error,\nsince a wrong prior moves every reader's number. One night, two passes, small counts; the asymmetry is a\nfinding to watch, not a rate. The helper's psychology is the same shape from the other side: minds\nover-generalize from a salient case (the law of small numbers) and under-apply a rule to a case (base-rate\nneglect), both from a case standing in for its class. **Consequence for the design**: the brake belongs on\nminting (BH5 mints only on a person's verdict; R8's four calls) and the generosity on filing (the membership\npass's two wordings, erring toward *no*, with the member's prior the stake), which is how the night's code\nalready leans.\n\n### 4d. Test three, from last night's faults: a case confused with a claim about its class\n\nTwo of the first nine filings put claims *about* the population under it as members (*it is really easy for\nfindings in psychology to appear real when they are not*; *the evidence for the crowd-within effect is\nmixed*). That is the downward reading mistaking the class's own counterexample, a generalization about the\nreference class, for a case of it. One sentence added to both candidate wordings (*a claim about the\npopulation as a whole is not one of its cases*) and a re-read refused both; a third of the same kind, with\nthe effect named, was still filed and waits for him (`priors-as-standing-work.md` § 11). The loop ran in\nboth directions inside one night: cases taught the rule, the rule re-read the cases. And the fault is the\none the literature names as the whole oscillation's hazard, a case and its class exchanging places.\n\n### 4e. Test four, from our own record: what a reading in one direction only costs\n\nEvery mistake in `the-kind-of-mistake-we-make.md` § 3 is a generalization that met no case for months. The\nbadge formula counted 1,567 attacks as supports from March to August: a formula read against no single\nclaim's badge. Part-of links were stored backwards from April to October behind a test that checked only\nthat two ids swapped: a rule tested on no real link. Dr Smith's credentials stood as parts in the canonical\nexample for six months: a drawing nobody ran a case through. The field `confirmed` read as *a person\nchecked this* for seven months: a word never asked of one record. **And this morning, in miniature**: the\nmethod's rule that *a ceiling he names is a ceiling* was applied to a 23-minute piece asked for at fifteen,\na rewrite begun and a second render started, before the case was asked back; his correction (*\"the\n\"ceiling\" is just a rough estimate of how long a walk I wanna take, but it doesn't justify spending\ntime/tokens/credits rewording and rerendering\"*) reshaped the rule within the hour. The 24 October piece\ncarries the same shape from the other direction: *\"deterministic code checks the form of every kind\"* said\nof two trial scripts run on cases the writer had made, a claim about a rule made before any real case. In\neach, the loop was running in one direction only, and the error lived exactly as long as the other\ndirection stayed silent.\n\n### 4f. Test five, with ground truth: the BCG trials and Lakatos's dated order\n\n`what-a-case-counts-for.md` § 7 already holds the thirteen BCG trials: one effect leaves 92% of the variation\nbetween trials unexplained; latitude, named as a moderator, carries about three quarters of it, and the\nnarrower reference class gives a child in Madras a risk ratio of 0.88 where the broad one says 0.49. AY5\nrules the guard: a moderator found by searching the same trials is a fit, kept only if named in advance or\nholding on bundles not used to find it. **Lakatos's progressive test is the same guard with dates**: a\ngeneralization is corroborated by a case found after its wording was fixed, and only accommodates a case it\nwas fitted to. The graph can tell these apart with no new judgment, because the record timestamps every\nreading and a minted generalization carries the pair it came from: the cases that proposed it are known, and\nevery case filed after can be told from them. ⚠ 2026-11-07: told apart, not thereby a test (Codex, § 7): a case\nfiled later may have been known, selected for agreeing, copied from the same study, or used while the\ngeneralization was revised; dates audit the order and cannot settle the role. Today nothing reports the two\napart; § 5.1 does, with the role recorded beside the date.\n\n## 4b. Set against what is settled\n\nExtends Y9 (the loop he ruled for his own reading, now named as the same loop the graph runs), AY2, AY3 and\nAY5 (the count, and the fit-not-test guard, given dates), AT2 (the exception question as the downward\nreading's trigger), BD2 (the membership pass as the standing downward reading), BH5 and BH6 (the structure\nand the finder), R8 (what a pass owes: a condition on either end), T1 (the depth cap as one of the anchors),\nAM12 (the method taught, never the contents, as Daniels's independence constraint). Consistent with S6, BC1,\nBC8, B2, AG14, the horizon ruling (read, never set). Reverses nothing. The outside reader's lines on BI1 are\nin its row.\n\n**Assumptions, weakest first.** That the asymmetry of § 4c is general (one night, two passes: weak). That\nLakatos's dated order maps onto *proposed by* against *filed after* (strong where a generalization is minted\nfrom a pair; weaker for one entered from a published count, whose proposing cases are the study's own). That\nthe project-level loop and the graph-level loop are the same shape and not an analogy (the anchors differ in\nkind, his verdicts against published counts; the test is whether a mistake at one level has the form of a\nmistake at the other, and § 4e says it does).\n\n## 5. Recommendations, in dependency order\n\n1. **BI1 (ledger, open): both ends key each other's readings, and the count tells proposing cases from cases\n   found after.** A member's reading (its membership, its prior) is outdated when the reference class's\n   wording, kind or band changes, not only when the pass's instructions change; a reference class's count is\n   recomputed when a member's text, outcome or standing changes (a member cut, superseded or refuted drops\n   out). And the count reports two figures: the cases the generalization was minted or entered from, and the\n   cases filed after its wording was fixed, so that a generalization corroborated by cases it never saw reads\n   apart from one that only accommodates the cases it came from (Lakatos's test, AY5's guard, as a number).\n   ⚠ Narrowed 2026-11-07 (Codex, § 7): the two figures are an audit pointer, never a verdict; a case carries\n   its role (used in formulation or tuning · evaluated under a fixed version without its outcome choosing it ·\n   exposure unknown), and a member that failed stays in the count as a failure while it remains eligible: only\n   a refuted eligibility or a refuted observation drops it, never a refuted outcome.\n   Build: a `generalization_version` on the reference class, carried on each `IN_CLASS` link at the time of\n   filing, compared by `standing_work` when it computes what the membership pass owes; `minted_from` and\n   `minted_at` on a generalization minted from a kinship (BH5, amended today); `class_count` reporting\n   `proposing` and `after` beside the total. No model call in any of it.\n2. **A resident line for the project (the Principles), graded demonstrated; BLESSED 2026-11-06 (*\"Bless the\n   line\"*, ledger BI2), queued for the sitting's one write (BD29) in `the-canon-distillation-pass.md` § 4k.** Put\n   to him in chat after this note: **A generalization and its cases calibrate each other, in the graph and in how we work**:\n   a case is read against the generalization it falls under and the generalization against the cases filed\n   under it, each reading keyed to the other's version; a rule of ours is tried on the next real case before\n   it is called settled, and a case we judge ends in a candidate rule (Y9); a reading in one direction only\n   is where the longest-lived mistakes lived. Demonstrated wrong by default: the badge formula met no case\n   from March to August, and this morning a rule on a piece's length was applied before the case was asked\n   back (§ 4e). Headline checks: over-broad, no (it forbids nothing we do; it names a reading we skip);\n   stranger, advisory, *calibrate* is a word readers bring their own sense to, so the body says *read\n   against*; collision, *case* is BD9's word (one instance under a generalization) and is used so.\n3. **For the global file, a sharpening and not a new directive, graded predicted.** The global corpus holds\n   both halves apart: *Case review starts one case at a time and ends in a rule he can bless* (the upward\n   half), *A guard that never fires looks exactly like a guard that works* and *A fixture is copied from\n   reality* (the downward half, for code and test data). What none says is that a rule of conduct, once\n   written, is tried on the next real case it meets, and that a rule that has met no case since it was\n   written is a hypothesis to be said so. Proposed as a clause on the case-review directive rather than a\n   line of its own, since the fourth gate (marginal value over the corpus) finds the pieces present; his\n   call, after the project line.\n4. **The provenance he asked about** (*\"The record would still show when/what analogy was proposed\n   originally, whether from a microtask or human contribution, I presume/hope\"*): yes today for the proposal\n   and the judging, and BH5 is amended so the minted generalization keeps it. Today: the hunt's proposal\n   sits on the resemblance link (`hunt_proposed`, `hunt_respect`, `hunt_difference`, `hunt_at`,\n   `hunt_sources`); the keeper's likeness and the names of its readings sit beside it (`likeness`,\n   `readings`); his verdict with `analogy_judged_by` and `analogy_judged_at`; every model call in the record\n   of machine judgments with its model, version and inputs; a person's seed, in the human-seeded channel's\n   design, attributed to the seeder. Added to BH5: the minted generalization carries `minted_from` (the two\n   claim ids and the judgment's uid), `minted_at` and `by` (the person who judged), so the generalization\n   says which pair proposed it and who confirmed it, and the analogy read from it can show its origin.\n5. **The Gentner episode**: the feed carries the audio (§ 2b); transcribe and read it on his word, since its description says it is mostly her path into the field\n   (§ 2b); until then her papers stand in for it.\n6. **Watch the asymmetry** (§ 4c) as the night's passes run again: a count of candidates that turn out to be\n   one generalization with cases, against a count of members the gate refused that a person later filed.\n   No build; the record already holds both.\n\n**Alternatives set aside**: a new conviction line (*nothing is permanently settled*): the convictions\n*Nothing is permanently atomic* and *Nothing is permanently terminal* already carry the openness, and a\nthird would restate them for generalizations; a separate pass that re-reads generalizations against their\ncases on a schedule: the membership pass and the count already are that pass once BI1 keys them to each\nother; a formal convergence criterion for the graph: the literature's only proofs rest on one shared score on\na bounded space, which a public graph of contested claims is not, so the honest instrument is the dated\ncount and the recorded choice points, not a promise.\n\n**What only he decides**: BI1; whether the project line enters the Principles; whether the global\ndirective takes the clause; whether the Gentner episode is worth an hour of the Mac.\n\n## 6. The same loop at every level: where each finding lands (2026-11-06 evening, Claude's reading for the ten-minute walkthrough, unruled)\n\nHis ask: *\"some more musings on how we might translate all these findings into Deliberus at different levels of\nabstraction/meta\"*. The loop is one shape, and it runs at five levels. At each the same three things are owed: an\nanchor outside the loop, the dated order of generalization and case (Lakatos's test), and a watch on the asymmetry\n(the upward move over-produces, the downward under-admits).\n\n| Level | The generalization | Its cases | Where the finding lands | State |\n|---|---|---|---|---|\n| The claim graph | a reference class (a `:Claim` with `claim_kind: 'reference_class'`) | its members (`IN_CLASS`), its counted bundles (`FILED_UNDER`) | BI1: a `generalization_version` on each `IN_CLASS` link, a recount trigger, `class_count` reporting *proposing* and *after* | proposed |\n| The microtasks | a pass's instructions (its version is their fingerprint) | the held-out pairs it is tried on (nine the keeper's examples never show, `tests/fixtures/microtask_cases/keeper-held-out.json`) | the evals loop already runs it: a fault found changes the instructions, which voids the proof (T8); the dated order is the record's | built |\n| The mind of Deliberus | the standing passes as organs (Gentner's two stages, § 2b) | what each pass reads | the hunt owes a second case before a kinship is minted (BH5); the keeper's `difference` as the alignable difference and first attack (AD27); the generalization's wording as the hunt's probe | partly built |\n| How we work | a ruling (a ledger row, a canon line) | the next real case it meets (a walkthrough's question, the outside reader's comparison, a session's act) | BI2 blessed; proposed, unbuilt: a `Met:` clause on a ruled row naming the dated case it last met, so a ruling that has met no case since it was written reads as such (the project's Lakatos test) | proposed here |\n| The convictions | the wager, the canon's lines | the horizon map's readings, each debate mapped | the canon pass's flow (asserted, ruled, compressed, resident) is the loop; the wager is scored by the horizon map, never by the loop running (convergence belongs to the score); one head's equilibrium gives none between heads, which is what the common prior (BC1) and the reading of evidence argued in the open (the theorems line) build toward | ruled, built in part |\n\nTwo musings past the table. **The `Met:` clause** is the one new proposal: a ruled row carries, beside its *Meets:* line,\nthe dated real case it last met and whether it held (*Met: 2026-11-06, the 23-minute piece, held after rewording*),\nwritten by the session that met it; a ruling with no *Met:* reads as untested (⚠ 2026-11-07: untested, which\nis not *a hypothesis* or *a guess*; an operational rule can be right without having fired, and a proof needs no\ncase; Codex, § 7), and `ruling_ripples.py` could list them. **The asymmetry may be the same at every level** (a hypothesis, § 0 point 4): the hunt over-produced\nkinships and the gate under-admitted members on one night; sessions over-produce resident lines (the canon pass's budget) and\nunder-apply them to the next case (this morning's length rule); the wager's danger is likewise two-sided, a horizon\nread as settled on an untested arrangement (over-production) and a reading refused where the moves have not been made\n(the conservative default, ruled). So the brake belongs on minting at every level and the generosity on filing, as\n§ 0 point 4 says for the graph.\n\n## 7. Codex's review of this note and the walkthrough (2026-11-07)\n\nFredrik had Codex review the whole session, this note and the walkthrough, after its October review of the\nindependence question; its reviews are born private: `.private/collab/codex-analogy-generalization-review-2026-11-07.md` (the critique, with a read-only check\nscript and its output beside it), `.private/collab/codex-gentner-and-deliberus-development-2026-11-07.md` (the\nconstructive note on Gentner and the mind of Deliberus), `.private/collab/codex-independence-review-2026-10-26.md`\n(October). Its judgment: keep the blessed principle (BI2) and the version-sensitive scrutiny; revise the counting\nand validation proposals before building them. What this note adopted the same day, each marked ⚠ above:\n\n- **Filed after is not tested** (§ 4f, § 5.1): a case's role (formulation, fixed-version evaluation, unknown\n  exposure) is recorded beside its date; the hunt's inclusion rule can make a 20% population read as 71%\n  among found cases with no arithmetic error.\n- **The resilience pair supplies no trials** (§ 4b): two general claims lie within the minted generalization\n  as narrower reference classes; the count waits for observed cases; 0.143 is withdrawn as a strength.\n- **Eligibility, outcome and evidential role are separate** (§ 5.1): a failed member stays a failure while\n  eligible; a named exception keeps the failed prediction under the old scope.\n- **One shape, not one object** (§ 0, § 4a): structural alignment, reflective equilibrium, hierarchical Bayes,\n  model revision and software evaluation connect without sharing a count, a stopping rule or a notion of\n  success; Gelman's Bayesian workflow (inference within a model against building and checking models) is the\n  literature this note lacked.\n- **The asymmetry is a hypothesis** (§ 0, § 4c, § 6).\n- **Three code findings reproduced** (its script runs green against `5fa01d7d`): a bundle reached through two\n  branches of the reference-class hierarchy counted twice (fixed the same day, `reference_classes.class_count`\n  counts each bundle once by id, with the diamond as a test); the within-argument range built from each reason's\n  effect on the conclusion as if it were the premise's own chance, so a conjunction of two premises raised to 0.6\n  reads [0.25, 0.30] where the exact answer is 0.36 (kept as a strict expected failure in\n  `tests/test_combination_models.py`; the fix needs each premise's own probability carried into `combine`, which\n  changes Q48's ruled formula, so it is his to rule); and three two-premise arguments among three claims read\n  pairwise as one three-premise argument (a strict expected failure; ledger BI3, an argument has an identity of\n  its own).\n- **Superseded sentences corrected**: `accrual.py`'s docstring (supports and attacks can share a story),\n  `reasons-and-cases.md` (*the safe end*), `a-tree-per-claim.md` § 5 (a product is not Q50's relevance), the\n  tools note (the Kelvin rubric; Krauss's lag distribution is no success rate), the daemon note § 10 (a\n  generalization's kind does not name its mechanism).\n\nWhat it adds past the corrections, unruled and kept there: a comparison yields four products (a reusable\npattern, a consequential difference, a candidate inference, a revised representation), and a difference becomes\nan attack only through an inference saying what it defeats (AD27 at one inference per link); the resilience\nexample sharpened by an exact model in which independent failures help under redundancy and majority rules and\nhurt under a series rule, so *independence improves resilience* needs the aggregation rule; the mind of\nDeliberus's question put as *what information must this mind retain to justify an inference, detect its\nfailure, and revise it without counting its own reconsideration as new evidence*; three experiments (one\ncomparison family worked through on the real resilience sources; representation improvements separated from\njudging improvements; whether an assisted reading of a real disagreement is more faithful); and four candidate\nprinciples, not adopted. The walkthrough's ten overstated lines are listed in its header.\n\n**2026-11-09:** the four products traced to their theory (Gentner 2010's four kinds of learning), translated to the data model and tested on eleven live candidates and seven histories: [what-a-comparison-produces.md](what-a-comparison-produces.md) (ledger BL3 recut, BL4 proposed).\n\n\n`what-a-rule-rests-on.md` (the spiral; the walkthroughs read against the older corpus) ·\n`what-a-case-counts-for.md` (what a case does, and the count's arithmetic) ·\n`the-evals-loop-and-our-machine-passes.md` and `three-minds-and-the-leap.md` § 5a (his method, one case at a\ntime) · `the-canon-distillation-pass.md` (the project-level loop) · `flywheels-in-building-deliberus.md`\n(ruling-to-canon, the answer key) · `the-kind-of-mistake-we-make.md` § 3 (what one direction costs) ·\n`overlaps-and-analogy.md` § 15 and `the-analogy-daemon.md` § 10 (the kinship structure this keys) ·\n`priors-as-standing-work.md` § 11 (the night's filings) · `tools-and-analogies-as-engines-of-discovery.md`\n(Mitchell, Krauss, the six analogies) · `ranking-reasoning-quality-without-bedrock.md` § 4 (the circle has\nno bottom and that is fine) · the listening pieces of 8 October, 23 October and 5 November · the helper's\nreport, verbatim, in `cases-and-generalizations-from-both-ends/literature-report-2026-11-06.md`.\n"}