{"path":"research/parsimony-and-the-core-data-model.md","content":"# Parsimony and the core data model: *listen to the architecture* run through the four checks\n\n**Date**: 2026-11-09 · **Type**: a conviction tested for fractal priors, with the gotcha it yields and one live\ncase · **Status**: the test's method and rulings BL1, BL2 and BL3 are his (BL2 recut and blessed, § 9; BL3 ruled in the evening in the wording of `what-a-comparison-produces.md` § 5, with BL4); the verdicts in\n§ 3 and § 4 are Claude's and unruled; BL3 is proposed. Home of ledger Arc BL.\n\n**His words, the same morning** (after the SOAR reading, `soar-listen-to-the-architecture.md`):\n\n> *\"Yes, make it a gotcha. It should be a strong directive/principle.\"*\n\n> *\"We should still have openness to data model extension/modifications/refactoring, but using it recursively\n> the way we have and leaning strongly toward developing and reusing the core data model, if need be sharpening\n> and adapting it to reality but with the least modification possible, ensuring elegance... kinda like when\n> physicists try to distill a theory of everything and aim for elegance in parsimoniousness or parsimony, I\n> guess. Research this a bit to see if we have fractal priors here undergirding these idea strands.\"*\n\n**Read for this note**: `convictions-across-scales.md` § 1 and § 6 (the four checks and what a result\nlicenses), `claim-or-system.md` § 3 (the sort), `the-kind-of-mistake-we-make.md` § 9 (the `confirmed` field),\n`what-belongs-in-the-ontology.md` § 5 (de-bake, confess, publish; fuzziness cheap, opinionatedness not),\n`saying-what-kind-a-claim-is.md` § 10 (the four products in the data model), `cases-and-generalizations-from-both-ends.md`\n§ 7, Codex's note of 2026-11-07 (private, § on design implications), `docs/graph-properties.md` (header),\n`deliberus/keeper.py` (the likeness fields). Outside sources opened this session: Einstein's 1933 Herbert\nSpencer lecture (*Philosophy of Science* 1934), MacKay's chapter on model comparison, the Stanford Encyclopedia\non Lakatos, Jacob's *Evolution and Tinkering* through the Embryo Project's summary and quoted passages,\nMathlib's *values* and *PR review* pages, reviews of Hossenfelder's *Lost in Math* (Ars Technica; Wilczek in\n*Physics Today*), Landau's notes and Satık (2022) on the strong minimalist thesis. Common law as a case is\nfrom Claude's training and marked so.\n\n---\n\n## 0. In short\n\nThe method, *listen to the architecture*, is **Raised**: it recurs at six independent levels with one mechanism\nunder it, that extending a shared representation is cheap exactly where the identity between old and new is\ncheap to decide, and that a part added outside it must be reconciled by hand forever. The aim he names,\nparsimony, is **Conditional**: it transfers where parsimony is measured as the Occam factor measures it, by\nhow many independent things a reader must know and how many free settings nothing reads, and it fails where\nit is measured as felt elegance alone, which led a generation of physicists astray and which biology never had.\nElegance as a goal of its own is **Lowered**; elegance of the *theory* behind a design, examined and defined, is\na criterion of near-equal weight with fit, on his correction (§ 9). The substrate condition every lower level needs is a core that\nis in fact fixed and small; ours is still being distilled, which is what the reviews are for, so the rule's\njob today is to keep the protective belt from thickening with stored properties while the core settles. The\nrule itself (BL1, ruled) is in § 5; the four products of a comparison, the case hot off the presses, are\nworked through in § 6: three have homes, the fourth is a claim, and the proposed property is a reading.\n\n---\n\n## 1. The conviction, stated so it can fail\n\nTwo strands, which his sentence joins and which the test takes apart:\n\n- **The method**: a new capability enters through what the data model already affords (a claim of an\n  existing kind, a link type that exists, a row in the record of machine judgments, a reading computed from\n  them), and a new kind, label, link type or stored property is admitted only where some operation the project\n  does to claims does not apply (BD13's test, read the other way).\n- **The aim**: when the core does have to move, it moves by the least change that fits reality, toward fewer\n  independent elements, with the openness he names kept: refactoring is allowed, patching is not.\n\nThe first is Newell's maxim as Laird gave it. The second is his, in the physicists' words. The test asks of\neach whether it recurs, whether the mechanism is the same or only the shape, what adversary our rung adds,\nand what the lower levels needed in order to work.\n\n---\n\n## 2. The levels\n\n| Level | The instance, in its own words | Mechanism | The case against, at the same level |\n|---|---|---|---|\n| **Physics** | Einstein, 1933: *\"the supreme goal of all theory is to make the irreducible basic elements as simple and as few as possible without having to surrender the adequate representation of a single datum of experience\"*; unification as *\"the reduction in the number of the logically independent conceptual elements\"* | fewer independent elements, each answerable to every datum | Hossenfelder, 2018: symmetry, unification and naturalness, extrapolated from the Standard Model's success, guided forty years of theory to nothing the collider found; *\"physics has a graveyard full of beautiful ideas\"*; Newton's gravity and quantum mechanics were called ugly on arrival. Wilczek's reply: the target is not beauty but *\"self-satisfaction, which encourages disengagement from reality\"* |\n| **Statistics** | MacKay: Bayesian model comparison *\"embodies Occam's razor automatically\"*: a model with more free parameters spreads its predictions thinner, so where both fit, the simpler is more probable *\"without our having to express any subjective dislike for complex models\"*; the Occam factor also penalises a model that must be finely tuned | probability mass; a free parameter nothing constrains costs evidence | none needed: this is the level that says what parsimony *is* when it is not aesthetics |\n| **Philosophy of science** | Lakatos: a research programme keeps a *hard core* that *\"we must use our ingenuity\"* to protect, and adjusts a *protective belt* of auxiliary hypotheses; the programme is *progressive* while each adjustment predicts something new that is then found, *degenerating* when adjustments *\"only accommodate\"* anomalies | a core protected because refuting it loses the programme's content; the belt bears the tests | Kuhn: anomalies accumulate until the framework is replaced, not extended; Lakatos's own exit is the degenerating shift, which Worrall reads as Kuhn's crisis given an objective test |\n| **Biology** | Jacob, 1977: *\"evolution does not produce novelties from scratch. It works on what already exists\"*; natural selection *\"works like a tinkerer … uses whatever he finds around him\"*; the molecular level shows *\"an underlying unity\"* (one code, shared pathways) and the organismal level *\"many imperfections\"* | descent: a new part is a modified copy of an old one, so compatibility is inherited; the conserved core (the code) is what every part must be legible to | the result is a patchwork, never elegant: the human brain as *\"an imperfect patchwork\"* of old and new; a system that reuses everything accumulates redundancy and waste |\n| **Mathematics (Mathlib)** | the library's values: definitions *\"very broadly applicable\"*, generality chosen *\"to reduce the likelihood\"* of breaking downstream projects; its review guide's first question, *\"Do the results already exist? possibly in a more general form with a different name?\"*, answered by trying `exact?` before writing; definitions settled *\"after a lengthy trial and error process\"* and refactored | dependency cost: a new definition that does not fit the hierarchy breaks what builds on it, and identity between a new lemma and an old one is decidable because Lean has types | Gouëzel's manifold definition took 1,600 lines: the right generality is found by use, not by taste, and a definition too general *\"you cannot work with\"* |\n| **Cognitive architecture** | Laird, 2026: *\"keep looking for capabilities the system doesn't currently have and ask how can we add those … in a way that respects everything that's already there, but also takes advantage of what's there\"*; the architecture *\"should be pretty solid so that … you're not … going down and saying, well, let's do this in the architecture\"* | a fixed level of abstraction that all knowledge is written in, so everything added stays learnable and usable by one engine | his own correction of SOAR's early self: three memories had to be added that the first architecture lacked; listening told him how to add them, not that nothing was missing |\n| **Linguistics** | Chomsky, 2000: *\"language is an optimal solution to legibility conditions\"* (the strong minimalist thesis), reducing what is proprietary to the language faculty to as little as Merge | the same aim as Einstein's, applied to a biological organ | Satık (2022) and Progovac (2019): the thesis is too strong and perhaps unfalsifiable (*syntax must be simple to be evolvable, and must have evolved, so it must be simple*); cartographic hierarchies do not reduce to the interfaces |\n| **Law** (training, marked) | common law extends by analogy from decided cases with the least departure from precedent, and a doctrine that needs a new exception for every new case is the sign it is wrong | the same shape, reuse by identity; a degenerating doctrine as the sign to recodify | codification: at some point the accumulated exceptions are replaced by a statute, Kuhn's move in legal clothing |\n\n---\n\n## 3. The four checks\n\n**1. Recurrence.** Raised. Six levels named by people not arguing for this project, from physics to a\ncognitive architecture, each stating the method in its own words; the aim, parsimony, is stated at four of\nthem (physics, statistics, linguistics, mathematics) and refused at one (biology).\n\n**2. Mechanism identity.** This is where the grand version dies and a narrower one survives. The levels do\nnot share one mechanism for *why* to prefer the existing architecture: in biology it is descent, in science\nit is the methodological cost of losing a programme's content, in a library it is downstream breakage, in\nstatistics it is probability mass. What they share is narrower and exact: **extending a shared\nrepresentation is cheap where the identity between old and new is cheap to decide, and a part added outside\nthe representation must be reconciled with it by hand, forever.** The genetic code, Lean's types, SOAR's\nworking memory and Mathlib's `exact?` are each the thing that makes identity cheap. That is the reuse\nflywheel's condition (`the-reuse-flywheel-prior.md`: *identity must be cheap to decide*) arriving as the\nmechanism under the method, which is why the two notes agree. The aim, parsimony, has a mechanism only at the\nstatistical level: MacKay's Occam factor says a free parameter nothing constrains costs evidence. At the\nother levels it is an aesthetic report of that fact, which is what Hossenfelder found when the report came\nloose from the fact.\n\n**His lens on the mechanism, the same morning** (*\"It occurs to me that BL1 is in a way an intentional form of the\nnecessarily narrow/provincial (or whatever words they use) reuse of complexity (molecules already existing in the\nenvironment etc) that Assembly Theory says happens as a natural result of accumulated complexity in the\nevolutionary/selection environment. No?\"*). Yes, with the two things the lens hands back. Walker's statement of\nit, already in the corpus (`assembly-theory-and-the-reuse-mechanism.md`): *\"at a certain complexity threshold, the\nspace is so large that there's no exhaustive search possible. So anything you build in that space has to be along\na constrained space where you build from parts that already exist.\"* So the narrowness is not a preference but\nthe only construction available, and what exists is a historically contingent path through the possible. BL1\nadopts that constraint on purpose, and the lens shows what the deliberate version must add that selection never\nhad: an exit. Nature does not refactor; it tinkers (Jacob), so the path it took stays locked in, which is the\ncorpus's frame lock-in class seen from physics. BL2's Lakatos clause (refactor by ruling when patches stop\npredicting) is the one move the intentional form can make that the natural one cannot. And the theory's own\nhardest problem, Walker's *\"identifying the boundaries of the selected units\"* (copy number), is the mechanism\nnamed above, identity cheap to decide, which is why the keeper sits where it does. The corpus's standing caution\nholds: the formalism of construction with reuse transfers, the theory's authority does not.\n\n**3. Adversary.** At our rung the defector is not a model or a user; it is **a session under pressure to\nbuild**, a motivated stakeholder in the three-class sorting: it has a stake (the capability shipped tonight)\nand is not modelling the detector. Its move is the stored property that encodes a judgment or a state, since\na property is quicker than a reading and needs no ruling. The record is unambiguous: `confirmed`, link\n`strength`, the stored terminus, the yellow link and `trust_tier`, each written in a hurry by a session,\neach silent for months, each paid at its readers (`the-kind-of-mistake-we-make.md` § 9). The lower levels\nhave no such defector: a gene does not add a property to the genome for convenience, and Mathlib's reviewers\nstand between a contributor and the library. We have a test (`tests/test_graph_properties_defined.py`, a\ndefinition per property) and no reviewer, which is why the rule in § 5 is a gate in prose and a check in code\nwhere code can hold it. No adversary of the strategy class yet: nobody outside writes the schema.\n\n**4. Substrate condition.** Every instance at a lower level requires **a core that is in fact fixed and small**:\nEinstein's irreducible elements, Lakatos's hard core, the genetic code, Lean's kernel, SOAR's architecture.\nListening to the architecture presupposes an architecture to listen to. Ours is not yet fixed: since\nSeptember the core has moved by ruling (the parts/reasons split of 2026-10-14, `confirmed` retired, strength\nreplaced by relevance, the terminus by a reading, the yellow link emptied, a reference class made a claim),\nand the canon distillation is the work of fixing it. So the condition is met in part, and the rule's job\ntoday is different from its job later: **today it keeps the protective belt from thickening while the core\nsettles** (a stored property added now is a property to retire later, as five were); **later it is Lakatos's\nnegative heuristic**, the discipline that sends every anomaly to the belt first.\n\n**Verdicts.** The method: **Raised**, with the mechanism named (cheap identity) and the substrate condition\nnamed (a fixed core, which the reviews are building). The aim, parsimony: **Conditional** on its measure being\nthe Occam factor's, the count of independent kinds and of settings nothing reads, never felt elegance.\nElegance as a goal: **Lowered**, on Hossenfelder's graveyard and Jacob's patchwork. Wilczek's reframing is the\none to keep: elegance proposes candidates and reality decides, which is *propose, never assert* applied to our\nown design.\n\n---\n\n## 4. What the corpus already held, and what the test adds\n\nThe method was already in the corpus, and once as a ruling: **BD21** (RULED 2026-10-25, his words: *\"Shouldn't this\nbe using existing argumentation structure instead? I always want to lean in that direction if possible/rational.\"*,\nand *\"which should also definitely be a standing habit/check, so we don't drift from this\"*), which says that before\na new link type, node label or reading is added the existing argumentation structure is tried first and the new kind\nships only with a ledger row saying what the existing structure could not carry; its canon decision was a check,\nunbuilt, and no resident line. The outside reader of BL1 and BL2 found this repetition; BL1 is therefore BD21 made\nresident and widened to stored properties a decision reads, with the mechanism, the exit and the record and readings\nas homes added. Three more times without its name: the sort in `claim-or-system.md` § 3\n(claim, record, reading, form), BD13 (two things are one only when every operation applies to both; its\nconverse admits a new kind only where one does not), and the three responses of `what-belongs-in-the-ontology.md`\n§ 5 (de-bake, confess, publish). The gotcha that *a stored property is defined where it is written* is the\nmethod at the level of one property. BI4 (a reference class is a generalization used as a prior's source,\n*\"never a second kind of object\"*) and AG12 (a machine's sentence is a claim) are the method applied.\n\nWhat the test adds:\n\n1. **The exit.** His openness to refactoring is Lakatos's degenerating shift made operational: when the\n   accommodations run ahead of what they predict, when each new register needs a new stored property and\n   none of them says anything a reading could not, the core is wrong and is refactored. The parts/reasons\n   split of 2026-10-14 was a progressive shift by this measure: it predicted the 24 reasons filed as parts\n   and the cap by the weakest part on things that were not parts, both then found.\n2. **The measure.** Parsimony is counted, not felt: kinds of object; link types; stored properties that a\n   decision reads; settings two people could want different that are not recorded settings. The corpus's\n   own asymmetry (fuzziness is cheap to fix later, opinionatedness is not; a lasting fact about registers not yet\n   seen, by the three-runs law, and so a claim about the design, not a cost of now) is Hossenfelder's warning in our\n   words: an over-committed elegant schema *\"silently distorts every source in a register it was not induced\n   from\"*. So the parsimony that is wanted is few kinds with confession slots, never few kinds with forced\n   fits.\n3. **The sequencing.** Because the substrate condition is a fixed core, the rule does not say *never\n   change the core*. It says the core changes by ruling, as one object at a time, with its theory settled\n   first (his own ruling of 2026-10-26, a quantity's meaning settled before its formula changes), and everything else enters\n   through what the core already affords.\n\n---\n\n## 5. The rule (BL1, ruled)\n\nThe line proposed for the Principles, queued in `the-canon-distillation-pass.md` § 4k for the sitting's one\nwrite, 122 words (the cap is 120; a word or two over is left as it is), with BD21 (ruled 2026-10-25) as the ruling it makes resident:\n\n> **A capability enters the graph through what the data model already affords: a claim of an existing kind, an existing link type, a row in the record, or a reading; a new kind, label, link type or stored property that a decision reads is admitted only where some operation the project does to claims does not apply; a run of accommodations that each add a property is the sign the core is wrong and is refactored, never patched.** *Listen to the architecture* (Newell). Demonstrated wrong by default: `confirmed`, link `strength`, the stored terminus, the yellow link and `trust_tier`, each a judgment stored as a property (BD21, ruled 2026-10-25; BL1, 2026-11-09, *\"Yes, make it a gotcha. It should be a strong directive/principle.\"*; `parsimony-and-the-core-data-model.md`).\n\nThe three headline tests: it names a use (*admitted only where*), not a noun, so it forbids nothing we do and\nwant: a cache of a reading marked derived and rebuilt from its source is a reading, and a property that holds\nwhat somebody said or who said it is content or record, both admitted. A stranger reads *data model*, *claim*,\n*link*, *record* and *reading* as ordinary words and *listen to the architecture* as the quoted maxim it is. It\ncollides with nothing: *architecture* is kept to the quotation, and the body says *data model*, since the\nglobal rules use *architecture* for a layer stack of their own. The procedure a session runs is the sort's\nfour questions in order: could two people disagree about it (a claim); does it say who did what (the record);\ncan it be computed again from those (a reading, and never a stored fact); what is left is form, kept small,\nwith a *does not fit*.\n\n**What is enforced in code already** (`docs/graph-properties.md`): a property the code writes has a line in\nthe table before its first reader, and a property nothing writes any more is removed from it. **What the rule\nadds that code does not yet check**: whether a new property is a reading in disguise. A census the test could\nrun: for each stored property, is there a function that computes the same value from claims and the record?\nNot built; proposed as the check this principle gets, by the ladder (a check in code when code can enforce\nit).\n\n---\n\n## 6. The case hot off the presses: the four products of a comparison\n\n**Origin, read whole.** Codex's note of 2026-11-07 (private): *\"A candidate analogy can have at least four\nproductive outcomes: a reusable abstraction, a discriminating difference, a new conjecture, or a revised\nrepresentation. A design that only records shared class membership loses some of these products … We should\nask which product a comparison produced and what subsequent work it enables.\"* The walkthrough of that\nevening spoke them as a pattern, a difference that matters, a guess to check and a better description. On\n2026-11-08 the analogy session placed them in the data model (`saying-what-kind-a-claim-is.md` § 10, last\nparagraph): a pattern is a generalization with the compared cases as its members (BH5); a guess is a claim\nnobody asserts (kind `machine_step`, AG12) with its link marked proposing; a sharper description is a\nquestion (a potential, rendered on read) or a concept distinction, which the concept layer holds; and *\"a\ndifference that matters is a claim about the pair … and has no home yet beyond a reason attached to the\nresemblance link's relevance claim\"*, with *\"recording which product a comparison made is one property on\nthe resemblance link, defined in graph-properties.md if he rules for it. Unruled.\"*\n\n**Run through the rule.** Three of four fit kinds that exist: the session listened. The two open items are\nthe rule's two cases, and it decides both without a ruling per case:\n\n- **The fourth product is a claim, and its home exists.** *In crowds diversity is an intention, in\n  governance an effect* is a sentence two people can disagree about, so by the first question it is a claim.\n  Which claim: Codex's own answer, kept in `cases-and-generalizations-from-both-ends.md` § 7, is that *\"a\n  difference becomes an attack only through an inference saying what it defeats\"* (AD27, one inference per\n  link). So a difference that matters is an attack on the positive analogy's generalization (*what holds of\n  crowds holds of institutions*), or on the relevance claim of the link that carried the pattern across, with\n  the inference written out: *the pattern depends on diversity being an effect; in crowds it is an\n  intention; so the pattern does not carry*. No new kind. The keeper's likeness already asks for exactly this\n  sentence (its `difference` field, *\"the negative analogy: the strongest relevant difference\"*, built\n  2026-11-06), which is the next item.\n- **Which product a comparison made is a reading, not a property.** By the third question: it can be\n  computed again from what the comparison wrote and from the record of the call that wrote it. A\n  generalization minted with the cases as members is the pattern; a `machine_step` claim with a proposing\n  link is the guess; a potential question or a concept split is the description; an attack by inference is\n  the difference; a stop answer in the record is a comparison that produced nothing. A `product` property on\n  the resemblance link would store that reading, and a stored reading goes stale the moment a person\n  reclassifies the claim it summarises. So the answer to *\"if he rules for it\"* is that he need not: the rule\n  rules against it, and a function `products_of(comparison)` over the graph and the record is the home.\n\n**A later reading, 2026-11-09 (the reference-class session), unruled.** Of the two targets § 6 offers the attack, the\nmembership built on 2026-11-07 names the nearer one: *the pattern depends on diversity being an effect; in crowds it is\nan intention; so the pattern does not carry* says that the crowds case is not one of the population the generalization\ncounts, which is an attack on the membership claim that files it (*The claim \"…\" is about one of: <population>*), the\nplace AT13 gives an exception (a claim of its own, never a list on the rule). Filed there, the difference that matters\ncounts once (BC8), lowers that case's standing under the pattern and nothing else, and is a case of BI2 (a case read\nagainst its generalization). The attack on the link's relevance claim stays for a difference that breaks the carrying of\none claim rather than a case's belonging.\n\n**The test's first live run, proposed (BL3).** The resemblance link carries, since 2026-11-05 and 06,\nfourteen stored properties for the analogy work (`graph-properties.md`: `likeness`, `likeness_mechanism`,\n`analogy_candidate`, `analogy_withdrawn_by`, `analogy_verdict`, `analogy_judged_by`, `analogy_words`,\n`analogy_judged_at`, `hunt_proposed`, `hunt_mechanism`, `hunt_respect`, `hunt_difference`, `hunt_at`,\n`hunt_sources`). They were written under his exemption of that night to get the work moving, and each has a\nline in the table, so they break no rule that stood. Read against this one: the positive analogy as a general\nsentence is a generalization claim (BH5 says so); the negative analogy is the attack above; a person's\nverdict with their words is a claim of theirs or telemetry (the 2026-08-31 ruling: bare judgments are\ntelemetry); who proposed the pair, when and from which sources is the record; `analogy_candidate` and\n`analogy_withdrawn_by` are states a reading gives from the rest. The proposal is not to rip them out tonight\nbut to run the census of § 5 on this link type first, since it is the freshest and the fattest, and to say\nfor each of the fourteen which of the four homes it belongs in. That is what *sharpening and adapting it to\nreality with the least modification possible* looks like on one object.\n\n**Astra's review, 2026-11-09, read whole** (`.private/collab/codex-arc-bl-catch-up-review-2026-11-09.md`, cut off\nat its § 4 by a usage limit; Fredrik asked for it because, in his words, *\"most of BL3 is still a bit up in the air\nfor me, I don't feel a 'hell yeah' for everything yet\"*). What it changes in § 6:\n\n- **The four products are achievements, not a partition.** Codex's note said *at least four productive outcomes*;\n  a comparison may yield several, one or none, and a new conditional generalization can be pattern, guess and\n  description at once. So `products_of(comparison)` cannot map new object → product mechanically; it reads the\n  outputs, their relations, their status and the record.\n- **A difference is not already an attack.** Codex's condition was *a difference becomes an attack only through an\n  inference saying what it defeats*; § 6 turned that into *a difference that matters is an attack*. The target\n  follows the inference as written: the case's membership claim (the case is not in the population), the\n  generalization (an eligible case fails it, and by `cases-and-generalizations-from-both-ends.md` § 7 an eligible\n  failure stays a failure), the transfer's relevance claim (the mapping lacks a condition this carrying needs), or\n  nothing, when the contrast sharpens the shared structure and refutes no proposition. Routing every failure to\n  membership would let a generalization escape by excluding its failures. This dissolves the fork the session put to\n  him (generalization or membership): the third option is *by the inference*.\n- **The crowds case does not establish its attack.** Intending diversity and obtaining diversity's effect are\n  compatible; the written bridge *the pattern depends on diversity being an effect* supplies an unwarranted premise;\n  and *what holds of crowds holds of institutions* is an unrestricted transfer, not the shared generalization. The\n  mechanism-specific candidate (collective failure depends on the joint distribution of local failures and the\n  aggregation rule) is the development case, read from the sources.\n- **A sharper description can change relations without a concept split**: a common cause found, an intention told\n  from its realization, a missing participant, two reports of one event. Those enter as claims and proposed links,\n  which exist; the before and after readings and the reason are kept.\n- **A case filed under a pattern by analogy is not an observed outcome.** Two statements instancing a schema, two\n  systems instancing a causal model and two observed outcomes informing a rate are three roles; the resilience pair's\n  withdrawn count must not come back through membership (BI4).\n- **Read, never store, keeps BL1's qualification.** The present reading of what a comparison achieved is a derived\n  view; that a person or microtask classified it so at a time is history, and history does not move when the\n  reading does. Where the reading takes a judgment, the judgment is recorded with inputs and version (the record of\n  machine judgments; a person's take in `reading_takes`) and the view derived under a declared policy.\n- **A code finding, verified by this session**: `scripts/analogy_candidates.py --judge` writes\n  `analogy_verdict`, `analogy_judged_by`, `analogy_words`, `analogy_judged_at` with `SET` on the resemblance link and\n  writes nothing to the record, so a second verdict overwrites the first and nothing shows a judgment was ever made.\n  The record has a table for exactly this, `reading_takes` (a person's take on one reading, his words verbatim).\n  Fixing that is the census's first step, and the census's acceptance test is a round trip of one real comparison:\n  recover its inputs, versions, outputs, judgments and present view, revise the generalization, and show the history\n  fixed while the view changes.\n\n**BL3 recut, proposed 2026-11-09 after that review** (the ledger row carries it beside its first wording):\n\n> **A comparison is a structural alignment of two arguments (Gentner), its mapping the one recorded judgment; each product is an element's relation to that mapping, read off it, and a comparison may yield several or none. A reusable pattern is a generalization with the compared cases as members, each case keeping the role the comparison gave it, never counted as an observed outcome; a guess to check is a claim nobody asserts, its link marked proposing; a sharper description is a potential question, a concept distinction, or a proposed link; a difference that matters is a different filler in a matched role, written out as an inference, and the role names what it attacks (the case's membership claim, the rule's scope as an exception, or the transfer's relevance claim); a difference at no role attacks nothing and stays in the record. Which product a comparison made is a reading over what it wrote and the record; the mapping is recorded with its inputs and version, never a `product` stored on the link.**\n\nThe census of the fourteen properties is work, not a ruling (it applies BL1), so it moves to `TODO.md` with the\noverwrite fix as its first step.\n\n**The theory the four come from, named 2026-11-09 on his question** (*\"I feel like maybe there's not enough connection\nto theory here for me … why the 'four products' are what they are and how they should be thought of best in terms\nof/translated to the core data model\"*). Codex's note of 2026-11-07 drew them from Gentner's *Bootstrapping the Mind*\n(Cognitive Science 34, 2010, pp. 754–755, read from the PDF this day): *\"Achieving a structural alignment sets the\nstage for four kinds of learning: abstraction, contrast, inference-projection, and rerepresentation.\"* Structural\nalignment (Gentner 1983; Falkenhainer, Forbus and Gentner 1989) maps two structured representations role to role,\none to one, preferring the match that carries the deepest connected system of relations. Given a mapping, every\nelement of either structure stands in one of five relations to it, and that is why the products are four and not\nsome other number: **matched** (the common system: *abstraction*, the shared schema that can be used again),\n**present in the richer side, connected to the matched system, absent in the other** (*inference-projection*: a\ncandidate inference, which Gentner's school always marks as to be checked), **in a matched role with a different\nfiller** (an *alignable difference*: *\"differences that occupy the same role in the two systems\"*, which is what\nmakes a difference matter, since a difference at no role is inert), **a predicate that should match and does not**\n(*rerepresentation*, by *minimal ascension*: *\"finding the most specific relation that is superordinate to the two\nnonmatching relations\"*), and **in no role** (a nonalignable difference, which produces nothing). Hesse's three\n(positive, neutral and negative analogy, the keeper's `respect`, `carried_over` and `difference`) are the first,\nsecond and third; Hesse has no fourth because his descriptions are fixed, and Deliberus's are not (nothing is\npermanently atomic; a claim can be cut and a concept split), so rerepresentation is the product this ontology\nspecially affords.\n\n**Translated to the core data model.** Deliberus's structured representation is the argument: premises, a rule,\na conclusion, and the chain under a link (AG14). A comparison aligns two arguments role to role, and the mapping\nis the one judgment, recorded (the keeper's likeness fields are it). Then each product is a reading off the mapping\ninto a kind that exists:\n\n| Gentner | What it is in the graph | Where it lands |\n|---|---|---|\n| abstraction (the common system) | the rule both arguments instantiate | a generalization claim with each case filed under it by a membership claim (BH5, BI1, BI4) |\n| inference-projection | a premise or conclusion the richer argument has and the other lacks | a claim nobody asserts (AG12), its link marked proposing |\n| contrast (alignable difference) | a different filler in a matched role | a claim at that role; the role names the target: a case lacking a condition the rule's population requires attacks its membership claim; an eligible case failing the rule is an exception, an attack on the rule's scope (AT13); a condition the carrying itself needs attacks the transfer's relevance claim (AD27) |\n| rerepresentation | two wordings or terms that should match and do not | a concept distinction or shared sense (the concept layer), a sameness verdict (two wordings, one claim), or a recut at the floor; never a link |\n| nonalignable difference | a difference at no role | nothing: a stop answer in the record |\n\nSo Astra's *route by the inference* has a mechanism: the inference names the role, and the role names the target.\nAnd Astra's *a common cause found* is not a rerepresentation but an inference-projection or an abstraction (a\nhigher-order relation projected or extracted), so the theory also sorts the additions.\n\n**One example whole: the atom and the solar system** (Rutherford to Bohr, the structure-mapping school's own case).\nTwo arguments as the graph would hold them:\n\n- Base: *the sun attracts a planet* → *the planet orbits the sun* → *the orbit is stable*, under the rule *a body\n  under a central inverse-square attraction follows a stable orbit*.\n- Target: *the nucleus attracts an electron* → *the electron orbits the nucleus*; nothing yet about stability.\n- **Abstraction**: the shared rule, minted as a generalization with both cases as members: *a body under a\n  central inverse-square attraction follows a stable orbit*.\n- **Rerepresentation**: *gravity* and *Coulomb attraction* do not match as predicates; they ascend to *central\n  inverse-square attraction*, a concept with two senses, so the rule above can be stated at all.\n- **Inference-projection**: *the electron's orbit is stable*, a claim nobody asserts, its link from the shared rule\n  marked proposing (and *an orbit may have any radius*, a second projected claim).\n- **Contrast**: in the matched role *what the orbiting body is*, the fillers differ: a planet is neutral, an\n  electron is charged, and by Maxwell *an accelerating charge radiates energy*. The inference: *the rule's\n  stability needs no energy loss; an orbiting charge loses energy; so the carrying fails here*. That attacks the\n  transfer's relevance claim (the link carrying *stable orbit* to the electron), not the electron's membership in\n  *bodies under a central attraction* (it is one) and not the rule for planets (which stands). Bohr's answer, a\n  new claim: *the electron's orbits are quantised*.\n- **Nonalignable**: *the sun is hot*; no role in the target; nothing written, a stop in the record.\n\nThe same machinery on a public debate: *rent control* ↔ *a minimum wage*, both price controls. Abstraction: *a\nbinding price control moves the quantity traded off the market's quantity*. Inference-projection: *rent control\nproduces a housing shortage* → *a minimum wage produces unemployment*, proposing. Contrast at the role *who sets the\nprice otherwise*: where one employer sets wages (a monopsony), the case lacks a condition the rule's population\nrequires, so the difference attacks the minimum wage's membership claim under *competitive price controls*, and\nthe projected unemployment does not follow (Card and Krueger). The same kind of difference, a different role, a\ndifferent target.\n\n**The all-angles pass on BL3, the same evening** (`what-a-comparison-produces.md`): the theory and the example above were tested against eleven live analogy candidates and seven historical analogies with known fates, and the literature read on where a difference lands. The recut above is replaced by that note's § 5 item 1, which names the default landing (the transfer's relevance claim) and the two others (the membership claim, where the difference is a condition the generalization's population states; the generalization, only by a counterexample case or a rival case). The ledger row carries the recut; BL4 (the difference names its role) is proposed beside it.\n\n---\n\n## 7. Proposals (ledger Arc BL)\n\n- **BL1, ruled**: the principle of § 5.\n- **BL2, ruled 2026-11-09** (*\"BL2 seems very sensible, blessed. But elegance, we need to ask ourselves what that\n  means rather than just disregard it.\"*) **and blessed the same morning in its recut form** (*\"Blessed.\"*): parsimony's\n  measure and exit, with the elegance clause recut in § 9.\n  As first proposed: *The core data model is extended by the least change that fits reality; parsimony is counted\n  (kinds of object, link types, stored properties a decision reads, unrecorded settings two people could want\n  different), never felt; and a run of accommodations that each add a property without predicting anything a\n  reading could not is the sign the core is wrong (Lakatos's degenerating shift), at which point it is refactored\n  by ruling, one object at a time, theory first.* The *never felt* was the part he corrected.\n- **BL3, RULED 2026-11-09 evening** (*\"Recommendations blessed. Build whatever is buildable.\"*, to the recommendations of `what-a-comparison-produces.md` § 5, whose first item is the ruled wording; recut twice that day after Astra's review, § 6, last part): the four products' homes, with the\n  difference's target following the written inference and *which product* a recorded judgment read, never a property.\n  The census of the resemblance link's fourteen analogy properties is work in `TODO.md`, its first step the fix of\n  `analogy_candidates.py --judge`, which overwrites a person's verdict instead of recording it.\n\n---\n\n## 8. Sources\n\n- Einstein, A. (1934). On the Method of Theoretical Physics (the Herbert Spencer Lecture, Oxford, 10 June\n  1933). *Philosophy of Science* 1(2), 163–169. The aphorism *as simple as possible, but not simpler* is a\n  1950 paraphrase by Roger Sessions, not Einstein's sentence.\n- MacKay, D. J. C. (2003). *Information Theory, Inference, and Learning Algorithms*, ch. 28, Model\n  Comparison and Occam's Razor; MacKay (1992), Bayesian Model Comparison and Backprop Nets, NIPS.\n- Lakatos, I. (1970). Falsification and the Methodology of Scientific Research Programmes; read through\n  Musgrave & Pigden, *Imre Lakatos*, Stanford Encyclopedia of Philosophy (2016).\n- Jacob, F. (1977). Evolution and Tinkering. *Science* 196(4295), 1161–1166; read through the Embryo Project\n  Encyclopedia's summary (Racine 2014) and quoted passages.\n- Mathlib community, *Mathlib's values* and *Pull Request Review Guide* (mathlib.org, 2026); Hartnett, K.\n  (2020), Building the Mathematical Library of the Future, *Quanta*.\n- Hossenfelder, S. (2018). *Lost in Math: How Beauty Leads Physics Astray*; read through Lee (Ars Technica,\n  2018) and Wilczek (*Physics Today*, September 2018).\n- Chomsky, N. (2000). Minimalist Inquiries; the strong minimalist thesis through Landau's MIT notes (2003);\n  Satık, D. (2022). The Strong Minimalist Thesis Is too Strong, *Biolinguistics* 16; Progovac (2019) as cited\n  there.\n- Laird, J. (2026). *The Cognition Project* ep. 24; `soar-listen-to-the-architecture.md`.\n- Corpus: `convictions-across-scales.md`, `the-reuse-flywheel-prior.md`, `claim-or-system.md`,\n  `the-kind-of-mistake-we-make.md`, `what-belongs-in-the-ontology.md`, `saying-what-kind-a-claim-is.md`,\n  `cases-and-generalizations-from-both-ends.md`, `docs/graph-properties.md`.\n\n---\n\n## 9. Elegance, examined: his correction of BL2, and what the literature gives it\n\n**His words**: *\"BL2 seems very sensible, blessed. But elegance, we need to ask ourselves what that means rather\nthan just disregard it. To me, as part of judging \"whether the core must change\", it's almost of equally primary\nimportance to let ourselves be moved by the beauty, parsimony and elegance of the THEORY behind why the\ncore/ontology/data model should be set/designed some way rather than whatever other conceivable ways it could be\ndesigned. Make sense? Research this a bit as well to sharpen BL2 wisely.\"*\n\nIt makes sense, and § 3 had thrown the word out with the fault. The fault Hossenfelder names is a particular one,\nand the literature separates it from what he means.\n\n**Where elegance sits among the criteria.** Kuhn's five characteristics of a good theory (*Objectivity, Value\nJudgment, and Theory Choice*, 1977): accuracy, consistency, scope, simplicity (*\"bringing order to phenomena that in\nits absence would be individually isolated and, as a set, confused\"*) and fruitfulness (*\"disclose new phenomena or\npreviously unnoted relationships\"*). They are *\"the shared basis for theory choice\"*, and they underdetermine it:\nthey are ambiguous (Copernicus was simpler than Ptolemy in some respects and not others) and they conflict, with no\nalgorithm for weighting them. So a choice among designs that fit the cases equally is made on the other four, which\nis what his sentence asks for, and *theory* is the right word: the data model is a theory of what a claim, a link, a\nreason and a record are, and a design is a theory choice, not a taste.\n\n**What elegance is, when it is not a feeling.** Four definitions from four disciplines, each countable or\ncheckable:\n\n1. **Enlightenment** (Rota, *The Phenomenology of Mathematical Beauty*): what mathematicians call beauty is the\n   property of a statement that makes one see *why*, as against a proof that merely establishes *that*;\n   *\"enlightenment, not truth, is what the mathematician seeks\"*. For a data model: the design follows from a\n   stated theory of what the graph is, so that each object is there for a reason one can give (the sort's four\n   questions give the reason for claim, record, reading and form), and not chosen among equals.\n2. **Unification** (Einstein, § 2): *\"the reduction in the number of the logically independent conceptual\n   elements required for the basis of the whole system\"*: the same few objects and operations cover many cases.\n   Countable: BL2's first list.\n3. **No interleaving** (Hickey, *Simple Made Easy*): simple means *one fold*, one role per thing, *\"not that\n   there's only one thing\"*; and *\"simple is actually an objective notion … we can probably go and look and see\"*,\n   where *easy* is relative to the person. BD13 is this test for our objects: two things are one only when every\n   operation applies to both. A kind that does two jobs is complected, whatever it feels like.\n4. **Conceptual integrity** (Brooks, *The Mythical Man-Month*): *\"the most important consideration in system\n   design\"*; a reader can hold the whole in mind because one coherent set of ideas runs through it. For the graph:\n   a stranger told the four questions can predict where any new thing goes.\n\nGruber's five criteria for a designed ontology (1993/1995), the discipline's own list, are the same four in\nontology-engineering words: clarity, coherence, **extendibility** (*\"define new terms for special uses based on the\nexisting vocabulary, in a way that does not require the revision of the existing definitions\"*, which is BL1),\nminimal encoding bias, and **minimal ontological commitment** (*\"as few claims as possible about the world being\nmodeled\"*, which is BL2's parsimony and the corpus's *fuzziness is cheap, opinionatedness is not*). Gruber adds the\ntrade-off we meet: clarity against extendibility, extendibility against commitment.\n\n**Where elegance misleads, and the mechanism.** McAllister (*Beauty and Revolution in Science*, 1996): a community\ncompiles its aesthetic canon by *aesthetic induction*, *\"attaching to each aesthetic property a weight proportional\nto the empirical adequacy attributed to current theories that exhibit that property\"*; so *\"scientists' aesthetic\nevaluations tend in the longer term to swing into line with their empirical appraisals\"*, and the canon lags.\nSuccessful theories *\"tend to inhibit revolutions\"*, since a new theory is judged by the aesthetic of the old\nsuccesses; Newton's gravity and quantum mechanics were ugly on arrival (Hossenfelder's own examples) and became\nbeautiful by succeeding. That is the precise fault: not elegance, but **yesterday's elegance applied to today's\nanomaly**. Wilczek's reply to Hossenfelder names the same thing from the other side: the target is\n*\"self-satisfaction, which encourages disengagement from reality\"*, and *\"we need more beautiful ideas, not fewer\"*.\n\n**So the sharpened rule.** Elegance of the theory is a criterion of near-equal weight with fit when judging\nwhether the core must change, and *how*; it is examined by the four definitions, not felt; it chooses among\ndesigns that fit the recorded cases equally (Kuhn); it proposes and reality decides (Wilczek, Poincaré); and it is\ndistrusted in exactly one place, the moment the core must change, where the old canon is what is being revised and\nthe new theory's **fruitfulness** (what it predicts that is then found: the parts/reasons split, § 4) outranks its\nfit with what used to look right (McAllister). The Hossenfelder graveyard is kept as the warning about that one\nplace, not as a verdict on the criterion.\n\n**BL2, recut, BLESSED 2026-11-09** (*\"Blessed.\"*; the ledger row carries this wording and points here for the first):\n\n> *The core data model is extended by the least change that fits reality. Whether, and how, the core must change\n> is judged by two things of near-equal weight: parsimony counted (kinds of object, link types, stored properties\n> a decision reads, settings two people could want different that are not recorded settings), and the elegance of\n> the theory behind the design, examined by four tests (each object there for a reason the theory gives; the same\n> few objects and operations cover many cases; one role per object; a reader can hold the whole), chosen among\n> designs that fit the recorded cases equally, proposing where the cases decide. A run of accommodations that each\n> add a property without predicting anything a reading could not is the sign the core is wrong (Lakatos), at which\n> point it is refactored by ruling, theory first, and the one place elegance is distrusted is there: what the new\n> theory predicts outranks its fit with the old canon (McAllister).*\n\nThe resident line for the Principles is the recut above cut to the cap, queued in `the-canon-distillation-pass.md`\n§ 4k for the next sitting's write; wrong by default, demonstrated: this note's first version wrote elegance off as\n*never felt*, and he caught it.\n\n---\n\n## 10. His question: is the core data model the theory?\n\n**His words**, the same hour: *\"Could one say that the core data model IS the theory? Or at least the relationship to the\ntheory is very close...\"*\n\nNewell said yes, of architectures, and the sentence is in the lecture this note's § 2 drew on: *\"unified theories take\nthe form of architectures\"*; *\"a unified theory is given by specifying the architecture of human cognition\"*, with the\ncontent held in memory kept apart from it as software is from hardware (*Unified Theories of Cognition and the Role of\nSoar*; Lehman, Laird and Rosenbloom's introduction: a cognitive architecture is *\"a theory of the fixed mechanisms and\nstructures\"*, *\"the framework into which the content theories must fit\"*). The corpus said it of the ontology in August:\n*\"the ontology is already a theory. Not 'risks becoming'; is\"* (`what-belongs-in-the-ontology.md` § 2). So for\nDeliberus the three words sit so: **the ontology is the theory** of what the graph can hold (a claim, a link, a row in\nthe record, a reading, and the operations that apply to each, BD13's test), **the data model is that theory's notation\nin the store**, and **the claims and links are the content**, Newell's software. One can say the model is the theory in\nthe sense Newell meant: the fixed part, read against the content that moves through it.\n\nTwo asymmetries keep the two words. **The theory overdetermines the model**: one theory admits several notations (a\nreading computed on request or a value stored; a reason as a link, or as a claim about the link), and the choice among\nthem is exactly where parsimony and elegance bite. **The model underdetermines the theory**: from the schema alone a\nreader learns that there are claims and typed links, not why a part link carries no strength (BD16) or why an author's\nhedge is stance and never strength (AM6); those live in the theory and in the code that reads the notation. The\nmeasured failures are the gap read from the model's side: `confirmed`, link `strength`, the stored terminus and\n`trust_tier` were notation with no theory under them, judgments stored as properties, which is what BL1 refuses. In\nNewell's words the architecture *\"shows through\"* in behaviour; here it showed through as a property nobody could give\nthe reason for.\n\nSo the recut's *the theory behind the design* can be read as *the theory the data model embodies*: one thing seen from\ntwo sides, and the first of the four tests, each object there for a reason the theory gives, is the test that the two\nhave not come apart. **His two answers the same hour**: the resident line carries both phrasings (*\"Well... why not\nboth? I can't convince myself to entirely skip the latter phrasing\"*), and the Newell distinction is woven into it\n(*\"Document/clarify in the resident line the bit about Newell and so on, that you said to me here. Weave it in.\"*);\nthe line is resident in `CLAUDE.md` § The Principles in One Breath, after BL1's line, since the eleventh write (2026-11-09, on his *\"Write it now yes\"*).\n\n**See also (2026-11-09)**: [historical-accident-and-channelled-self-interest.md](historical-accident-and-channelled-self-interest.md), where the assembly-theory lens on BL1 (§ 3) led to his decades-old counterfactual: contingent, path-dependent assembly read at the level of economic systems.\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- **\"Should this be a new property, label or link type?\" / listen to the architecture / parsimony and the core data model / the four products of a comparison** → `docs/research/parsimony-and-the-core-data-model.md` (2026-11-09; ledger Arc BL, BL1 and BL2 ruled and blessed, BL3 proposed). The method is Raised at six levels (Einstein's irreducible elements, MacKay's Occam factor, Lakatos's hard core, Jacob's tinkering, Mathlib's generality, SOAR) with one mechanism, identity cheap to decide; parsimony is Conditional on being counted (kinds, link types, properties a decision reads), and the elegance of the theory behind a design is a criterion of near-equal weight, examined by four tests and distrusted only where the core must change (his correction; § 9), the data model being the theory's notation (Newell, § 10); the substrate every level needs is a fixed core, which the reviews are building; Lakatos's degenerating shift is the exit, when accommodations stop predicting, refactor by ruling, theory first. BL1 is BD21 (existing argumentation structure first, ruled 2026-10-25) made resident and widened to stored properties. The four products of a comparison: three fit existing kinds; a difference that matters is a claim, an attack by inference (AD27); which product a comparison made is a reading, never a stored property; the resemblance link's fourteen analogy properties are the rule's first census (BL3).\n"}