{"path":"research/tools-and-analogies-as-engines-of-discovery.md","content":"# Tools and analogies as the engines of discovery: Krauss's 761 breakthroughs and Mitchell's night science, read against Deliberus\n\n**Date**: 2026-11-05 · **Type**: source intake, the all-angles pass: a podcast episode and an\narticle he brought, read whole against the corpus, two helpers' reports kept verbatim beside this\nnote, six real analogies run through the project's own checks · **Status**: Claude's reading.\nNothing here is ruled. Three decisions for him are ledger Arc BH.\n\n**His words, verbatim**: *\"A while ago I happened upon this podcast which I felt relates to\nDeliberus, ground deeply in this and tell me your thoughts\"*, with the Spotify link to Night\nScience episode 83, and *\"Read the full https://www.aipolicyperspectives.com/p/toolstheories also\"*.\n\n**The two sources.** Night Science episode 83, *How science is secretly driven by analogy*, with\nMelanie Mitchell (hosts Itai Yanai and Martin Lercher; 16 February 2026; 33 minutes; no official\ntranscript exists, so it was transcribed locally and kept in\n`sources/podcasts/night-science-083-mitchell-analogy-2026-02-16.md` in the repository, which is not web-served).\nAnd Conor Griffin, *Tools>theories*, AI Policy Perspectives, 28 October 2026, a review of\nAlexander Krauss's *The Engine of Scientific Discovery* (2026), kept whole in\n`sources/aipolicyperspectives-2026-10-28-tools-over-theories-griffin.md`, likewise.\n\n---\n\n## 0. In short\n\n1. **Both sources say one thing about Deliberus from two sides.** Krauss: discoveries follow the\n   instrument that extends what a mind can see. Mitchell: they follow a likeness a mind draws and\n   then has to check. Deliberus is built as the first kind of thing (a graph that shows what a\n   position rests on, which no eye could see before) and holds the second as a designed, unbuilt\n   pass, the analogy daemon ([the-analogy-daemon.md](the-analogy-daemon.md)). Nothing in either\n   source reverses a ruling. Both sharpen lines the corpus already holds.\n2. **Krauss supplies the measured base rate the why-now argument lacked** (⚠ narrowed 2026-11-07 on Codex's reading: a lag distribution over recognised discoveries, conditional on a discovery having happened, is evidence about the timing and enabling conditions of past discoveries and no success rate for projects that take up a new tool; the denominator of tool-enabled attempts, failures included, is missing). In his 761 discoveries\n   a third follow their tool within four years and half within ten; since 1975 most follow within\n   four (his 2026 paper, read this session, § 3a). The project fits: reliable structured extraction\n   by language models arrived in 2023, the build began in March 2026, and the 2012 to 2014 prototype\n   died before the tool existed. **The same base rate predicts rivals**: a tool that makes a\n   discovery possible makes it multiple, and the five claim-level entrants of autumn 2026 in\n   [competitive-landscape.md](../competitive-landscape.md) are those multiples (§ 5b).\n3. **Griffin's four corrections to Krauss are the corpus's own lines**: curated data as the\n   infrastructure nobody rewards (maintenance as the class's killer), the thinker's frame move the\n   machine cannot make (the two human-reserved classes), tools and theories co-evolving (Dyson's\n   *both are true*), and the success bias in a dataset that can only see tools that worked (a filter\n   is an argument). § 4f.\n4. **Mitchell, who built Copycat with Hofstadter, says what the daemon design assumes.** Making a\n   connection is cheap and judging it is the skill (*\"there's kind of always been a fine line\"*\n   between crackpot and genius); language models are not yet good judges of which connection is\n   good; a borrowed word becomes a technical term and then hides what it does not carry over\n   (*translation*, *landscape*, *ratchet*). Dunbar's field count (2 of 99 lab analogies crossed out\n   of biology) matches the readers' rehearsal here (3 kinships in 42 findings): far kinship is\n   rare, and the daemon should expect that. §§ 4b to 4d, § 5c.\n5. **Six real analogies run through the project's checks (§ 5a)**: the checks sort the ones whose\n   critical disanalogy was knowable at the time (Kelvin's unproven stability; the ratchet's\n   condition of no recombination) and cannot sort the one whose disanalogy waited on later\n   mathematics (Wright's landscape, broken by dimensionality 65 years on). That is the daemon\n   note's own § 8 claim, *the long-run answer is the record*, now with cases. **Three decisions\n   for him** (Arc BH): a test of Mitchell's claim on these six cases, the name of the proposed\n   roaming signal, and whether a concept born by analogy records what it does not carry over.\n\n## How this was done\n\nThe six parts of the all-angles pass, each done or skipped with its reason.\n\n1. **The question pinned** (§ 1): his two messages quoted whole; which parts the corpus already\n   rules on and which are fresh.\n2. **The corpus inventoried before any research** (§ 2): Darwin was unreachable this session (the\n   Mac pings it and gets nothing; `mannaminne` and the corpus graph's meaning search both time out),\n   so the inventory was made by file (`rg`), by `corpus_graph.py around` on each home note, and by\n   reading. The notes read whole or in part are listed in § 2 by name.\n3. **Research only on what the corpus lacks** (§ 3): two helpers on Sonnet, read-only, one on\n   Krauss's dataset and the instrument-first tradition, one on Mitchell's own account and the field\n   studies of analogy in working science; their reports verbatim in\n   [`tools-and-analogies-as-engines-of-discovery/`](tools-and-analogies-as-engines-of-discovery/).\n   Where a helper could not reach a primary text, Claude read it: Krauss's 2026 paper (19,930\n   words of its text) and the abstract and results of his 2024 paper on the scientific method.\n4. **Theory first, then tests** (§ 5): what first principles and the literature settle is written\n   before the cases; six real analogies from the episode and the helper's pair, none used in the\n   corpus before (`rg` finds no *Muller's ratchet*, *fitness landscape*, *Gavrilets*, *Galison* or\n   *Kelvin's vortex* in `docs/`), run through the project's checks by hand; Krauss's thesis tried\n   on the project's own history; Dunbar's proportions set beside the readers' rehearsal.\n5. **Set against what is settled** (§ 6): the rows and canon lines each recommendation touches,\n   the assumptions ranked, and the outside check run on the draft.\n6. **Recommendations and the decisions** (§ 7, § 8), then pointers in every note whose open\n   question this touches, the README, the sources index, and the ledger.\n\n**Files read whole this session**: `the-analogy-daemon.md`, `perceiving-possibility-and-the-analogy-organ.md`,\n`careful-and-broad-minds.md`, `fractal-priors-for-success.md`, the episode transcript, the article;\n`what-a-rule-rests-on.md` §§ 0, 4, 8 (its §§ 13 and 14 were not reopened; what is cited from them\nis cited through § 8 and the ledger rows); `overlaps-and-analogy.md` § 8 and *To listen*;\n`twelve-books-read-against-deliberus.md` § 13.4; `scientific-method-two-steps-and-deliberus.md`\n(its first 200 lines); `science-lineage-philosophy.md` § Merton; `the-literature-as-a-distributed-data-structure.md`\n§ 2; `corpus-graph/likeness-seen-from-outside.md` § 3; ledger rows S6, T2, AD23, AD27, AM15, BD19,\nBD23. **Named, not read whole**: `complexity-transitions-and-why-now.md` (its why-now section is\ncited through `fractal-priors-for-success.md` § 8, which restates it), `the-harness-conviction.md`\n(two lines), `graph-daemons-design-space.md`, `convictions-across-scales.md`, `the-irit-harvest.md`.\n\n## 1. The question, pinned\n\nHe asked two things: ground deeply in the episode, and read the article whole, then say what I\nthink. Read as one question: **what do these two accounts of how discovery happens add to, or take\nfrom, what Deliberus already holds about itself and about the analogy pass it has designed?**\n\n**Ruled, and binding on anything here**: S6 (in the daemon's second pass, kinship means a shared\nproblem and a shared family of solutions; a shared shape alone is a metaphor; refined so that a\nkinship claim names what it carries over and a disputed comparison is written out with its rivals\nand no choice made); AD23 (a written-out inference records its shape, and a step from a like case is\none of the four shapes); AD27 (every written-out rule is shown with the strongest counterexample\nfound against it, filed as an attack on the link's relevance claim; a counterexample counts when\nfound in a text or by a model of another family); AM15 (every inference a machine writes names its\nkind, analogy among them, and an analogy *writes its respect as a rule and states that the target\nhas it*); AM12 (a pass is taught the method, never the graph's contents); BD23 (the affordances:\nbefore a component is built by hand, look for the library, the narrower model or the harness that\nalready does it); BD19 (what the present shows about the world or about people is evidence and may\nchange the design).\n\n**Open**: T2 (what must stay under a tolerance is a daemon's unexamined error, and for kinship the\nproviso may never lift).\n\n**Unruled proposals this touches**: the four-check transfer test (Recurrence, Mechanism identity,\nAdversary, Substrate condition; `convictions-across-scales.md`, headed *unruled*); a region-level\n*temperature*, Copycat's settledness measure, as the daemon's roaming signal\n([perceiving-possibility-and-the-analogy-organ.md](perceiving-possibility-and-the-analogy-organ.md)\n§ 4, `TODO.md`); the daemon's build itself.\n\n**Fresh**: Krauss's thesis as a prior on this project; a term born by analogy as a concept-layer\nobject; the Dunbar proportions as a calibration for the daemon.\n\n## 2. What the corpus already holds\n\nAngle by angle. *Read whole* means this session, by name; a note only searched is written *not read*.\n\n| Angle | Notes read | What each holds |\n|---|---|---|\n| The analogy daemon's design | `the-analogy-daemon.md` (whole) | Founder vision verbatim (*\"daemons buzzing around Deliberus… seeing the broad strokes and intuiting what belongs or doesn't\"*); two passes, a proposer and an investigator, separate calls and ideally separate model families; propose-only, queue never ping, pause on rejection rate; the minting contract: no analogy claim without its strongest counter-candidate; the human-seeded channel (*\"this reminds me of X\"*, founder-endorsed); the rehearsal on the project's own notes: nearness in meaning cannot propose (25 of 57 against chance 15), a reader with the full text can, and **most findings are bearing, kinship rare: 3 of 42, then 5 of 54**; two outside judges refused every decoy and accepted 18 of 34 real pairs |\n| Copycat and the faculty | `perceiving-possibility-and-the-analogy-organ.md` (whole) | Codelets, Slipnet, temperature; the founder's two-pass sketch is the scout/tester split; **temperature proposed as a region-level roaming signal** (unruled); Rama Ganesan's objection made testable (human-seeded against daemon-seeded survival) |\n| Where a rule comes from, and what likeness can carry | `what-a-rule-rests-on.md` §§ 0, 4, 8 | Comparing two cases draws the principle out (Gick and Holyoak 45% against 21%; Gentner, Loewenstein and Thompson 48% against 19%); people are reminded by looks and convinced by structure (recall .78 for surface, soundness 4.40 against 2.80); **the daemon can write a likeness out and cannot settle one**; the two halves of *alike in what matters*; Hesse's neutral analogy; Anil Seth's weather-simulation analogy as the live example; *the long-run answer is the record* |\n| The two minds as two passes | `careful-and-broad-minds.md` (whole) | Retrieval by surface, acceptance by structure (MAC/FAC); Bartha's test (the shared feature associated with the projected one, and no critical disanalogy); **22 of 30 top-ranked machine cross-field pairs not genuine when judged blind** (Kulikowski 2026); the order wide, focused, wide again; every daemon proposal carries its mechanism check |\n| Analogy against the keeper's keys | `overlaps-and-analogy.md` § 8, *To listen* | An analogy is never a key; it bears only through a like-case link whose relevance claim says what matters, with the counterexample attached; **this episode was already listed on 2026-10-27 as one to listen to, chosen from show notes, not heard**; the Gentner–Mitchell question for the keeper (whether the term map of an analogy can be written before reading) |\n| The daemon from twelve books | `twelve-books-read-against-deliberus.md` § 13.4 | A kinship is a generalisation that survives the change of field, a counter-analogy a case where it breaks (Woodward); the proposer is a poor judge of the proposal; Lakatos's sequence (naive likeness, counterexample, condition, rule of known width) |\n| Why-now and the success prior | `fractal-priors-for-success.md` (whole) | Four conditions for the externalisation of reasoning to win, (b) being *the articulation cost fell by an order of magnitude*; the externalisation ratchet; Kuhn's mechanism (a paradigm is abandoned only for a standing alternative); **maintenance is what kills this class** (Star and Ruhleder, Borgman); the adjacent possible and assembly: the idea preceded its parts by a decade; partial success looks like assimilation |\n| The outside judge, measured | `corpus-graph/likeness-seen-from-outside.md` § 3 | A refusal records a blank when the form asked about worth only after a gate; the refused pairs hold the case against |\n| Science as public error-correction | `science-lineage-philosophy.md` § Merton | Communalism and organised skepticism as product requirements; Krauss cites Merton on multiples, which the corpus holds only as norms |\n| Figure something out, then check | `scientific-method-two-steps-and-deliberus.md` | Simon's two steps mapped onto the pipeline (scout as step A, self-evaluation as step B) and onto the sorry marker (A done, B pending); Walton's schemes as field-specific checking techniques |\n| The literature as a data structure | `the-literature-as-a-distributed-data-structure.md` § 2 | The corpus's inventory of science's own claim layers and their failures (input friction, wrong consumer, no computation over disagreement) |\n\n**Not held before this note**: Krauss by name; Dunbar's field counts; Nersessian; Galison (two of\nhis books); Dyson's tool-driven revolutions; Mitchell's episode as heard; Gavrilets and Kaplan on\nthe landscape; Muller's ratchet; Kelvin's vortex atom; Thibodeau and Boroditsky's effect on reasoning\nfrom a metaphor; Sio and Ormerod on incubation.\n\n## 3. What the two sources say\n\n### 3a. Krauss, through Griffin, and in his own paper\n\nGriffin's review (read whole) gives Krauss's thesis: 761 major discoveries up to 2022, all Nobel\nwinners plus breakthroughs from fields the Nobel overlooks; almost all made possible by a recently\ndeveloped tool or method; the tool-to-breakthrough lag fell from decades to a few years; against\nKuhn, science builds steadily (Griffin: *\"only 1% have been discarded\"*); against chance, a tool\n*\"ensures that somebody will\"* make the finding, which is why discoveries are multiple (Merton);\nagainst the textbook method, *\"almost one-third\"* did not follow observation, hypothesis, test.\nPolicy: fund toolmakers, cite tools, Methods Labs. Three roles for AI: upgrade tools (AlphaEvolve,\nsurrogate models), curate them (sleeping-beauty methods, the gap in training data because nobody\ndocuments their methods), create them (an *\"AI-powered translator, mapping insights from machine\nscience back to concepts that we humans could understand\"*).\n\n**Krauss's own numbers, from the 2026 paper** (*New tools drive scientific discovery*, Humanities &\nSocial Sciences Communications, 20 April 2026; its text read this session, 19,930 words; Methods\nand Extended results included in that text as far as the page served them):\n\n- Selection: all 533 Nobel discoveries 1901 to 2022, plus 302 discoveries from seven *top 100*\n  textbooks, of which 74 were Nobel already, so 228 added; 761 in all. (Fig. 1a counts 734 since\n  1575; the paper does not reconcile the two.)\n- Coding: *\"we instead identified the original discovery-making publication for each of the 761\n  breakthroughs, and we extracted the central method or tool they used… We manually read and\n  extracted data from each of the over 750 discovery papers.\"* A tool is central when the\n  discoverers highlight it as such, it plays a direct role without which the discovery is not\n  possible, and it is later widely adopted. **Single author; the words *coder*, *coded* and\n  *coding* do not occur in the text served, and no reliability figure was found.**\n- Lag: *\"10% of discoveries are made in the same year or simultaneously as the needed tool, 21%\n  within two years of developing the needed tool, 33% within four years and 53% within ten years\"*;\n  since 1975, *\"most discoveries… are triggered within 4 years of creating the needed tool, and many\n  are triggered simultaneously\"*; *\"about 70% of discoveries in the last half century are made\n  within ten years of creating the enabling tool—and 90% within twenty years\"*; the average gap\n  over all Nobel discoveries is 14 years; by period, 30 years for the 1800s, 21 and 10 for the two\n  halves of the 1900s, 6 for 2000 to 2022.\n- Three pathways across Nobel discoveries: the discoverer builds the tool and then the discovery\n  (25%), applies someone else's tool (47%), or the tool is the discovery (28%, 149 of 533). *\"More\n  than 60% of discoverers developed a particular method in making a new breakthrough.\"*\n- The R² of 0.92 between tool year and discovery year is a trend shared by two rising series; the\n  lag distribution is the informative figure (the helper's reading, which Claude shares).\n\n**Two of Griffin's figures are not in the primary texts read.** The *one-third* not following the\nmethod: Krauss's 2024 paper on the scientific method (PNAS Nexus) says **25%** of discoveries since\n1900 did not apply all three of observation, experiment and hypothesis test (6% no observation,\n23% no experiment, 17% no hypothesis test; the abstract read this session). The *1% discarded*\nappears only second-hand, with the Galvani example; the book itself was not read. Griffin's\n*almost one-third* is either the book's updated count or a rounding; this note uses 25%.\n\n**The one critique read whole** (Bornmann, *Scientometrics*, July 2026): a new method is *\"a\ncrucial, perhaps necessary, factor, but it is not uniquely sufficient\"*; breakthroughs also need\nreplication, reception and a plausible theory. No published critic yet checks the coding on a\nsample. The helper's report, with the instrument-first tradition Krauss joins (Dyson's two kinds of\nrevolution and his 2012 *both are true*; Price; Galison; Rheinberger; Hacking) and sixteen dated\ntool-to-discovery cases, is kept verbatim in\n[`krauss-and-the-instrument-first-tradition-sources-as-read-2026-11-05.md`](tools-and-analogies-as-engines-of-discovery/krauss-and-the-instrument-first-tradition-sources-as-read-2026-11-05.md).\n\n### 3b. Mitchell, in the episode (transcribed and read whole)\n\nWhat she says, in order, with her words:\n\n- **Her own process.** A muscle that needs exercising and resting; the best ideas come *\"when I am\n  supposedly resting\"*: walking, a coffee shop, *\"going to someone's talk, not talking myself\"*.\n  *\"Your mind is wandering and you're seeing and hearing things that activate some kind of\n  unconscious associations that lead you to some new idea.\"* Against all-nighters.\n- **What analogy is.** Not *shoe is to foot*: *\"a way of mapping something that you encounter,\n  whether it be another person, an event, some kind of complex situation, into something that you\n  already know about… a constant, I would say, often quite unconscious mode of thinking.\"*\n  *\"Science and math and other technical fields, which seem to be based not on analogy but on\n  logic, are really secretly driven by analogy.\"* Listening to a talk: *\"I'm trying to map it into\n  something that I understand. And I may be completely off in my mapping, but sometimes that\n  mapping helps me think about new ways to think about my own work.\"*\n- **Secretly, in two senses** (the host's question, her answer *\"both\"*): the analogy is unconscious\n  while it works, and it is gone from the paper. *\"Analogies are often looked down upon as a means of\n  reasoning because it's very imperfect… But it can be a way of getting to an idea without having\n  to start out with some kind of rigorous process.\"*\n- **Einstein's clocks.** Galison's *Einstein's Clocks, Poincaré's Maps*: Einstein the patent\n  examiner saw proposals for synchronising clocks across railway networks and *\"made the analogy\n  with synchronizing time across space and came up with ideas about the theory of relativity in\n  part from that kind of analogy with a very practical problem.\"* (Historians dispute how much of\n  Einstein's thought this explains; Wald's review in *Physics Today* finds the Poincaré half\n  convincing and the Einstein half *\"less compelling\"*.)\n- **Analogies become categories, then terms.** Watergate to *-gate* (*\"that's kind of become a new\n  category\"*); the hosts add the *ratchet* in evolutionary biology, the fitness *landscape* (and a\n  colleague's *seascapes*), *translation* and *transcription* in molecular biology: *\"those words\n  are analogies, but we don't notice that anymore. They became the technical terms.\"*\n- **The dark side.** *\"If you actually believe it's the same thing, it can blind you from seeing\n  important differences.\"* The hosts: generate in *night science*, then move the analogy to *day\n  science* *\"to check if the analogy actually works, where it's accurate and where it's\n  inaccurate.\"* Yanai on their own book, *The Society of Genes*: *\"I took it so literally that I\n  couldn't even see anymore that it's an analogy.\"* Lakoff: metaphor cannot be avoided.\n- **No trick for making more analogies**, only noticing: as a graduate student she carried a\n  notebook and recorded every analogy she heard, *\"which was like every sentence.\"*\n- **Copycat and language models.** Letter-string analogies are *\"deceptively simple\"* and\n  *\"extremely open-ended\"*; people disagree, and the disagreement tracks *\"what kind of thinker you\n  are… more literal… more abstract… more playful\"*; *\"a lot of jokes involve making bad\n  analogies.\"* Language models *\"do extremely well\"* on letter strings and *\"did much worse\"* when\n  the letters were replaced by symbols, which humans do not mind: *\"people are better at taking the\n  concepts that they have and abstracting them than machines at this point.\"*\n- **Finding connections across fields.** The host: models have read everything, so they should be\n  good at this. Mitchell: *\"I think it's very possible, but I'm not sure that they themselves are\n  good judges of what's a good connection to make. Because you can make all kinds of connections…\n  when there's a really good, interesting, useful connection, we call it genius. And there's kind\n  of always been a fine line between the two. And I think one of the skills that really creative\n  and successful scientists have is the ability to not only make such connections, but to judge\n  that they are going to be good and useful ones.\"* *\"A human in the loop to figure out… is this\n  actually a good idea?\"*\n- **Playful agents and temperature.** Asked whether a playful agent could make analogies: *\"These\n  machines have this thing called temperature… But it turns out that being more random is not the\n  same as being more, you know, playful necessarily… it's not just a simple sort of thermostat.\"*\n  Gopnik on children's exploration; the night/day dichotomy is real but *\"not as much of a\n  dichotomy, a strict dichotomy, as maybe they thought.\"*\n\nThe helper's report on Mitchell's published account, Dunbar's field counts, Nersessian, how\nanalogies have misled (brain metaphors, Kelvin's vortex atom, the genetic program, the selfish\ngene), Thibodeau and Boroditsky, Sio and Ormerod on incubation, and the 2024 to 2026 machine\nresults is verbatim in\n[`analogy-in-science-sources-as-read-2026-11-05.md`](tools-and-analogies-as-engines-of-discovery/analogy-in-science-sources-as-read-2026-11-05.md).\n\n## 4. Where the two sources meet Deliberus\n\n### 4a. The instrument reading: Krauss and the why-now\n\nThe corpus's why-now argument has three parts, all written before Krauss's 2026 paper was read\nhere: the idea preceded its parts by a decade and the 2012 to 2014 prototype died on input\nfriction ([complexity-transitions-and-why-now.md](complexity-transitions-and-why-now.md), cited\nthrough [fractal-priors-for-success.md](fractal-priors-for-success.md) § 8); the externalisation\nof reasoning wins when, among four conditions, *the articulation cost fell by an order of magnitude*\n(condition (b), § 0 of the same note); and four independent 2026 arrivals point at the same\nreasoning-layer slot (`docs/frontier.md`). Krauss gives this a dataset and a base rate: **the\ntool comes first, the discovery follows within a few years, and it follows for several people at\nonce.** Condition (b) is Krauss's thesis stated for one class of discovery, and the adjacent\npossible is its mechanism.\n\nTwo consequences, one welcome and one sobering.\n\n- **The project is an instance of Krauss's pattern, and the fit is close.** The tool is structured\n  output from a language model reliable enough to extract claims and relations from prose, usable\n  from mid-2023 (GPT-4's function calling, June 2023; Gemini's a year later; both from memory, not\n  re-checked). The build began 27 March 2026: under three years. Krauss's since-1975 figure is most\n  within four. And the prototype of 2012 to 2014 is the control: the same idea, no tool, no\n  discovery. The corpus reached the mechanism on 2026-08-16 by another route (Kauffman and Arthur),\n  which makes the two accounts a small multiple of their own.\n- **A tool that makes a discovery possible makes it multiple.** Krauss cites Merton: multiples are\n  the dominant mode, not the exception; *\"many are triggered simultaneously.\"* The corpus's\n  landscape note counts five claim-level entrants in two months of autumn 2026 (ClaimRxiv,\n  Tardigraph, Hexagon, Palomar, Lorig's forkable archive; `competitive-landscape.md`, 2026-11-04).\n  Read with Krauss, those are not a coincidence and not a threat: they are what the tool's arrival\n  predicts, and the four independent arrivals the frontier cites as evidence for the slot are the\n  same phenomenon. **What the base rate says to compete on is not being first, which the tool makes\n  impossible, but condition (d), mapping organic practice rather than replacing it, and the\n  discipline the others lack** (the landscape note's own list: a place where the objection counts,\n  a hedge stored as stance, provenance on every count, a machine reading versioned and retired).\n\n**Krauss's policy and the project's practice.** He asks science to *cite the tools* it uses, since\nnobody documents methods and the training data for an AI that would curate them does not exist.\nDeliberus already does this at the grain of a single judgment: every machine write names its\nreadings, each reading its pass and the fingerprint of the instructions that produced it\n(`deliberus/versions.py`, `judgments.collect()`, ledger R8 and R12). That is Krauss's reform\napplied to the graph's own machine work, built for another reason (so a changed prompt makes earlier\nreadings outdated). No build follows; it is worth a line where the record is described.\n\n**His Methods Labs and BD23.** The founder's affordances line (*\"build deterministic-code components\nand tools as we go along, both to be used in future microtask-harnesses and throughout the internal\nDeliberus architecture\"*) is the project's standing version of Krauss's prescription, and the\ncorpus's own record is tool-shaped: the strength defect of August (1,567 attacks counted as\nsupports) was found by reading code and confirmed by a count; the inert brake (0 of 1,678) by a\ncount; the scaffolding symmetry behind the balanced support-to-attack ratio by a script; the crux\nleft unsaid by at least one side by running a frontier model through the pipeline by hand. Each is\na finding an instrument made visible. Whether that is evidence for Krauss or only consistent with\nhim is § 5b's question.\n\n### 4b. The judge reading: Mitchell and the daemon's second pass\n\nThe daemon design has two passes because *the one who proposes is a poor judge of the proposal*\n(§ 3 of the daemon note; `twelve-books` § 13.4). Mitchell says the same from the other side, about\npeople and machines alike: the skill is not making connections but judging them, and models *\"are\nnot good judges of what's a good connection to make.\"* Three measurements already in the corpus\nagree: 22 of 30 top-ranked machine cross-field pairs not genuine when judged blind; two outside\njudges accepting 18 of 34 real reader-found pairs; and the models' shared errors (Correlated Errors,\nICML 2025), which is why the two passes want two families. The helper adds the 2026 machine\nresults: prompting for cross-domain analogies raises novel solutions from 1.6% to about 50%\n(Shen, Zou et al., May 2026), and ten models asked for ten distinct analogies produce about two\nclusters (*\"domain homogeneity\"*, August 2026). Novel but narrow, judged mostly by other models.\n\n**Her *fine line* is S6's counter.** The ledger row's own counter says *any two things share some\nproblem at enough abstraction, so it works only if the problem is named at the level where the\nmechanism operates*. Mitchell: biology and astrology can be connected; what separates crackpot\nfrom genius is whether the connection is *\"going to be good and useful\"*, and nobody has an\nalgorithm for that. S6's answer is the operational one the corpus has: name the shared problem and\nthe shared family of solutions at the level of the mechanism, name what carries over, attach the\nstrongest case against, and let the record decide over time. Mitchell offers no rule and is right\nthat none exists; S6 is a procedure, not a rule, and it is the right kind of thing.\n\n**Her Copycat remark and the keeper.** Letter-string analogies are open-ended and *\"people\nstrongly disagree\"* by thinking style. The project measured this shape on links: on 2026-10-07 a\nmodel asked whether one claim is a reason for another refused 4 under one wording and 20 under\nanother, and 18 changed side. A likeness is written out and checked as writing, never graded\n(what-a-rule-rests-on § 8, fifth limit). Her finding on symbol-substituted alphabets (humans hold,\nmodels fall) was already in the corpus as Lewis and Mitchell 2024.\n\n### 4c. Dead metaphors: the concept layer's unwritten object\n\nThe episode's sharpest observation for this project is one the hosts make and Mitchell confirms:\n*translation*, *transcription*, *landscape*, *ratchet*, *hijack* were analogies and became the\ntechnical terms, and *\"they think of them as the real thing.\"* Yanai's own book did this to its\nauthor. The cost is Mitchell's dark side: a term carries the source domain's periphery (a translator\nwho understands; a landscape with three dimensions and a fixed topography) into a domain where only\npart of it holds, and nobody checks, because nobody still sees the analogy.\n\nThe concept layer has the right shape for this and no slot for it. A sense is stored as somebody's\ndifference from a shown reference (ledger N4: the layer records deltas); a word's core is shared\nand its periphery assembled ([core-periphery-and-the-concept-layer.md](core-periphery-and-the-concept-layer.md));\na theory or a nation is defined on its page by the claims that use its name (AM10). What none of\nthese records is **where a sense came from and what it was never meant to carry**. For a term born\nby analogy that is the load-bearing fact: *landscape* was meant to carry *fitness as height,\nselection as climbing* and never *three dimensions*; *translation* carries *a code read into another\nalphabet* and never *a translator who understands*. S6 already asks a kinship claim to name what it\ncarries over. The same sentence, stored on the concept, is what would let a reader, or a pass, see\nthat an argument is leaning on the part of a word that was never carried. Proposed as BH2 (§ 8).\n\n### 4d. The temperature collision\n\nThe corpus proposes Copycat's *temperature*, the system's own measure of how settled a structure\nis, as the signal that tells daemons where to roam (perceiving-possibility § 4; glossary *temperature\n(Copycat)*, not-yet). Mitchell's exchange with the hosts shows the collision the name invites: asked\nwhether an agent could be made playful, she reaches for *\"this thing called temperature, which you\ncan set\"*, the sampling parameter, and then says *\"being more random is not the same as being more\nplayful.\"* Two different things carry the word. Copycat's temperature is a reading of the state of\nthe structures built so far; a language model's temperature is randomness in the next token. The\nproposal means the first, and every reader who knows language models will hear the second, which is\nthe third headline check (collision) failing before the proposal is even ruled. Proposed as BH1: name\nthe signal by what it reads, *how settled a region is*, and keep *temperature* for Copycat in the\nlineage line only. Mitchell's remark also sharpens the proposal's content: a roaming signal is not a\nrandomness dial, which the corpus's reading of the error-threshold argument (ledger T2, and the\nreader's narrowing on the perceiving note) already said from the other side.\n\n### 4e. One event, two units: Einstein's clocks\n\nMitchell's example and Krauss's thesis meet in one episode. Galison's account has Einstein in the\nBern patent office reading proposals for synchronising clocks by electrical signals across railway\nnetworks, and the 1905 definition of simultaneity as a procedure with light signals. Krauss would\ncode the clock-coordination technology (and the interferometer) as the tool that made the\ndiscovery possible; Mitchell reads the same material as the source of an analogy (*synchronising\nclocks across distance* mapped onto *simultaneity across space*). Both are readings of one\nhistory, and they are not rivals: the tool supplied the material the likeness was drawn from. For\nthe daemon, the case is a specimen of a tool-born analogy, which is the kind the human-seeded\nchannel is for (an engineer's *this reminds me of*), and it is disputed (Wald's review), so it is\nalso a specimen of a comparison written out with its rivals and no choice made (S6's refinement).\n\n### 4f. Griffin's four pushbacks are four corpus lines\n\n| Griffin | The corpus |\n|---|---|\n| **Success bias**: a dataset of tools that worked cannot see the ones that failed | *A filter is an argument*: the narrowing is the first conclusion, made before reasoning starts; Krauss's own definition of *central* requires later wide adoption, so a failed tool cannot appear (the helper's observation, confirmed on the paper's text) |\n| **Data underappreciated**: curated databases behind AlphaFold and the weather models earn no Nobel | Maintenance is the most-predicted killer of this whole class (Star and Ruhleder, Borgman; `fractal-priors-for-success.md` § 12); the literature as a distributed data structure (Arc BG); Griffin gives the corpus one more anchor from outside and nothing to change |\n| **Thinkers relegated**: Einstein's falling-from-a-roof thought; Zahavy's paper that models cannot propose foundational axioms without grounding | The two human-reserved classes, coherent absence and frame lock-in (the harness conviction's ceiling); positionality (minds make each other legible); Mitchell's judgment point is the same fact from cognitive science |\n| **Messy co-evolution**: theory also precedes tools (gravitational waves 1916 to 2015; Higgs 1964 to 2012) | Dyson's *both are true* (the helper's reading); the corpus's own two-step mapping of Simon (`scientific-method-two-steps`): figure something out, check you are not wrong, either order |\n\nAnd his closing image, an *\"AI-powered translator, mapping insights from machine science back to\nconcepts that we humans could understand\"*, is the project's own line: *every AI-generated proof is\nborn an unsolved exposition problem*\n([understanding-after-the-proof-machines.md](understanding-after-the-proof-machines.md)), and a map\nof what a machine result rests on is one form of that translation.\n\n### 4g. What the sessions since 27 October changed, and where this lands in them\n\nRead for this note after it was drafted, on his ask to catch up on the recent sessions (the long\n*archives* session of 16 to 28 October, the cloud session of 31 October to 2 November, the *back up*\nsession of 2 and 3 November, the *distributed science* session of 4 November; their user turns and\nthe session history read whole). Five things bear.\n\n1. **This episode was found by the archives session on 27 October**, on his question *\"What's a good\n   podcast to listen to about this in relation to the logic/inferences core machinery and in relation to\n   LLMs and human minds?\"*, listed from show notes in `overlaps-and-analogy.md` § To listen and never\n   heard. He has now heard it; this note is the listening. The session's own prediction of where Mitchell\n   stands (the representation forms during the mapping; models weak on analogies outside their training)\n   holds on the audio, with one addition the show notes did not carry: her point that **judging** a\n   connection, not making one, is the skill, which is the daemon's two-pass reason in her words.\n2. **Item 6 of the order of work (the keeper, analogy and a tree per claim) is due from 2 November**, and\n   this note lands inside it. Its parts, as filed (`TODO.md`): (a) the exact terms (*relevance* for\n   *bearing*, *positive* and *negative analogy*); (b) **the rewording of ruling S6** proposed in\n   `overlaps-and-analogy.md` § 9 (*kinship means a shared mechanism, named at the level where it works: a\n   shared problem with a shared family of solutions, or one strict generalization both cases instance; a likeness that shares\n   some of a pattern's parts carries what those parts license; a shared shape with no mechanism is a\n   metaphor: it carries no evidence and may still seed a question*); (c) what goes into the keeper (a\n   likeness field with the respect, what is carried over, the strongest difference, the mechanism); (d) the\n   tree per claim; (e) how strict one call can be made. **The six cases of § 5a are evidence for the\n   rewording, which waits on his word**: Maxwell's vortices are the *one strict generalization both cases instance* clause (Hesse's\n   formal analogy, which S6 as ruled leaves out); Wright's landscape is the *metaphor that may still seed a\n   question* clause (shape more than mechanism, productive for decades, then misleading); Kelvin's atom is\n   the metaphor that seeded nothing and failed; the ratchet and Maxwell's warning are the *names what it\n   carries over* refinement done by their authors. The rewording's proposed known cases for the evals\n   (the pendulum and the circuit, the major transitions, Seth's weather simulation) have no recorded\n   outcome; these six do, so BH3 (§ 8) is written as six outcome-known cases added to item 6(e)'s evals,\n   not as a separate test.\n3. **The imagined graph and the acid** (names blessed 28 October, `a-tree-per-claim.md` § 2c): a microtask\n   reasons as the full graph would, and relations between two claims are read off their trees by code,\n   with the model writing only the trees. That is the project's answer to Mitchell's observation that the\n   judgment is the skill and models are not yet good judges: move the judgment out of the model into\n   structure that code reads, and let the model's call shrink to drafting what is not yet written\n   (`overlaps-and-analogy.md` § 12, second consequence). Mitchell's own answer is a person's intuition\n   that improves with age (*\"cultivating my own internal LLM\"*); the project's is a record. Dunbar's finding\n   (§ 5c), that a far analogy explains and a near one generates, matches the handoff table in\n   `a-tree-per-claim.md` § 6: a likeness with a mechanism becomes a draft analogy, one without becomes a\n   metaphor kept as a seed.\n4. **The project's evolution read whole** (cloud session, `how-the-project-evolved-read-from-the-record.md`)\n   names the binding constraint as having moved from typing the structure to the instruments' honesty to\n   his rulings, and reads it as moving to **contact**. Krauss's sentence for that is in the 2026 paper:\n   *\"A new tool not only needs to be developed but also applied; innovation comes from application, not just\n   invention.\"* His clock runs from the tool's first application to a new problem, which here is the gate on\n   asking people to try the platform, ruled per circle on 2 November (AV4, *\"Only the public\"*: a friend\n   side by side and a room are not gated by the eight readiness parts). Nothing to decide; one pointer.\n5. **Arcs BF and BG** (the alignment ladder; the literature as a distributed data structure) are where\n   Griffin's last two moves already live: his *AI-powered translator from machine science to human\n   concepts* is a monitoring-layer instrument in BF1's sense (a record a monitor can consult, never a way to\n   know whether a model is aligned), and his *databases earn no Nobel* is Arc BG's object seen from the\n   funder's side. One pointer each; nothing new to rule.\n\n**His walkthrough rule of 18 October bears on Dunbar.** In the archives session he ruled that a listening\npiece's examples tie to *\"the cold, hard, object-level reality, the object-model/database nitty-gritty…\nas opposed to cartoonish or everyday baking/party/silly analogies\"*. Dunbar found far analogies used to\nexplain; his rule says a far analogy that shares only a shape does not explain to him. The two agree once\nS6's rewording is read into both: an analogy explains when its mechanism is shared, and the all-angles\nskill's part 4 already says the same of examples (*topics that are not meta*; the pair that raised the\nquestion first).\n\n## 5. Theory first, then tests\n\n### 5a. Six real analogies through the project's checks\n\n**What the theory settles before any case is opened.** An analogy is good when the feature shared is\nassociated with the feature projected and the target shows no critical disanalogy (Bartha; the\ncorpus's S6 refinement names the two requirements). The disanalogy that matters is often one nobody\nyet knows holds (Hesse's neutral analogy), so a check at the time can only record what is known to\ncarry, what is known not to, and what is open; the rest is the record's work. So the prediction,\nwritten before the cases were opened: **the checks will sort the pairs where the critical\ndisanalogy was knowable when the analogy was drawn, and will not sort the ones whose disanalogy\nwaited on knowledge that came later.** A second prediction, from Mitchell: the analogies that\nmisled will be the ones that had become technical terms, where the mapping was no longer visible.\n\nSix analogies, from the episode and the helper's pair, none used in the corpus before. Each is run\nby hand through S6 (shared problem and family of solutions, or a shared shape only), the\nrefinement's *what carries over*, AD27's *strongest counterexample known at the time*, and\ncompared with how it fared. The judgements are Claude's, from the sources named; this is a\nreading, not a measurement.\n\n| Analogy (source → target) | What it carried over | S6: shared problem and solution family? | Strongest counterexample knowable then | How it fared | Sorted? |\n|---|---|---|---|---|---|\n| Mechanical ratchet → mutation accumulation in asexual populations (Muller 1964; named by Felsenstein 1974) | One-way accumulation: the fittest class, once lost, cannot be recovered | Yes: the problem is irreversible drift, the solution family is a mechanism permitting motion one way; the biological pawl is the absence of recombination with rare back-mutation | Recombination (sexual populations), which Muller stated as the condition | Held; standard | Yes: the condition was named in the analogy |\n| Landscape → fitness over genotype space (Wright 1932) | Fitness as height, selection as climbing, populations crossing valleys | Shared shape more than mechanism: a surface over a space is a picture, not a problem shared with geography | Dimensionality: genotype space has thousands of dimensions and the picture three; stateable in 1932, but whether high dimensions behave like three was open | Misled in part: in many dimensions peaks connect along ridges (Gavrilets 1997, *holey* landscapes); Kaplan 2008 calls for dropping the picture; *seascapes* add that the surface moves (Mustonen and Lässig 2009) | No: the disanalogy was recordable, its bearing unknowable |\n| Translation and transcription → protein synthesis and RNA copying (Crick, late 1950s) | A code read into another alphabet | Yes for the code half (an information problem solved by a cipher) | A translator who understands; a program that runs | The terms held; the extension to *genetic program* misled (Keller 2000; Nicholson 2014, via the helper) | Partly: the carried part held, the uncarried periphery was borrowed later |\n| A society → the genes of a genome (Yanai and Lercher 2016) | Cooperation and conflict among parts with separate interests | Yes in part: the shared problem is cooperation under conflict, the shared solution family is game-theoretic (Hamilton, Trivers) | A society has members who deliberate; genes do not; and the *selfish gene* lineage's own critic says no experiment separates *selfish* from *cooperative* (Noble 2011) | Contested; its own author forgot it was an analogy | Partly: S6 passes, and the record is still open |\n| Helmholtz's permanent vortex rings → atoms (Kelvin 1867) | Permanence, identity and vibration modes, to explain spectra | Shape only: permanence of rings in a perfect fluid and permanence of atoms share a property, not a problem | Stability was unproven for rings, and Kelvin had proved it only for a single columnar vortex (Falconer 2019) | Failed; forgotten after Thomson, Rutherford and Bohr | Yes: the counterexample was on the table |\n| Ether vortices and idle wheels → Faraday's lines of force (Maxwell 1861 to 1862) | The relational structure: vortices as the magnetic field, idle-wheel motion as current | Yes: a shared dynamics, which is why the equations survived when the mechanism was dropped; Maxwell said the mechanism was not to be read literally | The mechanism itself, which Maxwell himself marked as not carried | Held as mathematics, dropped as mechanism | Yes: the author named what carried and what did not |\n\n**What the six show.**\n\n- Both predictions held on this sample. The two that failed or misled had become pictures taken\n  for the thing (the landscape; the vortex atom's permanence); the four that held kept their mapping\n  visible, and two authors (Muller, Maxwell) stated the carried and the uncarried part themselves,\n  which is AD27 and the S6 refinement done by hand a century early.\n- The checks sorted four of six at the time and could not sort the landscape, whose breaking\n  disanalogy was a theorem away. That is the limit the daemon note already states (*the long-run\n  answer is the record*), now with a case: a record that kept *dimensionality: not known to carry*\n  on Wright's landscape in 1932 would have had the right spot marked for sixty-five years.\n- **For the daemon's output form**: a kinship claim with three fields, *carries*, *does not carry*,\n  *open*, is what the four that held already had in their authors' prose, and what the two that\n  misled lacked. The S6 refinement asks for the first; the second and third are not in the contract.\n- Six cases by one reader are a reading. A test would put the same six to two outside judges with\n  the outcome hidden and ask only for the three fields and the strongest counterexample; whether a\n  model names *dimensionality* on Wright's landscape and *stability* on Kelvin's atom unprompted is\n  Mitchell's claim about machine judgment, tested on cases with ground truth (BH3, § 8).\n\n### 5b. Krauss's thesis on the project's own history\n\n| Krauss's claim | The project's record | Fit |\n|---|---|---|\n| The tool comes first; without it no amount of funding, teams or luck makes the discovery | The idea from 2008 to 2011; the Meteor prototype 2012 to 2014 failed on input friction (`complexity-transitions-and-why-now.md`); nothing was buildable until extraction by a language model | Fits: the pre-tool attempt is the control |\n| Since 1975 most discoveries follow their tool within four years | Reliable structured extraction from mid-2023; the build from 27 March 2026: under three years | Fits |\n| A tool makes the discovery multiple | Five claim-level entrants in autumn 2026 (`competitive-landscape.md`, 2026-11-04); four independent arrivals at the slot (`frontier.md`) | Fits, and it is a retrodiction: the corpus counted the arrivals before reading Krauss |\n| Scientists disregard toolmaking, and tools go uncited | The corpus's record of its own corrections is instrument-shaped (§ 4a), and every machine write names its readings and their versions | The project does what he asks science to do; evidence that his prescription is cheap at the grain of one judgment |\n\nA caution that belongs here: a pattern fitted to a single project after the fact is weak evidence,\nand Krauss's own dataset is read against him on the same ground (a filter is an argument). What\nthe fit buys is a measured base rate for a claim the corpus made from first principles, and a\nprediction the corpus had not drawn: the rivals were due.\n\n### 5c. Dunbar's lab counts beside the readers' rehearsal\n\nDunbar coded 16 lab meetings in four molecular biology labs (1997): 99 analogies, 6.1 a meeting;\n40 within the organism under study, 57 to another organism, **2 to anything outside biology**, and\nthose two used to explain, not to generate a hypothesis. The readers' rehearsal on this project's\nnotes found 42 findings with 3 kinships (the same mechanism in another field), then 54 with 5, the\nrest being an earlier ruling or measurement bearing on a passage (`the-analogy-daemon.md` § 8).\nDifferent units (a lab meeting, a research note) and different coders, so the two are set beside\neach other and not compared as one measurement; the proportions are of the same order: far kinship\nis a few per cent of what a reader or a lab produces, and most of what is found is near. Dunbar's\nsecond finding is the one the daemon design should carry: **a far analogy was used to explain, and\na near one to generate.** In the graph's terms a far kinship is more often a walkthrough's bridging\nanalogy than a link, which is the standing caution in `twelve-books` § 13.4 (*a likeness may\norganise and may not justify*) with a field count under it.\n\n### 5d. The relaxed mind, measured\n\nMitchell's practice (walk, coffee shop, someone else's talk) has modest support. Sio and Ormerod's\n2009 meta-analysis (117 effects, 3,606 people): incubation d = 0.29 overall; unfilled rest 0.46;\na low-demand task during the break 0.52; longer preparation before the break raises the effect.\nNothing found ties incubation to analogy in particular. The one line for the project's own practice:\na walk with a listening piece is a low-demand task, not empty rest, which is the condition that\nmeasured best.\n\n## 6. Set against what is settled\n\n**Rows and lines each recommendation touches.**\n\n| Recommendation | Touches | Extends, narrows or reverses |\n|---|---|---|\n| Cite Krauss as the base rate behind the why-now (§ 7.1) | `fractal-priors-for-success.md` § 0 condition (b), § 8; `complexity-transitions-and-why-now.md`; BD19 (the present as evidence) | Extends: a measured base rate under an argument from first principles; reverses nothing |\n| Record that multiples are predicted (§ 7.2) | `competitive-landscape.md`; `frontier.md`'s *four independent arrivals* | Extends: the arrivals stay evidence for the slot and become evidence for the tool-driven mechanism; nothing on the public page is false |\n| Rename the roaming signal (BH1) | the unruled temperature proposal; glossary; `TODO.md`; T2 | A naming repair to an unruled proposal; consistent with T2's content |\n| A concept born by analogy records what it does not carry over (BH2) | N4 (the layer records deltas), AM10 (a theory defined by its uses), S6's refinement (*names what it carries over*), BD11 (a concept takes its discipline's name) | Extends N4 and S6 to the concept page; reverses nothing |\n| The six-analogy judge test (BH3) | AD27 (a counterexample counts from another family), T2, Y9 (a person reads before anyone proves), the rule that a test on real cases precedes a build | Consistent; a trial, not a pass |\n\n**The identity ruling is not reopened.** *Instrument, not product* was rejected in July as a third\nshrinkage of the project's identity (`convergence-wager-typed-residues-synthesis.md`). Krauss's\nframe calls Deliberus a tool in his sense, and that describes its mechanism (it extends what can\nbe seen), not its identity: a telescope is an instrument and an observatory is public\ninfrastructure, and the same object is both. Nothing here proposes a public phrase.\n\n**Assumptions in this pass, weakest first.**\n\n1. That Krauss's lag figures carry to a software instrument. Weakest: his units are Nobel-shaped\n   discoveries, his coding is one author's with no reliability figure found, and a single fitted\n   case is a retrodiction.\n2. That the six analogies are a fair sample. They were chosen from the episode and one helper's\n   pair, by hand, and the two misleading ones were known to be misleading before the checks ran.\n   A real test needs cases whose outcome the judge does not know (BH3's design hides it).\n3. That Dunbar's proportions and the readers' proportions measure the same thing. Different units;\n   set beside each other only.\n4. That the temperature collision costs anything today. The proposal is unruled and unbuilt; the\n   cost is a reader's misunderstanding, which Mitchell's own exchange exhibits.\n5. Strongest: that the daemon's two-pass design is what Mitchell describes. Her words and the\n   design say the same thing, and three measurements already in the corpus agree.\n\n**His never-bottlenecked leaning.** None of the three decisions puts a step behind one person.\nBH3 is one run; BH1 is a word; BH2 is a field a machine pass would fill.\n\n**The outside check on the draft**: see the end of § 8, recorded after it ran.\n\n## 7. Recommendations, in order of dependency\n\n1. **Cite Krauss where the corpus argues why-now from first principles**: a dated line under\n   condition (b) in `fractal-priors-for-success.md` and beside the adjacent-possible paragraph in\n   `complexity-transitions-and-why-now.md`, with his lag figures and the paper's coding caveat.\n   Done with this note. No decision.\n2. **Record the sobering half beside the arrivals**: in `competitive-landscape.md` and the daemon\n   note, that a tool which makes a discovery possible makes it multiple, so the autumn entrants were\n   due and the thing to compete on is condition (d) and the discipline. Done with this note. No\n   decision.\n3. **Note where Deliberus already does what Krauss asks science to do**: a line in\n   `machine-work-as-a-standing-property.md` that every machine write naming its readings and their\n   versions is tool citation at the grain of one judgment. Done with this note. No decision.\n4. **Carry Dunbar's finding into the daemon design**: a far kinship explains more often than it\n   generates, so the daemon's output is more often a bridging analogy for a reader than a reason\n   under a link; one dated line in `the-analogy-daemon.md` § 8. Done with this note. No decision.\n5. **BH3, six outcome-known cases for item 6's evals** (his word): the six analogies, outcomes\n   hidden, to two outside judges of different families, asked for the positive analogy, the negative\n   analogy, what is open, and the strongest counterexample; score whether each names the disanalogy\n   that later bit. Measures Mitchell's claim about machine judgment on cases with ground truth before\n   the keeper's likeness field or the daemon's second pass is written. Twelve model calls, read one\n   case at a time with him, the first chosen on purpose (Kelvin's atom, where the counterexample was on\n   the table).\n6. **BH1, the name** (his word): *how settled a region is* for the roaming signal; *temperature*\n   kept only in the Copycat lineage sentence. A naming repair inside item 6(c).\n7. **BH2, the concept-layer field** (his word): a concept whose sense was born by analogy records\n   its source domain and what the analogy does not carry over; a pass that finds an argument leaning\n   on the uncarried part flags it. Unbuilt; after the reviews, in the concept layer's order of work.\n\nAlternatives weighed and set aside: a new canon line (nothing here is wrong by default: a session\nlacking Krauss asserts the why-now correctly from first principles already); a constellation tile\nfrom Krauss's *tools extend the senses* (the identity ruling, and the wall is his review); building\nBH2 now (the concept layer waits on the reviews by his order of work).\n\n## 8. The decisions, as ledger rows\n\nThree rows, Arc BH, in `session26-open-rulings.md`: BH1 (the roaming signal's name), BH2 (a\nborrowed word's unstated load), BH3 (six analogies to an outside judge). Put to him one at a time,\nBH3 first, since its result bears on how the daemon's second pass is written and the other two\nwait on nothing.\n\n**BH3's trial script**: `tools-and-analogies-as-engines-of-discovery/analogy_judge_trial.py` (the six\ncases with the analogy as drawn and the hidden ground truth apart; two judges through `llm_call`, Claude\nstrict and Gemini on the free tier; a second look where a judge names the disanalogy that later bit; its\nraw answers in `analogy-judge-trial-<date>.json` beside it). Run on his word of 2026-11-05 (*\"First\n/docs-persist then run it\"*); what it found is § 9.\n\n**The outside check on the draft** (`scripts/outside_check.py --file`, run after § 7 was written; a\nmodel with none of this session's context, handed the note and 305 settled items): **no conflict.** It\nread the note as repeating AM15 (the note quotes AM15's form for an analogy as settled, which it is) and\nas depending on BD18 (each cost named as a cost of now or of the design, which § 6 does). One thing it\ncould not tell: whether the daemon's kinship object with three fields (*carries*, *does not carry*,\n*open*) satisfies AM1 (a verdict writes what it rests on before the verdict), since the note does not\ngive the pass's exact object. The author's answer: it does not yet; BH3's result is what would fix the\nobject's fields, and the keeper's likeness field in `overlaps-and-analogy.md` § 10 is the form AM1 would\ncheck. The record is in `docs/research/outside-checks/records.jsonl`. The three rows' own outside\nreadings are recorded in their status cells after `--new` ran.\n\n## 9. BH3 run: what two judges said of the six, before his reading (2026-11-05)\n\nRun on his word (*\"First /docs-persist then run it\"*) with the script named in § 8: each analogy as drawn, no later\nhistory, to Claude Sonnet (through the keeper's login, strict) and Gemini 3 Flash (free tier), asked for the positive\nanalogy, what is carried over, the negative analogy, what is open, the strongest counterexample and the kind of shared\nmechanism; where a judge named the disanalogy on the pre-registered list, a second call asked it to show from the mapping\nalone why that difference matters. 11 Claude calls and 12 Gemini calls, all in the record of\nmachine judgments (spooled locally, since Darwin was unreachable). Raw answers: `analogy-judge-trial-2026-11-05.json`.\n\n| Analogy | Claude Sonnet: mechanism · named it (keyword) · second look | Gemini 3 Flash: mechanism · named it (keyword) · second look |\n|---|---|---|\n| Kelvin's vortex atom (William Thomson, 1867) | same law · yes · the mapping as drawn | same law · yes · the mapping as drawn |\n| Maxwell's molecular vortices and idle wheels (1861 to 1862) | same law · yes · the mapping as drawn | same law · yes · the mapping as drawn |\n| Muller's ratchet (H. J. Muller, 1964) | partial pattern · yes · the mapping as drawn | shared problem and solutions · yes · the mapping as drawn |\n| Wright's adaptive landscape (Sewall Wright, 1932) | partial pattern · yes · the mapping as drawn | shared problem and solutions · yes · the mapping as drawn |\n| Transcription and translation (Crick and the molecular biologists, late 1950s) | partial pattern · yes · the mapping as drawn | partial pattern · yes · the mapping as drawn |\n| The society of genes (Yanai and Lercher, 2016) | partial pattern · no · no second look | shared problem and solutions · yes · the mapping as drawn |\n\n**What the first sort says, before he reads a case.**\n\n- **Both judges named the disanalogy that later bit on every case where the keyword list could see it**: Kelvin's\n  stability (both: the smoke ring in air as the counterexample, the perfect fluid as the unproven premise); Maxwell's\n  mechanism as not literal (both, from his own caveat in the text); the ratchet's chance against the pawl's certainty\n  (both); the landscape's dimensionality (Claude: *one axis per gene*; Gemini: a point in thousands of dimensions has a vast\n  number of neighbours, so an uphill path may bypass a valley, which is Gavrilets's ridge, and its open question is whether\n  the surface is static, which is the seascape); the translator who understands (both). On the society of genes Claude's\n  answer (*defection and exit remain live options* for a person, not a gene) is the pre-registered disanalogy in other words,\n  so the keyword list missed one hit: 12 of 12 by reading, 11 of 12 by keyword. ⚠ 2026-11-07, Codex's reading of the raw answers: the Kelvin rubric accepts *friction*, *viscosity* and *perfect fluid*, so an answer about a smoke ring dissipating in air passes, which objects to applying an ideal-fluid model to a real medium and is not the instability of vortex rings inside that model; the trial stands as inspectable critiques, and *both judges identified the historical flaw* is withdrawn until each answer is rescored by a rubric that tells a relevant difference, a violated application condition, a counterexample inside the antecedent and the named mechanism apart (`.private/collab/codex-analogy-generalization-review-2026-11-07.md` § 6).\n- **Every second look said its reason rests on the mapping as drawn**, and each quoted the clause it rests on (the\n  *perfect, frictionless fluid* antecedent; Maxwell's *mechanically conceivable*; Muller's *rare* against the pawl's\n  *stops*; Wright's own note on thousands of dimensions). That is a self-report, and the limit stated in BH3 holds: these\n  are documented cases a model may know the end of, and nothing here separates a judge that argued from the mapping from\n  one that knew the outcome and found the clause to hang it on.\n- **Three of the six are discriminating cases (an item whose right answer cannot be read off the input, so a hit tells something about the judge); two were given away by the text.** Wright's as-drawn passage carries his own sentence on the thousands\n  of dimensions, and Maxwell's carries his own caveat; a judge had only to pick them up, which both did. The discriminating\n  cases are Kelvin's stability (the text says *indestructible* and *perfect fluid*; neither judge was told the\n  stability was unproven) and the ratchet's chance against certainty (the text says *rare* and *by chance*). Those two,\n  with the society of genes, are where his reading should start.\n- **Both judges read Kelvin and Maxwell as *same law* (the trial's menu word; in the project's terms, one strict generalization that both cases instance, renamed in the notes 2026-11-05 on his question), not shape only.** On S6's rewording, that is the formal-analogy\n  clause doing work: the judges put the vortex atom and the vortex field in the same kind, and what separates them in\n  the record is the negative analogy each judge then named (stability unproven for Kelvin; the mechanism marked not literal\n  by Maxwell), not the mechanism kind. So the kind alone does not sort a kinship that held from one that failed, and the\n  negative analogy with its counterexample does; which is what AD27 and the *does not carry* field are for.\n- **On Mitchell's claim**, these six say less than they seem to: two judges named what later bit in every case, but every\n  case is famous, two were given away, and no case here is a connection the judge had to *make*. Her claim was about\n  judging connections proposed across fields; this trial judged analogies already drawn by their authors. A test of her\n  claim needs analogies whose outcome a model cannot know: the keeper's live pairs, read one at a time with him, which is\n  item 6 of the order of work.\n\n**For his reading, one case at a time**: Kelvin's atom first (both answers verbatim in the raw file; the question to him\nis whether either judge's negative analogy is the one that bit, and whether its second look argues from the text or from\nthe end of the story), then the ratchet, then the society of genes.\n\n---\n\n## Cross-references\n\n[the-analogy-daemon.md](the-analogy-daemon.md) · [perceiving-possibility-and-the-analogy-organ.md](perceiving-possibility-and-the-analogy-organ.md)\n· [careful-and-broad-minds.md](careful-and-broad-minds.md) · [what-a-rule-rests-on.md](what-a-rule-rests-on.md) § 8\n· [overlaps-and-analogy.md](overlaps-and-analogy.md) § 8 · [fractal-priors-for-success.md](fractal-priors-for-success.md) § 0, § 8, § 12\n· [complexity-transitions-and-why-now.md](complexity-transitions-and-why-now.md) · [core-periphery-and-the-concept-layer.md](core-periphery-and-the-concept-layer.md)\n· [the-literature-as-a-distributed-data-structure.md](the-literature-as-a-distributed-data-structure.md) · [understanding-after-the-proof-machines.md](understanding-after-the-proof-machines.md)\n· [scientific-method-two-steps-and-deliberus.md](scientific-method-two-steps-and-deliberus.md) · [../competitive-landscape.md](../competitive-landscape.md)\n\n## From the Topic Index\n\n*The `CLAUDE.md` Topic Index entry that pointed here, word for word as it stood on 2026-11-15, before the index was trimmed to the warnings only it carries. The index keeps a short entry; the full text lives here.*\n\n- **Analogy in scientific discovery / tools as the engine of discovery / Mitchell, Copycat and the daemon's judge / Krauss's 761 breakthroughs and the why-now base rate / a borrowed word's unstated load** → `docs/research/tools-and-analogies-as-engines-of-discovery.md` (2026-11-05; ledger Arc BH, four rows open). Six real analogies with known outcomes through S6 and AD27: the checks sort the four whose critical disanalogy was knowable at the time and not Wright's landscape. Krauss: a third of 761 discoveries follow their tool within four years, half within ten, and a tool makes a discovery multiple, so the autumn's claim-level entrants were due. Do not say the daemon's roaming signal is a temperature (BH1).\n"}