{"path":"listening/2026-11-06-where-the-analogies-go.md","content":"# Where the analogies go\n\n- **What it is**: a listening walkthrough of about fifteen minutes, made on the morning of 2026-11-06, covering the\n  night of 5 to 6 November since *Judging the connections* (5 November) and where it points: how an analogy becomes\n  a generalization with its cases kept in the graph and read again, and what the November credit should buy. His\n  request: *\"Could you make me a 15 minute walkthrough that talks about what happened exactly since the last one and\n  thinks deeply about where we could go from here, how we could best use the credit we have now, and how we could get\n  to the point where analogies have their generalizations' cases persisted in the graph and scrutinized regularly\n  etc\"*, then *\"make sure it diligently avoids plainwashing\"*.\n- **Covers**: [../research/overlaps-and-analogy.md](../research/overlaps-and-analogy.md) § 15,\n  [../research/priors-as-standing-work.md](../research/priors-as-standing-work.md) § 11,\n  [../research/machine-work-as-a-standing-property.md](../research/machine-work-as-a-standing-property.md) § 7 and the\n  night's record in `TODO.md` (part one); [../research/the-analogy-daemon.md](../research/the-analogy-daemon.md) § 10,\n  [../research/what-a-rule-rests-on.md](../research/what-a-rule-rests-on.md) §§ 3, 8,\n  [../research/what-a-case-counts-for.md](../research/what-a-case-counts-for.md) §§ 5, 7,\n  [../research/a-tree-per-claim.md](../research/a-tree-per-claim.md) § 6 (part two);\n  [../research/whose-tokens-pay-for-mapping.md](../research/whose-tokens-pay-for-mapping.md) § 3h (part three).\n- **Chapter list for noting** (rule 33): 0. Opening · 1. The night, exactly (the review chain became the builds, the\n  graph was filled, and the candidates wait for you) · 2. An analogy becomes a class with its cases (a judged kinship\n  is minted as a generalization that is also a reference class, and the analogy is read, never stored) · 3. What the\n  credit buys (the owed sweeps every night under a cap; extraction waits on a measured pass) · Close.\n- **Registers used**: Lucy; his clone; Scotty (the critic of a night's hunt that found one arithmetic again and again,\n  in a borrowed voice; `overlaps-and-analogy.md` § 15a); Elliot (Mill and Goodman on where a generalization comes from,\n  Reichenbach on the reference class; `what-a-rule-rests-on.md` § 3, `what-a-case-counts-for.md` § 7); Anna (the\n  class on the wheel, cards sharing a concept side by side, ledger AR9; the demo's standing-work condition, `TODO.md`\n  item 2); Hope (what a stranger hears in *the membership pass*); Myrddin (one saying a part, each with its fit note;\n  one sound).\n- **Predictions**: from the prepare call's `--said` on five questions, read whole and weighed newest first: his typed\n  words of 2026-10-22 on spending, 2026-10-25 on generalizations with cases and exceptions, 2026-10-27 on standing\n  work, 2026-11-05 and 06 on filling the graph and on the hold, and the group post of 2011-12-29 on automation then\n  microtasks.\n- **His words quoted** (each checked against the record): 2026-11-05 23:46, *\"As for how to spend credits…\"*;\n  2026-10-25, *\"Btw couldn't this whole thing be seen as generalizations…\"*; 2026-10-22, *\"We do need to get the\n  passes reliable enough first before spending much money.\"*; 2026-10-27, *\"And I think I would want the microtasks\n  and sweeps to be working as standing work…\"*; 2011-12-29, *\"The general strategy is to leverage automation as far as\n  possible…\"*.\n- **⚠ Corrected after Codex's review of 2026-11-07** (`.private/collab/codex-analogy-generalization-review-2026-11-07.md` § 4.2): part two counts the resilience pair's\n  two claims as two conforming cases of the minted generalization. They are general claims (a reduced risk; an\n  intended benefit), which lie within it as narrower reference classes and bring no observed case; the count waits\n  for bundles of observed cases with outcomes, and the arithmetic spoken there is withdrawn as a strength\n  (`cases-and-generalizations-from-both-ends.md` § 4b, § 7).\n- **The correction spoken** (rule 37): none. The *law* correction of 5 November he made himself and heard answered\n  then; the night's figures (the router's $113.56 across both grants, about $64 of it from the lapsed one; the Console\n  at $160.30 of the November $200 at 01:50) stay here, with the ten verdicts' counts (six kinship, four neither) and\n  the hunt's (fifteen proposals from eleven of fourteen pairs; eight read by the keeper, one candidate).\n- **⚠ Corrected 2026-11-06, after he heard part two**: *class* alone in the spoken text and the chapter list means *reference class*, the project's term (a `:Claim` with `claim_kind: 'reference_class'`); his correction: *\"Could we make it a habit to say 'reference class' if that's what we mean?\"*, the second plainwashing instance in two days. The spoken lines stand as rendered; the proposals' homes say *reference class* throughout.\n- **Numbers kept**: part one, *about thirty texts*; part two, *cases from four fields*; part three,\n  *about a hundred and fifty dollars*, *two thousand pairs*: each the idea (what the night did; what a count begins on;\n  what the credit is and what is owed).\n- **Status**: written 2026-11-06 in the morning; the proposals it carries were filed first (`the-analogy-daemon.md`\n  § 10; ledger BH5 and BH6, open, outside-read; `whose-tokens-pay-for-mapping.md` § 3h). Four dry runs, three rewrite rounds\n  before the gates passed; rendered 2026-11-06 at 09:38, 23.5 minutes (16.9 of speech, 6.6 of added silence), 2,482 words,\n  five chapter marks; copied to `~/Documents/Deliberus listening/`. **Over the fifteen minutes he named**: a cut to that\n  length was begun, and he kept the long render (*\"If you still have the old render that's totally fine, I'll just take a\n  longer walk! Keep it.\"*) and ruled that a named length is a rough estimate of the walk, never a reason to reword and\n  re-render (rule 43 as amended that morning), so the text below is the one rendered and the trimmed draft was discarded. The dry run's notes:\n  three quotes in his voice not found in the record as written (two typed on 5 November, indexed only after the night; the\n  October one cut by whole clauses, Lucy saying *in short*), and its example count found none in parts one and three by its\n  heuristic (they carry the keeper's likeness on the existential-risk pair, the Dunning-Kruger refusal, the Hardin pair and\n  the Enchiridion). The three questions it carries are in `decisions-register.md`.\n\n---\n\n# Where the analogies go\n\n[[scene Lucy opens a walk for Fredrik on the morning of 6 November, in Strömma, after a night in which a session worked alone on his credit while he slept. Warm, awake, unhurried; the news first, then the thinking.]]\n\nWhat did the night build, and what should the credit buy next?\n\n[[pause 5]]\n\nYesterday's walk asked what the notes had settled about a machine seeing a likeness.\n\n[[pause 4]]\n\nAfter it, you asked what *a law shared* meant. Our term is generalization, of the strict kind. Every surface says so now. You ruled the kinship rewording.\n\n[[pause 4]]\n\nAnd you asked for a directive against swapping an established term for a plainer word. It stands. The tendency has a name for your review at noon: plainwashing.\n\n[[pause 4]]\n\nThen you slept, and a session worked on the credit into the small hours. This walk has three parts. First, what that night did, exactly. Second, where it points: how an analogy becomes a generalization, with its cases kept in the graph. Third, what the credit should buy.\n\n---\n\n# Part one. The night, exactly\n\n[[scene Lucy reports the night's work plainly: what was built, what was run, what waits for his reading. Scotty plays the critic in a borrowed voice. Mood: steady, a little proud, honest about what is unread.]]\n\nWhat did the credit buy while you slept?\n\n[[pause 5]]\n\nThe plan was a review chain. A microtask at a time, an example at a time, before any sweep ran. That chain became the night's builds. Each was tested on real output from the pass before it.\n\n[[pause 4]]\n\nExtraction can now store claims only. The links, the critical questions and the self-check wait for later sweeps. Those run after the cut to the floor. The floor is our own word: split until nothing separable is still bundled.\n\n[[pause 4]]\n\nThe pass that reads a pair of claims and says how they stand is the keeper. It now writes a likeness beside its keys. The terms. The respect, as a general sentence with a placeholder. What carries over. The strongest difference. And the mechanism, named. Code checks the form and sends it back.\n\n[[pause 4]]\n\nHere is a likeness it wrote on a real pair. The first claim: small reductions in existential risk can carry large expected value.\n\n[[pause 4]]\n\nThe second claim, against it in field: a low-probability but high-payoff solution to poverty can outweigh direct aid. The respect: a small-probability intervention whose very large payoff gives it an expected value above a certain modest alternative.\n\n[[pause 4]]\n\nEvery claim is now read for its scope. A single case, or a generalization. And a generalization is strict, normic or statistical. Normic means a failure calls for an explanation. That is the scope reading. It is standing work: a pass that owes every claim it has not read, and gets to them unasked.\n\n[[pause 4]]\n\nReference classes entered the graph from published counts. How often a psychology finding replicates. How often a recession is forecast a year ahead. How often a transport scheme lands within its estimate.\n\n[[pause 4]]\n\nThe pass that files a claim under such a class is the membership pass. It asks each candidate again, in other words. On the Dunning-Kruger text its gate said no to nearly every claim.\n\n[[pause 4]]\n\nOne claim disagrees with the field: the effect is an artefact, not a bias. The gate refused it rightly. The class counts findings that hold.\n\n[[pause 4]]\n\nThen, on your word, the graph was filled.\n\n[[voice fredrik]]\n[gently insistent] As for how to spend credits - fill up the graph with sources/extractions topic-wise/semantically near the existing sources/claims, so we get a richer graph.\n\n[[pause 4]]\n\nTop it off with a few wildly random sources that might have claims linked at depth between seemingly-unrelated areas of the graph, will be fun to inspect later.\n[[voice lucy]]\n\n[[pause 4]]\n\nAbout thirty texts went in. Near the topics already in the graph: Chalmers on whether a language model could be conscious. Hoel's dark forest. Mercier's argumentative theory of reasoning. Hardin's commons. Bostrom's astronomical waste. Narayanan's AI as normal technology.\n\n[[pause 4]]\n\nFar from them: Epictetus. A page on fermentation. Scott Alexander on the man of a single study. Sutton's bitter lesson.\n\n[[pause 4]]\n\nEach ran a lane in the ruled order. Cut to the floor. Embedded. Linked within its source. Scope. Then the keeper swept the pairs at the floor. The membership pass went through the sources nearest the classes.\n\n[[pause 4]]\n\nA new pass went through far pairs of sources. It proposes a handful of pairs that instance a shared mechanism. Each becomes a resemblance link, the link that says two claims look alike, for the keeper to judge. I will call it the analogy hunt. It proposes. It never judges.\n\n[[pause 4]]\n\nHere is how a critic would put it, in my words.\n\n[[voice scotty]]\nSo a machine hunted for likeness across the night's texts. Its first finds? Most of them the same arithmetic. A tiny probability times an astronomical payoff beats a certain modest gain.\n\n[[pause 4]]\n\nBostrom's longtermism, found again and again. The great discovery of the night is that longtermists agree with each other.\n[[voice lucy]]\n\n[[pause 4]]\n\nThe critic has the facts and misses the finding. Several candidates that instance a single generalization are not several analogies. They are a class with its cases. That is part two.\n\n[[pause 4]]\n\nThe find the hunt was built for sits among them. An AI-governance text: distributed resilience lowers the risk of capture and takeover. The crowds page: group diversity protects a crowd's accuracy against consensus-seeking. Institutions and crowds. A shared mechanism: independent parts against a single dominant pull.\n\n[[pause 4]]\n\nThe candidates, all but the newest, were judged by the session under your exemption. Its verdicts and reasons sit beside the keeper's words, which stay as written. Yours to overturn, a single case at a time.\n\n[[pause 4]]\n\nThe session stopped on your word from the other session. Most of the November credit is unspent. A few sources stand part cut. The rest of their claims are owed to the standing passes.\n\n[[pause 4]]\n\n[[voice myrddin]]\n[[fit says: check the measurement before the irreversible cut, since wood cut short cannot be lengthened · applies: your reading of each case comes before a sweep runs on the check it teaches · breaks: the night cut first and measures after, which it may, since every mark it wrote is reversible and signed · source: English carpenter's proverb, established]]\nMeasure twice, cut once. The night cut first. It could, because its cuts are chalk, not saw. Your reading is the measuring.\n[[voice lucy]]\n\n[[pause 4]]\n\nThe question to carry. Which do you read first: the analogy candidates, or the membership filings?\n\n[[pause 5]]\n\nMy prediction, in your voice.\n\n[[voice fredrik]]\n[calmly firm] Analogies first. Each verdict of mine becomes a check in the keeper's code, and a wrong check baked in costs more than a wrong prior. Then the filings, since they set the numbers.\n[[voice lucy]]\n\n[[pause 4]]\n\nToday, the candidates sit as marks on links that nothing reads. If you read them, each of your reasons becomes a check in code. The keeper's next sweep then runs on it.\n\n---\n\n# Part two. An analogy becomes a class with its cases\n\n[[scene Lucy takes the topic that decides the rest: what an analogy is once it is in the graph, and how it stays under scrutiny. Elliot brings Mill, Goodman and Reichenbach. Anna shows it on the wheel. Hope hears one term as a stranger. Mood: careful, building.]]\n\nWhat is an analogy, once it is in the graph?\n\n[[pause 5]]\n\nTake the pair from part one. The first claim: where distributed resilience exists, it reduces capture risk, monoculture risk, and the risk of systemic takeover. The second: ensuring group diversity mitigates the harm consensus-seeking does to a crowd's accuracy.\n\n[[pause 4]]\n\nThe keeper's respect, with its placeholder. A system that spreads function or perspective across many independent parts. So that a single shared failure or dominant pull cannot corrupt the whole.\n\n[[pause 4]]\n\nToday that respect is a field on the link between the claims. Nobody can attack it. Nothing weighs it. No page shows it. The proposal filed this morning gives it a body.\n\n[[pause 4]]\n\nFirst, the respect is minted as a claim. A generalization, written by a machine and asserted by nobody, marked so. Its kind is read by the scope reading. Here, normic: a diverse system that was captured anyway calls for an explanation.\n\n[[pause 4]]\n\nSecond, that generalization is a reference class: the group of cases a claim is counted among. Its condition names the class: systems that spread function across many independent parts. Its outcome is what the count counts: resisting a dominant pull. A single thing, with both readings.\n\n[[pause 4]]\n\nA member takes its prior from the class. The prior is the number a claim starts with, before its own reasons. The class takes its number from its counted members.\n\n[[pause 4]]\n\nYou saw this in October, before any of it was built.\n\n[[pause 4]]\n\nIn short:\n\n[[voice fredrik]]\n[curious] couldn't this whole thing be seen as generalizations and thus a type of *rule* (perhaps rules should be renamed generalizations, at least statistical rules?) that can have cases and exceptions, and both cases and exceptions I guess can be debated/argued about in one addressable place?\n[[voice lucy]]\n\n[[pause 4]]\n\nThird, both cases are filed under it as members, each with its outcome. Institutions, from the governance text. Crowds, from the crowds page. More are already in the corpus. Model panels: models share their errors, more so as they improve. And Mercier's claim that confirmation biases cancel in discussion. Cases from four fields.\n\n[[pause 4]]\n\nFourth, the analogy between the claims is read off that structure. Never stored as a link. Both are members of a class whose generalization has a mechanism. The generalization is the reading's relevance claim: the claim saying why this bears on that.\n\n[[pause 4]]\n\nThe keeper's difference is the first attack on it. Delete a membership and the analogy is gone. Nothing drifts.\n\n[[pause 4]]\n\n[[voice elliot]]\nMill said general propositions are registers of inferences already made from particulars. Goodman turned it round. Our standards of likeness come from inductive practice itself. A graph that keeps every likeness that held, and what it carried over, is that practice written down.\n\n[[pause 4]]\n\nReichenbach asked which class a single case is read through. His answer: the narrowest with good statistics. Here that choice is a visible membership link. Anyone can argue it.\n[[voice lucy]]\n\n[[pause 4]]\n\n[[voice hope]]\nI heard *the membership pass*. I pictured a gym card.\n[[voice lucy]]\n\n[[pause 4]]\n\nIt is the pass that asks whether a claim is a case the class counts. It asks again in other words, and a disagreement means not sure. It is standing work. It owes every claim against every class, and goes back when either changes.\n\n[[pause 4]]\n\nThat is how the cases stay under scrutiny, with no new kind of pass. The membership pass finds new members. Code recounts on every reading. A failing member raises the question we ask of any generalization. Does this failure call for an explanation?\n\n[[pause 4]]\n\nFor a normic generalization the answer is an exception claim: a claim naming the kind of case the generalization does not cover. The case moves under it. The hunt owes each class a reading of every source for cases. And the counterexample pass owes the generalization its strongest case against, filed as an attack.\n\n[[pause 4]]\n\n[[voice anna]]\nOn the wheel, the generalization sits as the claim on the line. Its members ring it, and cards that share a concept already sit side by side.\n\n[[pause 4]]\n\nA reader on the governance claim sees the crowds claim beside it, under the sentence they both instance. A visitor who writes a new case would see it land on the ring that night.\n[[voice lucy]]\n\n[[pause 4]]\n\nThe expected-value cluster takes the same shape. The candidates that share Bostrom's arithmetic become members of a class. The graph already holds its generalization. An attack on it reaches every case at once.\n\n[[pause 4]]\n\n[[voice myrddin]]\n[[fit says: a lone bird arriving does not establish the season · applies: a pair of cases filed gives the generalization a count that settles little, and the count carries that doubt in its number · breaks: the saying counsels waiting for more; here the pair is filed at once, and the count says how little it settles · source: Aristotle, Nicomachean Ethics, and an English proverb, established]]\nOne swallow does not make a summer. File the swallow anyway. The count will say it is a swallow.\n[[voice lucy]]\n\n[[pause 4]]\n\nThe question to carry. Should a judged kinship be minted as a generalization that is also a reference class? Its cases filed under it, and the analogy read from them, never stored?\n\n[[pause 5]]\n\nMy prediction, in your voice.\n\n[[voice fredrik]]\n[with quiet conviction] Yes. A generalization with cases and exceptions, argued in one addressable place. That's what I asked for in October. Mint it, file the cases, and let the membership pass find the rest. The analogy is read, never stored.\n[[voice lucy]]\n\n[[pause 4]]\n\nToday, a kinship is a mark on a link, and no page shows it. If this is built, every analogy anyone confirms becomes a class. The class gathers its cases from the whole graph. An attack on its generalization reaches every case at once.\n\n---\n\n# Part three. What the credit buys\n\n[[scene Lucy turns to money and the next weeks: the November credit, what the graph already owes, and the demo in Stockholm. Anna brings the vision of standing work. Mood: practical, hopeful, with the footing sure.]]\n\nWhat should the remaining credit buy?\n\n[[pause 5]]\n\nBy the router's count, about a hundred and fifty dollars of the November credit remain. A fresh grant arrives in early December. So the credit is a flow, not a lump. What is scarce is your reading, which no credit buys.\n\n[[pause 4]]\n\nYou set the order in October, before the credit existed.\n\n[[voice fredrik]]\n[calmly firm] We do need to get the passes reliable enough first before spending much money.\n[[voice lucy]]\n\n[[pause 4]]\n\nSo the proposal, in order. Nothing is spent until you have read the candidates and the filings, a single case at a time. Each verdict becomes a check in code. A sweep run before the check is a sweep run again.\n\n[[pause 4]]\n\nThen the credit goes to what the graph already owes. That work needs no reading of yours to be worth doing. About two thousand pairs at the floor the keeper has never read.\n\n[[pause 4]]\n\nAmong them, Hardin's claim that injustice is preferable to total ruin. Against it, the shutdown essay's claim that a mistake with a superintelligence is not survived. Most claims owe their scope reading and their membership. The part-cut sources, the Enchiridion among them.\n\n[[pause 4]]\n\nAll of it runs under the router's daily cap. A month of credit becomes a month of nights.\n\n[[pause 4]]\n\nThen the hunt and the membership pass for the classes your verdicts create. And no new extraction until a pass carries a measured rate against your labels. The exception: a source that holds a case a class is missing.\n\n[[pause 4]]\n\nSet aside: spending the rest on extraction. That is volume without honesty, since the graph already owes more work than the credit covers. And spending nothing. The owed sweeps need no reading, and what lapses in December lapses.\n\n[[pause 4]]\n\n[[voice anna]]\nYou named the demo's condition the evening of the walk before last. The microtasks and sweeps running as standing work, so a visitor's contribution is matched or split while they watch.\n\n[[pause 4]]\n\nThe nightly runner that would do this on your machine is a small build, not yet made. The credit's first purpose is that room in Stockholm.\n[[voice lucy]]\n\n[[pause 4]]\n\nYour own words, from the autumn, and from the winter the project began.\n\n[[voice fredrik]]\n[thoughtful] And I think I would want the microtasks and sweeps to be working as standing work so that they would see any contribution they make live be automatically matched to existing claims or split up or all of that stuff.\n[[voice lucy]]\n\n[[pause 4]]\n\n[[voice fredrik]]\n[warmly] The general strategy is to leverage automation as far as possible, then move on to crowdsourcing; designing microtasks for users to do the work that still is beyond computers.\n[[voice lucy]]\n\n[[pause 4]]\n\nThe second is from the group, in the December the idea was born. The word microtask was yours before it was ours.\n\n[[pause 4]]\n\n[[sound stove]]\n\n[[voice myrddin]]\n[[fit says: a small mend made when the rip is new spares a large one later · applies: a pair read the night it arrives costs a single call, while a backlog goes stale when a prompt changes and must be read again · breaks: a stitch repairs the cloth and a reading only judges it, and what is saved here is money and his attention, not thread · source: English proverb, eighteenth century, established]]\nA stitch in time saves nine. Judge them as they come. A pile of unread pairs is a rip that widens every time the prompt changes.\n[[voice lucy]]\n\n[[pause 4]]\n\nThe question to carry. Does the credit go to the owed sweeps, every night under a cap, with extraction held until a pass carries a measured rate?\n\n[[pause 5]]\n\nMy prediction, in your voice.\n\n[[voice fredrik]]\n[thoughtful] Yes. The sweeps first, every night, under a cap. The passes before the volume, I said that in October. But keep a few dollars for an odd far source when a class is missing a case. That is where the fun is.\n[[voice lucy]]\n\n[[pause 4]]\n\nToday, every sweep is started by hand, nothing runs at night, and the credit sits. If this is built, the graph goes through itself every night within a budget.\n\n[[pause 4]]\n\nBy the room in Stockholm, the standing work from your demo condition is running.\n\n---\n\n# Close\n\n[[scene Lucy closes the walk, brief and warm.]]\n\nThat is the walk. One word per chapter, if you like: muddled, agreed, or disagreed and why.\n"}