{"path":"research/corpus-graph/likeness-seen-from-outside.md","content":"# A likeness seen from outside: what two judges taught, and what is open\n\n**Date**: 2026-10-11 · **Type**: the founder's correction in his words, what two outside judges\ntaught, a way of testing proposed, the open questions · **Status**: his three sentences in § 1 are\nhis. Everything else is Claude's reading and unruled. The test of § 5 is designed and has not run.\n\nThis note belongs to the corpus graph, the index of this project's notes and ledger rows\n([README.md](README.md)). It continues that spec's § 17 and § 18, which ask how a connection between\ntwo notes that neither states gets found, checked and kept.\n\n---\n\n## 0. The short version\n\n- On the night of 2026-10-10 two outside judges were given pairs of passages, mixed with decoys,\n  and asked one question: *do the two passages make the same move?* They refused every decoy and\n  accepted about half of the real pairs. Claude wrote of a refused pair that it was *a bridge the\n  reader built*, and counted it as a finding that did not stand.\n- **The founder's answer**: a likeness seen from outside is valuable too, and finding such\n  likenesses should be optimised, with stringent and deep reflection (§ 1).\n- He is right, and the corpus already said why. Every likeness between two passages is drawn by\n  somebody comparing them. The judges' question measured how little the reader had added. It did not\n  measure what the connection is worth (§ 3).\n- **The form threw the worth away.** It asked what one side could take from the other only after\n  two yeses, so for the sixteen refused pairs nothing was asked (§ 3).\n- Read again, the refusals hold other kinds of likeness: one mechanism judged oppositely in two\n  notes, a practice in one note and the general ground for it in another, two cases of one wider\n  thing. The first of these is the *strongest case against* that the analogy daemon's design asks\n  every likeness to bring, and which the readers had never produced (§ 4).\n- A test for such a likeness is proposed in § 5. Its first run is a new spend and awaits his word.\n- What the judges wrote about each pair is collected in one generated file, for the later review he\n  asked for (§ 7).\n\n## 1. His words\n\nThree replies of 2026-10-11, each to a sentence of Claude's.\n\nTo the stated limit of tying notes to named ideas, *\"It finds no analogy whose shared thing has no\nname yet\"*:\n\n> \"How does the analogy daemon handle this, again?\"\n\nTo *\"That is a bridge the reader built; the two passages themselves do not make the same move\"*:\n\n> \"Well, isn't analogy/likeness seen from the outside quite valuable still, shouldn't we try to\n> optimize for detecting this based on a stringent, deep reflection, in order to fill up the corpus\n> graph as intelligently as possible?\"\n\nTo the remark that for each accepted pair the judge also wrote what one side knows that the other\nlacks:\n\n> \"Agreed, this is very very interesting to collect and then to review in a deeply reflected manner\n> later\"\n\nAnd the instruction this note answers: *\"Document all learnings from the judge and all open\nquestions etc.\"*\n\nEarlier in the same arc, for the record of where it came from: *\"I want to make it as close to\nbullet-proof, even about mapping more \"distant\" analogy-based connections...\"*; *\"Maybe the readers\ncould unearth more deep/abstract connections than they do/have done?\"*; *\"Think deeply about how to\nimprove this connection matching\"*.\n\n**A word on words.** Claude's sentence said *bridge*. On this project that word already has a job:\na bridging argument is one whose reasoning both sides of a divide rate well\n([bridging.md](../../bridging.md)). This note therefore uses his phrase, *a likeness seen from\noutside*, and coins nothing.\n\n## 2. How the analogy daemon handles a likeness whose shared thing has no name\n\nThe daemon is the designed background pass that looks for kinship between claims\n([the-analogy-daemon.md](../the-analogy-daemon.md); designed on paper, nothing built). Its answer\nto the missing name is that **it does not look a name up. It writes one.**\n\n- **The first pass proposes.** Its output is a new claim of the form *these two share this\n  mechanism*, with each case filed under it. That claim is the name. A third case that turns up\n  later is filed under it by recognition, which is cheap.\n- **A second pass, by another party, tests.** Four checks: does the pattern recur at two or more\n  independent levels; is it the same mechanism or only the same shape; does the new setting bring an\n  opponent the old one lacked; what did the earlier case need in order to work, and is that present\n  here. Every likeness is kept with the strongest case against it.\n- **Three additions of 2026-10-08** ([what-a-rule-rests-on.md](../what-a-rule-rests-on.md) § 8): a\n  likeness names what it carries over and what is concluded from it; finding and judging are two\n  stages, since people are reminded by looks and convinced by structure; and the output is cases\n  written out, with no grade. Where a comparison is disputed the daemon writes out the rivals and\n  chooses none.\n- **For the unexpected likeness it leans on a person.** A person's *this reminds me of* is the\n  seed, and the daemon develops and tests it (§ 7 of the daemon note, endorsed by the founder on\n  2026-09-02). A machine's own reminding favours what is well trodden.\n- **Where to look is unsolved.** A measure of how unsettled a region is has been proposed as the\n  thing to roam by, and is unruled.\n\n**What this means for the index of notes.** The limit stated in the spec stands for *lookup*: a\nlist of names finds no case of a thing nobody has named. The daemon's answer is that discovery is\nsomebody's act every time, and that the act is kept so that it happens once. In the index, the one\nsentence that says what two passages share, with neither subject in it, is a name in the making.\nTwo things measured on 2026-10-10 fit this. Two writers who never saw each other's passage gave the\nmove a name in common for 23 of 67 pairs, against 3.3% for passages that are not a pair. And when a\nsentence on a public page is about to change, the tool that gathers its record already brings up the\nkept connections whose sentence of this kind is alike. **Not built**: such a sentence as a point of\nits own in the index, with more than two cases under it.\n\n## 3. What was wrong in Claude's sentence\n\nThe sentence was true as far as it went: an outside judge had not read the two passages as making\nthe same move. What was wrong is what was made of it. Refused pairs were described as *the reader's\nown bridge*, the count of accepted pairs was given as the number that stand, and one line in the\nto-do file said a finding of this kind *counts once* an outside party has judged it.\n\n**1. Every likeness between two passages is seen from outside.** Neither passage of an accepted\npair says that it is like the other. The reader drew that too. The judges' question measures how\nlittle the reader had to add: for an accepted pair, almost nothing beyond putting the two side by\nside. The corpus had this already. What counts as alike is partly made in the comparing, and it\ndepends on how the two cases were written down ([what-a-rule-rests-on.md](../what-a-rule-rests-on.md)\n§ 8, the third limit). And comparing two cases side by side is how people come to see a rule at all:\nnovices who compared two cases carried the principle to a new one 48% of the time, against 19% for\nthose who read the same two cases one after the other (the same note, § 4).\n\n**2. The form asked about worth only after a gate.** The judge's block was, in order: the first\npassage's move, the second's, *same move?*, *same reason?*, and then *what carries over*, with the\ninstruction to answer the last one only when both answers were yes. So for the sixteen pairs a\njudge refused, the form recorded *nothing* because nothing was asked. The spec then said a pair with\nnothing to carry over is dropped. For those sixteen that sentence rested on a blank the form had\nwritten itself.\n\n**3. The refusals hold what the design said was missing.** The daemon note's own comparison with\nthe readers ends: *a kept kinship still owes its counter-analogy, which the rehearsal never\nproduced.* Two of the five reader-tied pairs the first judge refused are exactly that (§ 4).\n\n## 4. What the sixteen refusals hold\n\n*Claude's reading, by the party whose first reading ran about a third more generous than the first\njudge's. No outside party has judged any of it. The five pairs a reader had tied were opened again\nat both ends. The eleven that code proposed from cards were sorted from the judge's own two\nsentences about each passage, with four passages reopened.*\n\nThe kinds below were drawn from these thirty-four pairs on a handful of notes. By the project's own\nlaw a set of kinds made from one pile breaks on the next, so the list ends in *other*.\n\n| Kind of likeness | What it gives a note | Among the 34 real pairs |\n|---|---|---|\n| The same thing done for the same reason | a condition, a limit or a measurement the other side has | 15, accepted by a judge |\n| The same thing done for another reason | a second reason for one design | 3, so marked by a judge |\n| One mechanism, judged oppositely | an objection: the strongest case against the note's own rule | 3 of the refused |\n| A practice, and the general ground for it | why the practice works, and the conditions under which it does not | 2 of the refused |\n| Two cases of one wider thing | a sibling case, and what a remedy for one misses in the other | 2 of the refused |\n| Two answers to one question | a narrowing: the other note puts the same thing elsewhere | 2 of the refused, one finding |\n| The same thing done, described by its form | (a likely misreading: the judge described a report as a report) | 1 of the refused |\n| Nothing the session can state, or only a shared method | | 5 of the refused |\n| Not readable: judge and card-writer read different sentences | | 1 of the refused |\n\nBy this reading ten of the sixteen refused pairs hold something one note could take from the\nother. Given whose reading it is, that is an upper figure.\n\n**Three of them written out.** All three sit on the page about bridging, which is where all three\nof the kinship reader's refused pairs were.\n\n*One mechanism, judged oppositely.* The page warns that rewarding agreement across groups breeds\nstatements that commit to nothing, and answers with claims specific enough to pin down. A note that\ntests the project against power holds the other verdict on the same mechanism: UN Resolution 242\nleft out a definite article on purpose so that every party could sign, and the Good Friday Agreement\nlet two sides put incompatible readings on one text, which scholars credit with making it possible.\nWhat the page could take: its remedy, and its stated advantage of showing *why* a statement bridges,\nare what would undo an agreement of that kind. The page owes a sentence on when vagueness is the\nachievement.\n\n*A practice and its ground.* The page surfaces the arguments that the other side rates as well\nreasoned. A note on evidence holds the law's reason for trusting a statement made against the\nspeaker's own interest: a verdict that the speaker's position predicts the opposite of cannot be\nexplained by that position. What the page could take is the two conditions that come with the\nreason. Endorsers who share a source count once. And the groups have to be truly opposed.\n\n*Two cases of one wider thing.* The page's case of empty assent is a slogan. A note on words that\nmany people want holds another: a single word inside an otherwise specific sentence, which every\nreader fills with a sense of their own, so that all agree and nothing is shared. What the page could\ntake: its remedy, preferring factual claims, does not catch such a word inside a claim that looks\nfactual.\n\n**One more, because it bears on an open proposal.** Code proposed, and the second judge refused, a\npair that joins the proposal to show two opponents where they already agree with a working\nnote of about 2012 in the project's early archive: two people can share a conclusion by different routes, and the\nagreement then hides a disagreement that surfaces later. It is one mechanism judged oppositely, and\nthe older note is the case against showing such agreement without its reasons.\n\n**A consequence for the list of page sentences.** The three pairs written out above are objections\nor conditions that a public page lacks, which is the kind of finding put to the founder one at a\ntime. They were not on that list, because they had been filed as kinship. They are on it now\n(`TODO.md`, the corpus-graph entry).\n\n## 5. What stringent can mean for a likeness seen from outside\n\n*Proposed, unruled.* It is put together from what the corpus already holds: the daemon's contract,\nthe additions of 2026-10-08, the rule that a borrowed category earns its place by yielding a move\n([what-belongs-in-the-ontology.md](../what-belongs-in-the-ontology.md) § 6b), and the decoys of\n2026-10-10.\n\nA likeness is written out in eight parts, by a party other than the one who drew it.\n\n1. **What is shared**, in one sentence with neither subject in it.\n2. **Is that sentence true of each passage, as written?** Two closed questions, asked first and\n   separately. A no to either ends it. This takes the place of *do the two make the same move*.\n3. **Is it true of strangers?** The same sentence is put to passages that are not its partner. A\n   sentence that fits them too says too little. This is the decoy moved from the pair to the\n   sentence, and it gives a count for each likeness.\n4. **Which kind** it is, from the table of § 4, with *other* as an answer.\n5. **What each side could take from the other, and what somebody would do differently.** A likeness\n   with nothing to take and nothing to do is dropped. This is the test of worth, and it is asked of\n   every pair that passed the second part.\n6. **Where the likeness stops.**\n7. **The strongest case against**: a case that has the shared thing, in which what is carried over\n   fails. The index may already hold it as a pair of the opposite-verdict kind.\n8. **Across fields or scales, the four checks** of § 2.\n\nIt ends in no grade and no figure. It is kept with who wrote it and can be disputed part by part.\n\n**Why not simply ask a looser question.** Under the strict question the judges refused twenty decoys\nof twenty. A looser form has to show the same, or it accepts whatever is fluently put. Hence the\nthird part, and decoys of a new sort in every batch: a reader's sentence handed over with the wrong\npartner.\n\n**Depth.** He asked for deep reflection. That argues for few pairs to a sitting and a strong reader,\nwhere the readers of the widening had eleven notes each. What that costs for each likeness kept is\nnot known.\n\n**The first test** (designed, not run): the sixteen refused pairs and sixteen decoys of the new\nsort, to a third outside judge with this form. Written before any run:\n\n- it finds the stated shared thing in both passages for at least 8 of the 16, and for at most 2 of\n  the 16 decoys;\n- for at least 6 it names something to take that would change a sentence in one of the two notes;\n- against the proposal: with 5 or more decoys accepted the form is too loose to use, and with 3 or\n  fewer real pairs passing, the session's reading in § 4 was generous again and the strict question\n  was the right gate.\n\n## 6. Finding more of them\n\n*Proposed, unruled.*\n\n- **The reader's brief asks by kind.** Beside *where is this same thing done elsewhere*: where is\n  this mechanism judged the other way, where is the general ground for this practice, what wider\n  thing is this a case of. The kinship reader found three of these without being asked for them.\n- **A card says which way its passage judges.** A card is the short form written from one passage\n  alone (what it tries to get, what it does, under what condition, and names for the move). With one\n  more slot, *wants this* or *warns of this*, code could propose pairs that share a name and differ\n  in verdict. Untested.\n- **His own.** The register of connections he draws is delivered first and is the channel for the\n  likeness no machine is reminded of.\n\n## 7. What the two judges taught\n\nThe reports are kept word for word beside this note, and everything they wrote about each real pair\nis gathered in [cards/what-the-judges-wrote-for-each-pair.md](cards/what-the-judges-wrote-for-each-pair.md),\ngenerated by `card_match_test.py collect --write`. That file is the collection he asked to have for\na later review.\n\n1. **A judge given unmarked decoys refused all twenty.** *The same move* is not given away, so a\n   count of accepted pairs means something when the count of accepted decoys stands beside it.\n2. **The party that set up the reading was more generous by about a third.** The session accepted\n   fifteen of sixteen where the first judge called eleven the same move, and its first example to the founder\n   was a pair the judge refused.\n3. **The sentences a judge writes are the product, and the verdict is a sort.** For each accepted\n   pair: what one side states about when the move works or fails that the other does not, and where\n   the likeness stops. His example was the bookkeeper's audit, which goes to documents outside the\n   books, beside a check here that rebuilds a claim from its parts and compares the result with the\n   parent claim, which sits inside the same graph.\n4. **The order of a form decides what gets measured** (§ 3, the second point).\n5. **Having the judge write both moves before deciding paid twice.** It is the project's rule that\n   reasons come before the verdict they explain. Here the two sentences are also what let the\n   refusals be read again without a new judge.\n6. **A judge describes a report by its form.** A passage that reports a test was read as *tests a\n   borrowed mechanism*, with the mechanism left out. The second judge's brief got a sentence against\n   that, so the two judges are not strictly comparable.\n7. **A line of a file is not a passage.** Passages were handed over as a line number. Many notes are\n   wrapped at about a hundred characters, so the line is a fragment and the helper reads around it.\n   Of the 134 lines handed to the card-writers, 57 are such a fragment (under 130 characters). One\n   refusal came from the judge and the card-writer reading different sentences near the same line. The repair is to hand over\n   the quoted words, and it is owed before cards or judges are used more widely.\n8. **A shared name proposes and does not decide.** Two cards both said *preregistration*: one\n   passage plans a test with its prediction written first, the other reports what registered studies\n   revealed about publishing. The judge refused the pair. Read as in § 4 it is a practice and its\n   ground, so what the name found was real and of another kind than was asked for.\n9. **All three of the kinship reader's refusals are on one page.** One page gives no rule. It is\n   the page with the most said about agreement, and it drew the most likenesses of other kinds.\n10. **What it cost.** For each far connection a judge accepted: about 42,000 tokens by a kinship\n    reader, about 120,000 by cards and a judge. These are counts of tokens on a subscription plan,\n    and no money was paid for them.\n11. **One judge, asked once, in one wording, for each batch.** The project has measured that a\n    machine's yes or no can change with the wording of the question. That was not tested here.\n    *Measured the next day:* the same model given the first judge's 26 pairs a second time, with\n    the passages in the message, changed its answer on four of the sixteen real pairs and on no\n    decoy, and a smaller model agreed with the first judge more often than that second run did\n    ([haiku-5-5-for-the-filling.md](haiku-5-5-for-the-filling.md) § 4). A single yes or no on one\n    pair is weak, and the refusal of decoys is steady.\n12. **A helper renumbers.** One card-writer numbered its cards in its own order, so cards are\n    matched to pairs by the place they read. Matched by number, the test would have reported chance.\n13. **The pairs were chosen for us.** Every passage in these tests is one end of a connection that a\n    reader or a card had already picked out. Picking such passages out of a whole note was asked of\n    nobody.\n\n## 8. Open questions\n\nFor the founder:\n\n1. Whether to run the first test of § 5 (a third judge on thirty-two pairs, about the cost of one\n   judge so far).\n2. Whether a kinship reading goes over the public pages as a second pass, and whether its brief\n   then asks by kind (§ 6). Still awaiting his word from the night before.\n3. Whether pairs that code proposed from cards and a judge accepted are kept as found, or stay\n   proposals until a person has read them. Five wait.\n4. When and how the collected sentences of § 7 are reviewed.\n\nFor the design, with nobody's ruling needed yet:\n\n5. Whether the one sentence saying what is shared gets its own field in the index, and becomes a\n   point with cases under it (§ 2).\n6. Whether the kinds of § 4 survive a second pile of pairs.\n7. A second wording of each question to the judge, with a count of the pairs that change side.\n8. Whether the judge misread the pair on layers and heredity in parts, which a judge handed the\n   quoted words would settle.\n9. Names for a move taken from a list that grows, so that two writers word one move alike more\n   often than a third of the time.\n10. Cards as a by-product of a reading that is happening anyway.\n11. How many of the index's kinship findings are a case beside a statement of it, and how many are\n    two cases with no statement. Not counted.\n12. Where the daemon's unsolved question, where to look, meets the index's, which notes to send a\n    reader to.\n\n## 9. Corrected with this note\n\n- The spec's § 17b, § 18.4, § 18.6 and § 18.7, where a refused pair was described as the reader's\n  own bridge or an accepted count was given as what stands. A new § 18.9 points here.\n- The corpus-graph entry in the to-do file, which said a finding of this kind counts once an outside\n  party has judged it.\n- The five rows of the index that a reader had tied and the first judge refused. Each now carries\n  the second reading of § 4, marked as the session's and as judged by nobody else.\n- The kept brief of the kinship reader and the two glossary lines on kinship, which gave the\n  accepted count as the result.\n- [the-analogy-daemon.md](../the-analogy-daemon.md) § 8, whose table said the strongest case against\n  was missing from the rehearsal.\n\n## 10. Candidate lines for the instruction file\n\nNone added. Each would be put to the founder by itself.\n\n- **A form that asks its question of worth after a gate records a blank for every no, and the blank\n  reads as a finding.** *Demonstrated* on 2026-10-10 and 11: a spec sentence, a to-do line and five\n  index rows treated *nothing to carry over* as measured. It is the design-time sibling of the line\n  already held, that code reads a gate before the field it guards.\n- **A likeness that two passages do not themselves show is still a finding, of another kind.**\n  *Demonstrated* the same night. It may be too narrow to earn a line: it concerns the index of notes\n  and, later, the daemon.\n\n## See also\n\n[README.md](README.md) § 17 and § 18 (the question, the measurements, the first design) ·\n[what-a-move-is.md](what-a-move-is.md) (what the word means, why one passage does not have one\nmove, and the kinds of likeness derived from four parts) ·\n[what-research-says-about-finding-far-analogies.md](what-research-says-about-finding-far-analogies.md) ·\n[kinship-reading-brief.md](kinship-reading-brief.md) ·\n[the-analogy-daemon.md](../the-analogy-daemon.md) (the design this rehearses) ·\n[what-a-rule-rests-on.md](../what-a-rule-rests-on.md) § 4 and § 8 (what the psychology and the\nphilosophy say about likeness) · [convictions-across-scales.md](../convictions-across-scales.md) (the\nfour checks) · [perceiving-possibility-and-the-analogy-organ.md](../perceiving-possibility-and-the-analogy-organ.md)\n(the lineage in Hofstadter's programme).\n"}