{"path":"research/corpus-graph/reading-for-every-kind-of-insight.md","content":"# Reading for every kind of insight: how the judge is told to work today, and what a read-through would need\n\n**Date**: 2026-10-11 · **Type**: the founder's question answered from the files, the corpus on\nanalogy read through, three web researchers' reports, and a design · **Status**: § 1 is his words.\n§ 2 and § 3 are read from files and transcripts. Everything in § 5 to § 8 is Claude's proposal and\nunruled. Nothing in § 7 has run.\n\n**How it was read.** Claude read the corpus's notes on analogy itself (listed in § 4). Three\nread-only web researchers (Claude Sonnet 5.5, no access to this repository) read outside sources;\ntheir reports are kept word for word in [reading-for-insight/](reading-for-insight/), each finding\nmarked with how far its source was read. Claude then opened five of the sources itself, at the\ndepth of the abstract: the Wikipedia inconsistency study, Klein and Jarosz, AbsenceBench, DocPrism\nand FIND. Everything else from outside is as a helper reported it.\n\nThis note belongs to the corpus graph, the index of this project's notes and ledger rows\n([README.md](README.md)), and continues [likeness-seen-from-outside.md](likeness-seen-from-outside.md).\n\n---\n\n## 0. The short version\n\n- **The judge of 2026-10-10 was a narrow sorter.** It saw two passages with fifteen lines either\n  side, nothing else, and was asked whether they make the same move. It was told nothing of what\n  the corpus holds on analogy, and it was never asked whether a connection is worth having (§ 2).\n- **The corpus holds a good deal on how to test a likeness, and none of it was in the judge's\n  brief** (§ 4).\n- **Dissonance is already half of what the readers return**: of 601 findings that stood, 324 say that\n  something contradicts, corrects or narrows something else. What the readers were never asked for\n  is tension between two things the project holds, and a clash that shares a subject and few words\n  (§ 3).\n- **The readers saw more than the form kept.** Most reports end in a free line. In those lines the\n  readers name 58 further pairs that are not in the index, read at both ends and cut for a cap of\n  eighteen. About fifty of them are filed nowhere. They also name inconsistencies on public pages (the 24 pages written\n  for visitors; every note under `docs/` is served on the site, the research notes too),\n  some of which a rule of the brief forbade them to report; the session had put most of those on the\n  founder's list at the time, and two were new today (§ 3.3).\n- **Research agrees on four things**: a reader finds the kind of thing it was sent for and little\n  else; a permissive reader followed by a strict judge beats one careful pass; machines are weak at\n  noticing what is absent unless the empty place is marked; and no benchmark measures how much of\n  what a careful person would find a machine reading recovers (§ 5).\n- **The design**: one kind of finding to a pass, the open question asked first and alone, no cap,\n  a card for each load-bearing passage whose slots are paired by code, a second party with a form\n  that fits the kind, and an answer key made from this project's own corrections (§ 6).\n- **First steps**, cheapest first, in § 7. The first costs nothing and is owed.\n\n## 1. His words\n\n2026-10-11, after the note on a likeness seen from outside:\n\n> \"How is the judge instructed, which docs is it told to read, which model does it run, etc? Tell\n> me more about how it is told to evaluate connections/analogies etc. Read everything the corpus\n> has on analogies. And I guess I also want to find connections that maybe are more dissonant, like\n> inconsistencies. Really I would like these corpus read-throughs to surface any insight that WE\n> would find if we read through everything by hand and thought deeply about it. Any kind of\n> insight. Think deeply about this. Research this extensively online also.\"\n\n## 2. How the judge was instructed\n\nRead from the two judges' own transcripts and from their briefs, which are kept word for word at\nthe end of [the first report](cards/2026-10-10-outside-judge-with-decoys-as-returned.md) and\n[the second](cards/2026-10-10-second-outside-judge-as-returned.md).\n\n| | |\n|---|---|\n| Model | Claude Sonnet 5.5, run as a read-only helper of the kind that does not load the project's instruction files. It knew nothing of the project beyond its brief |\n| What it was given | A list of pairs, each two places (a file and a line). 26 pairs for the first judge, 28 for the second |\n| What it was told to read | Each file at its line, with fifteen lines before and after. Nothing else: no search, no other file, nothing in the folder that holds the answers |\n| What it was told about the pairs | Some were put together by a reader (or a method) that thought they make the same move, others at random, unmarked, in a share *not what you might guess* |\n| What it wrote for each pair | The first passage's move, with its subject taken out. The second's. Then: the same move, yes or no. For the same reason, yes or no. What carries over (only after two yeses). Where the likeness stops |\n| The standard | *Be strict.* A shared topic, word or general virtue is not a shared move. A move has something it is trying to get, something it does, and a condition under which that works. If a sentence must be stretched to match, the answer is no |\n| How much work | 42 and 57 model calls. One pass, one wording |\n\n**What it was not told.** What an analogy is for. What the corpus holds on testing one (§ 4). That\na likeness may be of another kind than *the same move*. Any example. And it was not asked for the\nstrongest case against a pair, for what would be done differently if the pair holds, or for the\nfour checks.\n\n**Checked today: the judges ran blind.** A reminder fires in any session that opens a note, saying\nwhat other files have judged about it. It reached the helpers too. The first judge got it once,\nabout a note in a decoy, and the second never. The two writers of cards got it for twelve and three\nnotes, and for 14 of the 67 pairs it showed a card-writer a kept finding that names the partner\nnote. Scored again without those 14, nothing moves: the partner stands in the first three for 32%\nby the passages' words and 41% by the cards, as before, and the two cards share a name for the move\nin 21 of 53 (`card_match_test.py cards`; the 14 are listed in\n[cards/pairs-the-reminder-showed.json](cards/pairs-the-reminder-showed.json)). A helper in a blind\ntest should be handed the words in one file and open nothing in the repository. That also repairs\nthe fragment problem of [likeness-seen-from-outside.md](likeness-seen-from-outside.md) § 7.\n\n**The readers.** Also Claude Sonnet 5.5, read-only, 183 to 273 model calls each. A reader is given\neight to eleven notes to read in full, the search tools, and\n[one brief](reader_brief.md). That brief asks for five kinds of finding at once, best first, at\nmost eighteen, and a last line beginning *Noticed:* for anything that fits no kind. It excludes any\npair where the note already names the other note. The reader that found fourteen kinships had\n[another brief](kinship-reading-brief.md): it first read the notes where the project names its\npatterns and the three lists of held ideas, and was sent for kinship only.\n\n## 3. Dissonance: what the readers already return, and what they were never asked\n\n### 3.1 What is kept\n\nCounted from the index on 2026-10-11, over the 601 findings that stood (578 kept and 23 since written\ninto the notes; the index file has 661 lines, which also hold later marks, 22 pairs set aside and 5\nproposals):\n\n| Kind, as the reader labelled it | Findings |\n|---|---|\n| narrows | 162 |\n| a later correction | 95 |\n| contradicts | 67 |\n| supplies a missing piece | 63 |\n| the same mechanism elsewhere | 48 |\n| already answers | 53 |\n| the same idea earlier | 19 |\n| tension | 4 |\n| labelled in the reader's own words, or unlabelled | 90 |\n\nSo about half of what is kept is dissonant already. The project also has two instruments aimed at\none kind of dissonance each: the outside check, which compares a new proposal with everything\nsettled ([../outside-checks/README.md](../outside-checks/README.md)), and the list of passages a\nruling has overtaken (`repair.py candidates`).\n\n### 3.2 What the form cannot return\n\n1. **A conflict with a note the target names.** The brief excludes such pairs, to keep readers on\n   connections nobody wrote. A note that cites another can misstate it, and several readers say in\n   their free line that they saw exactly that and could not report it.\n2. **A conflict inside one note, and between a note and the code.** Readers were sent to pairs of\n   notes.\n3. **A clash that shares a subject and few words.** The outside check measured this on its own\n   cases: it named the plain reversal, where the two texts share their words, and missed three of\n   three collisions between texts that share a subject and little wording.\n4. **Tension between two things the project holds.** Four findings of 601. Nobody was sent for it.\n5. **One mechanism judged oppositely in two notes.** Found only by accident, as kinship that a\n   judge refused ([likeness-seen-from-outside.md](likeness-seen-from-outside.md) § 4).\n\n### 3.3 What the readers saw and the form did not keep\n\nOf the 29 reports, 26 end in a free line, and none of the 26 says *nothing*.\n\n- **Pairs cut for the cap.** Twenty reports say that further pairs were read at both ends and left\n  out because eighteen was the limit. A count by code of the clauses that name two notes gives 67\n  such pairs, of which 58 are not in the index. Eight of the 58 stand together in one passage of the\n  to-do file as it was before today, so about fifty are filed nowhere. The true number is higher,\n  since many clauses name the second note by a section and no file name.\n  **Kept the same day, on his word** (*\"Ok and these are documented so we can go through them?\"*):\n  the lines are gathered whole in\n  [what-the-readers-noticed-beyond-their-forms.md](what-the-readers-noticed-beyond-their-forms.md),\n  and the session opened both ends of every pair a line names with a line number on each side and\n  kept 76 in the index ([README.md](README.md) § 18.12). A stricter count than the first one, by\n  the two notes of each pair, is what gave 76 where the first gave 58.\n- **Inconsistencies on public pages.** Eleven are named. Seven were opened today at the ends stated\n  and stand. *(⚠ Corrected the same day: this note first presented all seven as never kept. The\n  session that filed the readers' reports had already put five of them on the founder's list in the\n  to-do file; the last column says which. The free lines about public pages were read at the time.\n  The pairs between research notes were not.)*\n\n| Public page, and its sentence | What stands against it | Checked | On his list before today |\n|---|---|---|---|\n| the page on bridging: *The people in question shared a room while being shot at* | the research note on that study: anonymous written responses and votes, a self-selected sample, no control | both ends | **no** |\n| the page of open problems: evidence shared between parts *is discounted so it isn't counted twice* | the discount is built and not switched on | the page; the other end from the instruction file | yes |\n| the architecture page: the stream path *writes NO forensics row*, marked open | the code writes one, with a comment that it closes that gap | both ends | yes |\n| the vision page: *two-axis evaluation* as part of what is new | the page of interaction principles records the decision of 2026-03-31 for one axis | both ends | yes |\n| the design guide: *23 UX principles* | that page holds 25 | both ends | yes |\n| the history page: *250 users insufficient for network effects* | the analysis of the prototype's database: *10 real users* | both ends; which figure is right is his to say | **no** |\n| the page on the earlier company: *Internal reference document — not listed* | the repository's front page: everything in that folder is published when committed | both ends | yes |\n\n  Four more are named and are not counted here: a figure of *at least 84%* given for an item where\n  the research note has 90% (true as written, and loose); a sentence that draws a prediction from\n  assembly theory; a claim that nobody has broken the circularity of a checker of the same model\n  family; and a graph said to be on claim pages that was taken off them. The last two were not\n  opened.\n- **Statements that are stale on their face**, where a note contradicts its own later section.\n- **Searches that found nothing**, which is a record nobody else keeps: the kinship reader lists\n  fifteen patterns from other fields that it looked for and did not find in the corpus.\n\nTwo of the seven also stand in the instruction file every session loads (the stream path, and the\ncount of principles). Those are repairs for a fresh session and are listed in `TODO.md`.\n\n## 4. What the corpus holds on analogy, and what it says a judge should ask\n\n**Read for this note**: [the-analogy-daemon.md](../the-analogy-daemon.md) (the design, its contract\nand its rehearsal), [what-a-rule-rests-on.md](../what-a-rule-rests-on.md) (philosophy, psychology,\nthe live graph, the strength code; whole), [convictions-across-scales.md](../convictions-across-scales.md)\n(the four checks and their verdicts), [perceiving-possibility-and-the-analogy-organ.md](../perceiving-possibility-and-the-analogy-organ.md)\n(Hofstadter's programme), [the-harness-conviction.md](../the-harness-conviction.md) § 8 (two\nanalogies worked through by hand), [the-irit-harvest.md](../the-irit-harvest.md) § 3 (four\nconsiderations from argumentation theory), the questions the code hangs on a link of the analogy,\nexample and precedent kinds, and last night's\n[research note](what-research-says-about-finding-far-analogies.md). Not reopened: the dossier under\n`what-a-rule-rests-on/`, the note on major transitions, and the thirteenth listening piece, each of\nwhich the notes above summarise.\n\n**What it comes to, as questions for whoever tests a likeness.**\n\n1. **What do the two cases share, said with neither subject?** This half can be answered from the\n   cases, and lining them up is what picks out the candidates.\n2. **Does what they share bear on what is concluded?** This half is a claim about the world, it\n   changes with the conclusion, and it is where people disagree. In the strength code it is the\n   link's relevance.\n3. **What is carried over, and what is concluded from it?** The same earlier case supports opposite\n   conclusions depending on what is mapped onto what.\n4. **What does the likeness predict?** The features of the first case not yet known to hold in the\n   second are candidate questions for the second. In the worked example of the sea chart, this is\n   how a gap was found: charts push corrections to those who relied on them, and the graph has no\n   such channel.\n5. **Is the thing claimed true in the earlier case?** One of the three questions the code already\n   asks.\n6. **Is there a difference between the cases that breaks it, and which is the deepest?**\n7. **Is there a like case where the conclusion fails?** The design requires every likeness to be\n   kept with its strongest case against.\n8. **Does it recur at two or more independent levels? Is it the same mechanism or only the same\n   shape? What opponent does the new setting add? What did the earlier case need in order to work,\n   and is that present?** The four checks, with four verdicts, one of which is *lowered*.\n9. **Which rival comparisons are there?** Where a comparison is disputed they are written side by\n   side and none is chosen.\n10. **No figure.** A likeness gives plausibility and no degree.\n11. **The likeness is asked about before the conclusion is shown.** People judged an earlier case\n    more like the present one when its outcome suited them, and training did not protect them.\n\n**How a likeness was tested by hand here**, in the comparison of the sea chart and the audited\nledger: the properties of each are listed and mapped one by one; what makes the earlier thing work\nis named and asked of the new one (*double entry works because exchange conserves value, and\nreasoning has no conserved quantity, so it carries over as a shape and never as arithmetic*); the\ndeepest difference is named; the attacks the earlier thing has met are listed; and the two rivals\nare set in one table. That is a different act from a yes or a no, and it is what a judge who\n*thought deeply* would return.\n\n**Set beside the judge's brief of § 2**: of these eleven, the brief asked the first in a narrow\nform, part of the third, and the sixth.\n\n## 5. What research says\n\nThe three reports are in [reading-for-insight/](reading-for-insight/). What bears on the design:\n\n**On kinds of insight, and how people read for them.**\n\n- **Three routes to insight in 120 real cases**: noticing a contradiction, seeing a connection, and\n  being forced past an impasse by dropping a weak assumption. Those who gained the insight took the\n  odd fact seriously where others with the same information explained it away. The authors set the\n  effort to gain insights beside the effort to reduce mistakes as two different efforts (Klein and\n  Jarosz 2011; abstract read by Claude, the counts by a helper).\n- **A statement is interesting when it denies something its audience assumes**, and dull when it\n  affirms, is unrelated, or denies everything (Davis 1971, as a helper read it). The project's own\n  test for what earns a line in its instruction file is the same test.\n- **Reading many texts on one subject has known steps**: find the passages, bring the authors to\n  one vocabulary, frame the questions that all of them can be read as answering, range the answers\n  against each other, and rule on nothing (Adler and Van Doren, second-hand).\n- **Syntheses of qualitative studies have three moves**: translating one account into another's\n  terms, explaining why two accounts refute each other, and building a whole that no account holds.\n  The second is the one most often skipped (Noblit and Hare, through a methodological review).\n- **Looking for the case that does not fit is a recognised step and is rarely done.**\n- **Questioning what a literature assumes is a method with steps**, distinct from spotting gaps:\n  state the assumptions as assumptions, weigh them, and build the alternative (Alvesson and\n  Sandberg 2011, as a helper read it).\n\n**On machines reading a whole corpus.**\n\n- **A reader finds what it was sent for.** In software inspections, a reviewer given one role found\n  more of that role's kind of fault and no more of the others, and roles overlapped less than\n  readers with one checklist (Basili and colleagues, as a helper read them).\n- **Coverage of planted insights is at the level of people, and telling where each came from is far\n  below** (Laban and colleagues 2024). On questions that need a whole novel in mind, readers scored\n  about 97% and the best model about 56% (Karpinska and colleagues 2024).\n- **Noticing what is absent is weak, and thinking longer hardly helps. Marking the empty place\n  helps a great deal** (AbsenceBench; the headline figure read by Claude, the rest by a helper).\n- **Naming themes in advance and applying them is at the level of people. Finding themes nobody\n  named is not**, and the themes a machine proposes are too broad.\n- **Different models find different things, and discussion between agents makes them alike.**\n- **Ideas a model generates look new and weaken when carried out, and a model cannot rank its own.**\n- **Nothing measures how much of what a careful person would find a machine reading recovers.**\n\n**On inconsistencies.**\n\n- **They are common and findable with help.** In Wikipedia at least 3.3% of facts contradict\n  another fact, and editors with a tool that fetches the candidates found 64.7% more in the same\n  time. The best fully automatic system reached 75% on the standard measure (Semnani and colleagues\n  2025; abstract read by Claude).\n- **More than half were numbers**, and about half of the flags were no contradiction: two things\n  with one name, rounding, an undated fact compared across periods, a belief reported as a fact.\n- **Asked plainly, a model flags nearly everything.** On pairs of code and documentation the flag\n  rate was 98%, since documentation says less than code. Separating *says something wrong* from\n  *says less* brought it to 14% (DocPrism; abstract read by Claude).\n- **The best model found 64% of inconsistencies planted in long documents**, fewest where the clash\n  is in structure and not in a number, and it also found real ones the authors had missed (FIND;\n  abstract read by Claude).\n- **Law sorts conflicts by how much of their scope overlaps**, and calls a conflict apparent when\n  the later, the higher or the more specific rule settles it.\n\n**On judges.**\n\n- **A model judge leans lenient, is moved by the order and wording of what it is shown, and agrees\n  with experts less than experts agree with each other.** A panel of three model families did\n  better than the single strongest judge.\n- **Planted items measure a judge.** In one study judges missed more than half of planted drops in\n  quality.\n- **People only work out their criteria while grading real output**, so a judge's instructions\n  cannot be fixed before a person has graded examples (Shankar and colleagues 2024, a small study).\n  This project holds the same as a principle: a person reads what a machine pass produced before\n  anyone proves how often it is wrong.\n\n**Not found by anyone**: a study of judging a pair of passages with and without the wider document;\na comparison of a strong model reading few documents against a cheap one reading many, scored on\nwhat is not obvious; a system that has *tension* as a kind.\n\n## 6. The design\n\n*Claude's proposal. Each part names what it rests on.*\n\n### 6.1 Kinds of finding, in four families and an open one\n\nDrawn from § 3, § 5 and the kinds of [likeness-seen-from-outside.md](likeness-seen-from-outside.md).\nBy the project's own law the list will break on the next pile, so the open kind is asked first.\n\n| Family | Kinds | The index today |\n|---|---|---|\n| **It fits** | answers or narrows; supplies a missing piece; the same thing done elsewhere; a practice and its ground; two cases of one wider thing; the same thing found several times and never collected; what follows from two notes together that neither says | most of the 661, the last two kinds hardly at all |\n| **It does not fit** | a number, date or status that differs; overtaken by a ruling; one word in two senses; a rule and a practice that breaks it; one mechanism judged oppositely; two held ideas that conflict in some case; the case that breaks a pattern; an odd result explained away | the first two well; the rest hardly |\n| **It is not there** | a question never answered; a prediction never scored; a test designed and never run; the objection nobody raised; what a group of notes all assume and none argues | not asked for |\n| **It rests on little** | a claim others build on, with one case or a machine's own reading under it; a citation that does not say what it is cited for | found by hand, a few times |\n| **Something else** | whatever the reader would tell the founder that none of the above covers | the free line, last, and taken up only where it spoke of a public page |\n\n### 6.2 Ten parts\n\n1. **One family to a pass.** The reader that was sent for kinship alone returned fourteen where the\n   reader asked for everything returned three of that kind. The inspection studies say the same, and so does the\n   remark in the insight study that hunting errors and gaining insight are two efforts.\n2. **The open question first, in a pass of its own.** A free line at the end of a capped form got\n   real findings, and only its remarks on public pages were taken up. The same thing happened to the judge's question of worth. What is\n   asked last is answered last.\n3. **No cap, and no *best first*.** The reader lists everything it read at both ends, each in two\n   lines with its quotes. The second stage is the strict one.\n4. **No exclusion of notes the target names**, for the two dissonant families.\n5. **The reader carries what the project holds.** The helpers do not load the instruction file, so\n   every plain reader read the notes without the convictions, the principles or the index of\n   topics. The kinship reader was told to read them first. For the dissonant families a reader gets\n   the settled items in full, since the outside check found the conflict with the full text and\n   missed it with the headlines.\n6. **A card for each load-bearing passage, with more slots, paired by code.** Last night's card has\n   an aim, a move, a condition and names for the move. Five more slots, each written with the\n   subject taken out: what kind of claim it is; what it rests on; what would have to be true\n   elsewhere for it to be false; what follows if it holds; what it leaves open or promises. Each\n   family is then a pairing of slots between cards:\n\n   | To find | Pair |\n   |---|---|\n   | the same thing done elsewhere | move with move (tested 2026-10-10: a second net, no better than the passage's words) |\n   | a clash in other words, and one mechanism judged oppositely | a claim with another card's *what would make it false* |\n   | a practice and its ground | *what it rests on* with another card's claim |\n   | what follows from two notes | *what follows* with another card's *what it rests on* |\n   | a promise nobody kept | *what it leaves open* with nothing later that answers it |\n\n   Only the first row has been tested. The second is the method that found far kinships, turned\n   round: say the denial without its subject, and search in other fields' words.\n7. **Numbers, dates and words for a thing's state are compared by code.** More than half of the\n   Wikipedia inconsistencies were numbers. In this corpus the counterpart is counts (principles,\n   tests, sources, users), dates of rulings, and the words *built*, *live*, *open*, *unbuilt*. The\n   checks the project already runs by hand before a commit are this, and three of today's seven were\n   of this kind.\n8. **A second party with a form that fits the kind, and worth asked before the sort.**\n   - *For a likeness*: the eight parts of [likeness-seen-from-outside.md](likeness-seen-from-outside.md)\n     § 5, with the questions of § 4 above, and for a likeness that matters the property-by-property\n     table.\n   - *For a dissonance*: what each passage commits to, in a sentence each, written first. Can both\n     be true at once? If they can, is there a case where they pull apart, and which? Is one later,\n     and does it name the earlier? Is a word used in two senses? Does one speak of what is designed\n     and the other of what is built? Does one say less, or say something else? Which stands, by\n     what authority: the code, a ruling, a measurement? What should change, in which file? The\n     verdict *apparent, explained by …* is kept, since that explanation is a finding too. The code\n     already holds two of these questions for an argument from inconsistent commitment: was the\n     commitment really made, and has it been withdrawn.\n   - *For both*: decoys of the same kind in every batch; the quoted words handed over in one file;\n     the two passages' dates given; each question put in two wordings and both orders, with the\n     pairs that change side reported as unsteady; a verdict worked out by rule from the answers.\n9. **An answer key from our own record.** The log holds 81 commits that begin *correction:*, and\n   the notes hold 362 marks of something corrected, in 99 files. Each is an inconsistency somebody\n   found by hand, with the corpus as it stood just before. Checking out the commit before a\n   correction and sending a reader through the notes involved tells whether a read-through would\n   have found it. The founder's 27 recorded connections, and the times he had to ask twice, are the\n   same kind of key for the consonant families. Beside it, a small set read by hand: eight to ten\n   notes that he and Claude read slowly, with every insight written down, which no read-through has\n   seen.\n10. **What is still unfound is estimated only with four or more independent passes of one kind**,\n    and given as a floor. Two further readers on four notes found 28 findings with one in common\n    with the first reading, which says that the whole is large and gives no estimate of it.\n\n### 6.3 Which model, and how much thinking\n\n- The judge would ideally be of another model family than the readers, since a panel of families\n  judged better than one strong judge and a model favours what reads like its own. The second\n  family at hand is paid for from the founder's prepaid account and is not to be spent unasked, so\n  for now the judge is of the same family and that is said beside every count.\n- More thinking helps where a whole has to be held in mind and hardly at all for what is absent,\n  where a marked empty place does more.\n- Cards and tables are wide and cheap work. The pass across notes, the open pass and the judge are\n  the few places where a strong model with time is worth its cost. No study found says by how much.\n\n### 6.4 What no read-through will give\n\n- **What the corpus lacks as a whole**, and **the frame it was written inside**. These are the two\n  classes of error the project keeps for people, and the research on absence and on whole-book\n  reasoning says the same from outside.\n- **The connection only he can be reminded of.** Of the seven connections that bear on the sixth\n  sentence of his walk-through, the readers had found one. The other six came after he asked twice.\n- **A ranking of its own findings by what they are worth.**\n\nSo the reading by hand stays, and the read-throughs can aim it: the notes with the most dissonance\nfound are where he and Claude reading slowly will find the most.\n\n### 6.5 What reaches the founder\n\nHis attention is the scarce thing, and *any kind of insight* over four hundred notes is thousands of\nfindings. The order follows what he is deciding: sentences on public pages, which he goes through\none at a time; lines of the instruction file; the notes behind proposals that await his ruling; and\nthe rest, which is kept in the index for sessions to be handed at the moment they open a note.\nWhere a ruling of his own has overtaken a passage in a research note, Claude adds a dated line\nwithout asking, as he allowed on 2026-10-10.\n\n## 7. First steps, cheapest first\n\n1. **Owed, no cost**: go through the 26 free lines, open both ends of each pair they name, and keep\n   what stands. The two public-page sentences that were new today are on his list.\n2. **Owed, no cost**: the revised briefs, written out: one for each family, and the two forms for\n   the second party.\n3. **His time**: the sitting he asked for, over\n   [what the judges wrote](cards/what-the-judges-wrote-for-each-pair.md). By the finding on\n   criteria, this is the step that turns the judge's instructions from a first guess into\n   something tested, so it comes before any wider judging.\n4. **A pilot on the same four notes as last night** (three of them public pages), so the counts\n   compare: one reader for *it does not fit*, with the settled items in hand and no exclusion, and\n   one reader for the open question alone. About the cost of last night's two test readers.\n   Written before any run:\n   - the dissonance reader returns at least ten findings that stand when both ends are read, at\n     least three of them of a kind the index holds fewer than five of (a tension, one mechanism\n     judged oppositely, a clash with a named note, a clash with the code);\n   - the open reader returns at least five things that fit none of the four families;\n   - against the design: if the dissonance reader returns mostly what the free lines of § 3.3\n     already hold for those notes, the old form saw as much, and the gain is in keeping what is\n     seen and not in a new brief.\n5. **The answer key from corrections**: a script that lists each correction with the notes it\n   touched and the commit before it, and a first replay on ten of them.\n\n### 7a. The pilot, run 2026-10-11\n\nAsked whether to run step 4, the founder: *\"Sounds great\"*. Four readers went through the four\nnotes: two Sonnet helpers of the app, as proposed, and the same two briefs on Haiku 5.5 through the\ncommand-line tool, which Claude added without asking so that the two models could be compared on one\ntask ([haiku-5-5-for-the-filling.md](haiku-5-5-for-the-filling.md) § 6). The briefs are\n[dissonance-reading-brief.md](dissonance-reading-brief.md) and\n[open-reading-brief.md](open-reading-brief.md); each reader was handed everything the founder has\nsettled, in one file. The reports are in `readers/`, word for word, dated 2026-10-11.\n\n**The open question, Sonnet.** Sixteen remarks, and all 40 passages they rest on stand at the lines\ngiven. What came back that no earlier reader returned:\n\n- **Three faults in the code, found by reading it beside the page.** The bridging feed reads the\n  links that leave a claim and not the ones that arrive at it, two votes are enough for the top\n  score, and the check for sacred commitments has one caller and guards neither the splitter nor\n  the pass for unwritten premises. The session confirmed each in the code. They are filed in\n  `TODO.md` with the build they belong to.\n- **Sentences stronger than what they rest on**: research that *suggests* in the note it came from\n  and *confirms* on the page; a cause stated as fact that another note lists as open; a claim in\n  bold with no source, beside a cited study that found nothing for arguments alone.\n- **An example that shows what it was built to show**: the page's demonstration is a chain written\n  by hand, with its last step in one side's voice.\n- **A feature whose success is described in the words the same page uses for its failure.**\n- **A defence against a danger the design does not have**, while the one it has goes unnamed.\n- Unbuilt things in the present tense, and four surfaces still using names he has replaced.\n\n*Predicted: at least five things that fit none of the four families. Held, on Claude's own sorting:\nthe example built to show its result, the feature and its failure, the defence at the wrong door,\na test that performs the harm it looks for, and the three faults in code.* The sorting is one\nreader's, and the families are loose enough that another would sort some of these under *it rests\non little*.\n\n**The open question, Haiku.** Its first five remarks are lost: it wrote its report in two messages\nand the tool returned the last. Of the eight that came back, all 23 passages stand at their lines.\nThree are also in the Sonnet report and five are not, among them that a page's sharpened line (the\nsoftware as the lens) and the same page's later finding (the interface is not the instrument)\ncannot both be its position.\n\n**What does not fit, Haiku.** Ten findings, all twenty passages in place, and all ten stand when\nboth ends are read. By its own labels: two of a thing's state, two overtaken, two counter-cases,\ntwo tensions, one rule against a practice, one of a kind it named itself (a qualifier dropped).\n*Predicted for this brief: at least three of a kind the index holds fewer than five of. Held: two\ntensions, two clashes with the code, one inside a single note.* None of the ten is among what the\nearlier readers' closing lines held for these notes, so the check written against the design did\nnot fire. **Six of the ten cannot be kept in the index**, which holds pairs of notes and ledger\nrows: their other end is the code, the to-do file, the instruction file or the same note.\n\n**What does not fit, Sonnet.** Fourteen findings, all 28 passages in place, and all fourteen stand\nwhen both ends are read (two were checked further, in the code and in a results file). By its own\nlabels: four overtaken, three of a thing's state, three of one word in two senses, and one each of\na counter-case, a number that differs, a rule against a practice, and a kind it could not name.\n*Predicted: at least ten that stand, at least three of a kind the index holds fewer than five of.\nHeld: three words doing two jobs (scaffold, bridging, and what the bridging score rewards), three\nclashes with the code, and a note contradicted by its own results file.* Seven are kept in the\nindex. The other seven have the code, the to-do file, the instruction file or the same note as\ntheir other end. It made 180 tool calls in 24 minutes.\n\n**The two models on this brief.** Haiku returned ten and Sonnet fourteen, and they share one\nsentence (a note saying the corpus lacks a citation it has since gained) and one fault seen from\ntwo sides (the check for sacred commitments). Twenty-three distinct findings from two readers of\nfour notes, all standing.\n\n**What the pilot says (Claude's reading).**\n\n1. **The open question, asked alone, returned the most useful things of the day.** None of them\n   is a pair of notes. Each is a remark with a consequence, and the index has no place for one.\n   They need a home of their own, and a way of reaching the founder that is not a list of pairs.\n2. **A reader with the code in reach finds faults in the code.** No earlier brief sent a reader\n   there. The three found are in shipped behaviour and had survived every test.\n3. **Two readers of one brief hardly overlap.** On the open question they shared three remarks, and\n   on what does not fit one finding of twenty-three. Most of what can be said about four notes was\n   said by one reader only, which is § 6.2 part 10 again: the whole is large, and a second reading\n   by another model is worth more than its price.\n4. **Most of what came back is about pages written for visitors**, since three of the four notes\n   are such pages. Each sentence goes to him one at a time, with its whole record gathered first.\n   The new candidates are listed in `TODO.md`.\n5. **The cost.** The two Sonnet readers made 82 and 180 tool calls, in fifteen and twenty-four\n   minutes. Each Haiku reader took six to eight minutes and, at list price, well under a dollar. No\n   money was paid for any of them: all four run on the subscription.\n\n### 7b. Both briefs over the other pages written for visitors, begun 2026-10-11\n\nAsked whether both readers should go over all 24 pages at the top of `docs/`, the founder: *\"Sure do\nit\"*. Three of the 24 were read in the pilot. The other 21 hold about 78,000 words, and go out in six\nbatches of 12,000 to 14,000 words, each read once for what does not fit and once for the open\nquestion, by Sonnet helpers of the app, three at a time:\n\n| Batch | Pages |\n|---|---|\n| a | `architecture-reference.md` |\n| b | `ux-principles.md`, `design-guide.md` |\n| c | `vision.md`, `civilizational-vision.md`, `the-missing-layer.md` |\n| d | `conceptual-threads.md`, `convergence.md`, `analysis-and-attunement.md` |\n| e | `technical-direction.md`, `object-model.md`, `extraction-pipeline.md`, `inductive-reasoning-research.md`, `academic-foundations.md` |\n| f | `open-problems.md`, `frontier.md`, `competitive-landscape.md`, `history.md`, `history-delibr-era.md`, `community-and-outreach.md`, `sketches.md` |\n\nThe briefs are the two of the pilot, unchanged, with the settled items written out afresh (135 items).\n`frontier.md` is generated from the project's instruction file, so a finding about it is a finding\nabout its source.\n\n**Written before any run.**\n\n1. At least 95 in 100 of the quoted passages stand at the lines given (the pilot: all of them).\n2. Each open reader returns at least eight remarks. Across the six batches at least six faults in\n   shipped code or scripts are named and hold when the code is read (the pilot: three, from four\n   notes).\n3. Each reader for what does not fit returns at least ten findings that stand when both ends are\n   read, and at least half of all of them have as their other end something the index cannot hold:\n   the code, the to-do file, the instruction file or the same note (the pilot: 13 of 24).\n4. Batch a, the description of the built system, returns at least twice as many clashes of state\n   with the code as any other batch.\n5. Fewer than one in five of the page sentences that come back are already among the 39 on the\n   founder's list.\n\n*Against the plan.* If more than a third of the findings for what does not fit fail when both ends\nare read, a batch of this size is too large for the brief. If the open reader of batch f, the\nhistorical pages, returns fewer than five remarks, or mostly remarks that would be true of any old\npage, then such pages do not need an open reading and a check of their stated facts by code would do.\n\n**A place for what comes back, built first.** The pilot left two things with no home: a remark\nfrom the open question, and a finding whose other end is no note (§ 8, questions 7 and 9). With\nsome 150 more on the way, `remarks.py` now gathers both from the reports, in the readers' words,\ninto [what-readers-said-that-is-no-pair-of-notes.md](what-readers-said-that-is-no-pair-of-notes.md),\nordered by the note and line each is about. Under each stands what a session found on opening it\n(`remarks.jsonl`). `remarks.py --check` lists quoted passages that are no longer at their lines,\nwhich is how a changed sentence shows up. That page is the list for the founder's walk-through of\nthese pages. The numbered list in `TODO.md` stops at 39.\n\n**The open question, batches b and c** (the page of UX principles with the design guide; the three\nvision pages). Eighteen and nineteen remarks. All 102 quoted passages stand at their lines. The\nreaders made 54 and 55 tool calls in nine and twelve minutes, on 207,000 and 249,000 tokens. Opened\nby the session: 36 stand and one stands in part, its other half overtaken by a build of 30\nSeptember. None repeats a sentence already on the founder's list, though two are about lines that\nare (the walk-through's sentence 6, and the sibling of sentence 13). What they are, in kinds:\n\n- **A correction that reached one surface and not its neighbours.** Quality is computed and never\n  voted since March, and three places on the same page still have people rating it. The positioning\n  page was changed on 10 October to say where structure-as-output holds, and the vision page makes\n  the same promise unchanged. Outward copy was asked to drop one image of the lens, and three\n  surfaces still use it.\n- **A page that describes what is not built as how the product works**: four of the UX principles,\n  a landing page that was since reordered on his word, a colour system the code does not follow\n  (the four semantic colours are written out 58 times and used through their names 6 times).\n- **A claim about the whole that the project's own notes narrow**: predecessors failed on\n  experience alone; the prototype died on one obstacle; four arrivals called independent, where the\n  project chose which sources to read; a count of what humanity shares.\n- **A limit stated on one page and missing on the page that needs it most**: lowering the cost does\n  not settle a conflict of interests, said on the vision page and absent from the page about\n  consequences at the scale of a civilisation.\n- **The page's own voice**: on the vision page the first person is the founder in some sections and\n  Claude in another, and the page tells the story of its own drafting.\n- **A named person.** A research note names the author of a public post, with the party her brother\n  works for. Put to the founder the same hour. **Ruled 2026-10-11** (founder: *\"The FB post name is\n  fine\"*): she stays named, and the note is unchanged.\n\n*Prediction 2 so far: held for remarks (18 and 19 against eight). No fault in shipped code was\nnamed by these two; what they found against the code is pages that describe it wrongly.*\n\n**Batch a, both briefs** (the page that describes the built system). Fifteen findings and fifteen\nremarks; 183 and 66 tool calls in eighteen and ten minutes, on 349,000 and 228,000 tokens. Opened by\nthe session: two findings are pairs of notes and went into the index, twelve of the rest stand and\none stands in part; thirteen remarks stand and two in part. This is the batch where the readers\nfound faults in the code itself, mostly by looking for the caller of a function the page describes:\n\n- The read log never deleted anything: the function that forgets after ninety days had no caller.\n  And seven places in extraction printed a failure where the project's rule asks for a logged one.\n  Both repaired the same day, with tests.\n- A web address was checked, and then looked up a second time when it was fetched, so the check\n  could be passed with one answer and the fetch sent to another. Repaired later the same day\n  (`deliberus/safe_fetch.py`: one lookup a hop, and the connection goes to the address that passed).\n- The paid model route keeps its spending record under a home folder that the production container\n  does not have. It matters on the day that route is switched on.\n- The page is public and carried operations detail about the machine it runs on, and signed-in\n  readers are not told that their reads are logged. Both were the founder's to decide. On the first\n  he ruled the same day (*\"Ok sure split the server info into public/private, makes sense\"*) and the\n  page was split; the second is open (cluster A2 of [the decision clusters](decision-clusters.md)).\n\n**What does not fit, batches b and c.** Nineteen findings each; 128 and 158 tool calls in fifteen\nand nineteen minutes, on 325,000 and 385,000 tokens. Every quoted passage was found, one of them\nfourteen lines from where the reader put it. Opened by the session: 21 are pairs of notes and went\ninto the index (7 and 14), 15 of the rest stand and 2 stand in part. In kinds:\n\n- **The design guide describes styling the code no longer has**: one width for every page, where\n  widths of 860, 900, 980 and 1040 pixels also exist; a blur that was removed in March; a solid\n  card colour that is translucent; a landing layout from before 19 September; a graph that is on no\n  page.\n- **A sentence with *every*, *never* or *cannot* that the project's own measurements narrow.**\n  Every badge asserted the opposite of the evidence (3,401 of 3,828 had no evidence to oppose). A\n  fake citation cannot be made in a claim graph (it can, and the check is unbuilt). Strength is\n  never a model's opinion (the number printed beside a link is one). No tool has ever tested the\n  thesis (the page beside it calls one study a confirmation).\n- **A number, date or source that differs from the note the page relies on**: a profile dated six\n  months late, a figure on trust credited to authors no note cites, one stay placed in 2012 on one\n  line and in January 2013 on the next.\n- **A principle the code breaks**: saving a reworded machine-written part sets its text in place,\n  so the earlier wording is gone, against *contributions are added, never overwritten*.\n- **A rule the page breaks elsewhere on the same page**: no gap without a move, and the maps show\n  gaps with no move; people vote agree or disagree only, and the first-use path asks a newcomer to\n  rate quality.\n\n**The open question, batch d** (three concept pages: the cross-cutting threads, convergence,\nanalysis and attunement). Seventeen remarks; 68 tool calls in fifteen minutes, on 312,000 tokens.\nAll 40 quoted passages stand at their lines, and all seventeen remarks stand on opening; three rest\nin part on the reader's own knowledge of Polis, Lean and Rorty, which was not checked. In kinds:\n\n- **A correction placed on one sentence**, while three passages on the same page still need the\n  design that was dropped (a second vote, on quality).\n- **An item presented as open that is closed**: twice on the convergence page (a horizon is dated\n  since 16 September; the kind the founder was asked to choose exists in the code), and once on the\n  threads page (the split between is and ought is called invisible, and was published as a claim on\n  22 August).\n- **The wager's losing condition has no number**: no count, depth or cut-off is given anywhere, and\n  the reading counts leaves, so splitting finer lowers the fraction with nobody converging.\n- **The page's lead evidence is not read by the page's own newer rule**, which asks at which level\n  an agreement was found.\n- **One word for three things**: attunement is a person's practice, a space of embeddings, and the\n  acts kept for people.\n- **A study cited for more than it carried**: the canvassing result is narratives in one\n  conversation, cited as sustained exposure to reasons; and a cost the ledger withdrew on 27 August\n  still stands on the page as a cost.\n- **A remark about aim**: the feed is described as surfacing what is most likely to change the\n  reader's mind, which is what a persuader aims at.\n\n**The five predictions, on seven runs of twelve.**\n\n1. Held. All 183 passages quoted by the four open readers stand at their lines, and every passage\n   quoted in the three readings for what does not fit was found, one of them fourteen lines off.\n2. Held for remarks (15, 17, 18 and 19 against eight). Five faults in shipped code were named and\n   hold when the code is read: the read log, the unlogged failures, the address looked up twice,\n   the spending record's folder, the reworded part. Six was the mark, with five runs not sent.\n3. Held. 15, 19 and 19 findings stand, three of them in part, and 30 of the 53 have an other end\n   the index cannot hold. The vision pages are the exception: 14 of their 19 are pairs of notes.\n4. Did not hold. By the readers' own labels the description of the built system returned two\n   clashes of state, the vision pages six and the UX pages two. Counting every finding in which a\n   page is behind the code or behind a later note (state, overtaken, number or date), the three\n   batches returned 9, 12 and 14.\n5. Held, and the check is weak: one of 122 targets lies within two lines of a page line the to-do\n   list names, and the list of 39 is mostly about other pages.\n\n*Against the plan.* No finding failed when both ends were read, so a batch of this size is not too\nlarge for the brief. **That zero needs a caution.** Of 122 findings and remarks, 23 went into the\nindex as pairs that stand, 93 were judged to stand and 6 to stand in part, and none was judged not\nto stand; the judge is the session that sent the readers. A reader that is never wrong and a judge\nthat never says so look the same from here. The founder's walk-through is the check that counts. A\ncheap one beside it: a second reader given only the findings and the passages, asked which do not\nhold. The test on batch f, the historical pages, was not run.\n\n**Held on the founder's word** (*\"Let's think it over before sending the last readers\"*): what does\nnot fit over batch d, and both briefs over batches e and f.\n\n## 8. Open questions\n\nFor the founder:\n\n1. ~~Whether to run the pilot of § 7, step 4.~~ Run on his word, 2026-10-11 (§ 7a). Now: whether the\n   same two briefs go over all 24 pages written for visitors, with what they return put to him one\n   sentence at a time.\n2. Whether a judge of another model family may be paid for from his prepaid account, for a small\n   batch, to learn how much the family matters.\n3. Whether he wants to read eight to ten notes by hand with Claude as the answer key of § 6.2,\n   part 9, and which.\n\nFor the design:\n\n4. Whether the five further slots of a card can be written steadily by a small model.\n5. Whether the denial of a claim, said without its subject, finds a far clash more often than the\n   claim's own words do. Untested.\n6. What a decoy is for each kind of dissonance.\n7. How the index should mark a tension, which is neither a contradiction nor a correction.\n   And where a clash with the code, with the to-do file or inside one note is kept, since the\n   index refuses a pair whose other end is not a note (§ 7a).\n8. Whether the open pass returns more than the free lines already did. *It did (§ 7a).*\n9. Where a remark from the open pass is kept, and how it reaches the founder.\n\n## 9. Corrected with this note\n\n- The spec's § 18.6, which did not say that a reminder had reached the helpers of the blind test.\n  Checked and recorded there: the counts do not move.\n- [likeness-seen-from-outside.md](likeness-seen-from-outside.md) § 7, point 7, gave a figure for\n  how many quoted passages stand in the line handed over that could not be reproduced. It was\n  corrected before that note was committed.\n- Two statements in the project's instruction file that the code and a public page have overtaken\n  (§ 3.3). Listed in `TODO.md` for a fresh session, and not changed here.\n- **The word *public page*.** This note and the reply that reported it called the 24 pages at the top\n  of `docs/` the public pages, as the scripts do. The founder read that as saying the other notes\n  are not public. Every note under `docs/` is served. Said so here, at the top of the spec and in\n  the glossary.\n- **This note itself, the day it was written.** Its first form, and the reply that reported it, said\n  that seven inconsistencies on public pages had never been kept. Five were already on the\n  founder's list. The claim that something was filed nowhere had been checked against the index and\n  not against the to-do file. Checked against both, two of the seven were new, and of the 58 pairs\n  about fifty are in neither.\n\n## 10. Other uses of the same findings (his question, 2026-10-11)\n\n*\"Any other ways you can think of, of applying what's in the reading-for-every-kind-of-insight\nwriteup?\"* Claude's list, most at stake first. Each is a proposal, none is built, and where one\ntouches the product it waits behind the open reviews like everything else.\n\n**Before something goes out to a reader who will check it.**\n\n1. **An application, read against the site.** A funder reads the application and then the pages.\n   The reader for what does not fit, sent through the draft with the pages and the code in reach,\n   returns the sentences where the two disagree. The pilot's cheap reader found ten such sentences\n   in four notes in six minutes. This is the use with a date on it.\n2. **The pages written for visitors, all of them, before any wider showing.** The same pass over\n   the 24 pages at the top of `docs/`, one family at a time. What it returns goes to him one\n   sentence at a time, as now.\n\n**In the product.** These are the same findings turned on claims in place of notes.\n\n3. **A linking pass that looks only for what clashes.** The linking step asks one question for\n   support, attack and qualification together, and a reader finds what it is sent for. Between\n   claims taken from texts the graph holds about four supports for every attack. A pass sent only\n   for attacks, with the three tests of the new brief (saying less is no clash; name the case; read\n   both ends), is the same design.\n4. **Find a clash by searching for a claim's denial.** Nearness by wording cannot tell a claim from\n   its opposite and does not reach a clash said in other words. Writing for each claim what would\n   make it false, and searching for that, is untested here and cheap to try on two opposed sources.\n5. **Numbers compared by code.** More than half of the inconsistencies found in Wikipedia were\n   numbers. Two sources giving different figures for one quantity can be paired with no model.\n6. **A marked empty place.** A model asked what is missing does poorly, and far better when the\n   gap is marked. The critical questions are marked places, and the proposal to write every link out\n   in one form is another. The pass for unwritten premises still asks the open question.\n7. **How much did extraction miss?** Nothing measures it. Two or more independent readings of one\n   source, by different models, give a floor for what all of them missed, by the same arithmetic as\n   § 6.2 part 10. At Haiku's price that is affordable for the first time\n   ([haiku-5-5-for-the-filling.md](haiku-5-5-for-the-filling.md)).\n\n**In the reviews and the instruction files.**\n\n8. **One word doing two jobs, before a tile is ruled.** The third of the headline tests asks whether\n   a word already does a job in the corpus. That is one kind of not fitting, and a reader sent for\n   it alone could list the words in a candidate tile that do. Today gave two: *move*\n   ([what-a-move-is.md](what-a-move-is.md)) and *public page*.\n9. **The instruction files themselves.** They are loaded into every session, so a pair of rules that\n   give opposite orders in one case misleads every session. The same brief, with the rules as the\n   notes.\n10. **Numbers that decay, checked by a script.** Counts of tests, sources and principles go stale in\n    prose. A list of each such number with the command that produces it turns a hand check before\n    every commit into one run.\n\n**For measuring readers and models.**\n\n11. **The answer key from corrections** (§ 6.2 part 9) grades any new reader, brief or model, as the\n    card test and the judges' decoys did for Haiku.\n12. **Themes from the index.** With about 750 findings kept, the notes can be grouped by code from\n    what was found between them. This is the one part of the graph-and-summaries method\n    (*GraphRAG*) that the index now has the material for: a view of which notes cluster and which\n    stand apart. A written summary of each cluster is the part to be careful with, since a summary\n    drops the rare item first.\n\n## See also\n\n[likeness-seen-from-outside.md](likeness-seen-from-outside.md) ·\n[what-a-move-is.md](what-a-move-is.md) ·\n[dissonance-reading-brief.md](dissonance-reading-brief.md) and\n[open-reading-brief.md](open-reading-brief.md) (the two briefs of the pilot) ·\n[what-the-readers-noticed-beyond-their-forms.md](what-the-readers-noticed-beyond-their-forms.md) ·\n[README.md](README.md) § 17 and § 18 ·\n[reading-for-insight/](reading-for-insight/) (the three research reports) ·\n[the-analogy-daemon.md](../the-analogy-daemon.md) ·\n[what-a-rule-rests-on.md](../what-a-rule-rests-on.md) ·\n[../outside-checks/README.md](../outside-checks/README.md) ·\n[the-residual-error-taxonomy.md](../the-residual-error-taxonomy.md) (the classes of error no single\nstep owns).\n"}