{"path":"listening/2026-10-13-three-days-to-the-floor.md","content":"# Three days, down to the floor of a step: a listening walkthrough (2026-10-13)\n\n- **What it is**: the sixteenth listening piece and the second catch-up, on the founder's request:\n  *\"Can you make me an audio walkthrough of about an hour detailing everything that happened since\n  (and including a very short recap of) the last audio walkthrough?\"* The last one was\n  [the chain folded inside an arrow](2026-10-10-the-chain-folded-inside-an-arrow.md), ten minutes\n  on the evening of 10 October. This one covers that evening to the afternoon of 13 October.\n- **Rendered**: 6,688 words, 56 minutes 42 seconds, twelve parts and a closing view; 179 marked\n  pauses, 43 breaths, 11 slowed sentences. Written from the session record\n  (`SESSION-HISTORY.md`, the entries of 10 to 13 October) and from the notes each part names\n  below, every one committed before it was written.\n- **Covers**: a recap of the last piece and two corrections to it · what the graph holds, the\n  four sorting questions, and the ruling that a machine's sentence is a claim the moment it is\n  written ([claim-or-system.md](../research/claim-or-system.md)) · the readers' findings as the\n  reviews seen from the other end ([the corpus-graph spec](../research/corpus-graph/README.md)\n  § 14) · a likeness seen from outside, and reading for every kind of insight\n  ([reading-for-every-kind-of-insight.md](../research/corpus-graph/reading-for-every-kind-of-insight.md))\n  · the purposes and the outside model's reading\n  ([what-deliberus-is-for.md](../research/what-deliberus-is-for.md),\n  [an-outside-reading-of-the-thesis.md](../research/an-outside-reading-of-the-thesis.md)) · the\n  twelve books and the *Tractatus*\n  ([twelve-books-in-plain-words.md](../research/twelve-books-in-plain-words.md)) · a part and a\n  reason ([a-part-and-a-reason.md](../research/a-part-and-a-reason.md)) · checks on machines and\n  people ([checks-on-machines-and-people.md](../research/checks-on-machines-and-people.md)) · the\n  floor of a step, the ten links and the ruling of 13 October\n  ([the-chain-inside-a-link.md](../research/the-chain-inside-a-link.md) § 10, § 11) · the tools.\n- **Status**: a snapshot of 13 October. The purposes are not read out in it, since neither list\n  is ruled. How these pieces are written: [how-to-write-for-listening.md](how-to-write-for-listening.md).\n- **Left out, or given one clause**: the four books read at a first level (Ziliak and McCloskey,\n  Woodward, Paul, Jullien); the decision model and the seven keys; the week read together and the\n  merged branch, with the count of duplicate commits that came from a shallow checkout; the\n  repair to how the site fetches an outside address; the five rulings that shaped the corpus\n  graph's copy; the fifth finding of the twelve books (a friend's design of one line with fixed\n  slots); the correction about the stop rule of 25 August. Each is in the session history.\n- **Where the spoken text is rounder than the notes**, so nobody reads more into the audio than\n  the notes hold.\n  1. *I was wrong about twenty times*: about twenty-two corrections in the session history's\n     entries for those days, counted by hand, most of them the session's own.\n  2. *Five kinds of thing, and fifteen kinds of arrow*: five kinds of node and fifteen kinds of\n     link, nine of them between claims, read from the live graph on 2026-10-10.\n  3. *Four things already sit under an arrow*: the linking step's reason is a text field under\n     1,481 of 1,636 support and attack links between claims from texts or people; a relevance\n     number under 190; a standing question, as a node; and a slot for the id of a claim about the\n     link, built 2026-09-28 and unused until that night.\n  4. *That list of four is my proposal*: ledger AG13, open.\n  5. *About two hundred machine sentences*: 203 claims, 174 sentences and 29 relevance claims; the\n     strength of all 52 claims at the ends of those links was the same before and after.\n  6. *A few dozen decisions, in seven groups*: 46 clusters. *More than fifty*: his words were\n     *\"postpone it until we have more than 50 users or whatnot\"*.\n  7. *Seven notes … kept one*; *a few far connections … several times as many*: three against\n     fourteen. *About half*: the two judges called 18 of 34 real pairs the same move and none of\n     20 random pairs. *Most of the refused pairs*: about ten of sixteen, in Claude's reading.\n  8. *A well-known experiment*: Drew, Võ and Wolfe, 2013, on radiologists and a gorilla in a lung\n     scan. Recalled for this piece and not opened.\n  9. *More than doubled in a month*: 133 open ledger rows on 13 September, 345 at midday on\n     11 October.\n  10. *Three of our public pages*: by the outside model's own account. *Less a truth machine than\n      a reasoning observatory* is the model's phrase.\n  11. *Four findings*: the note has five. The books were read from reviews and critics, none of\n      them themselves. *Mother* is given three tests where the note gives four, two of them\n      recalled.\n  12. *Nearly a hundred claims of that shape*: 93, none cut by the splitter yet.\n  13. *Ruled on the thirtieth of September, for now*: ledger Q34. *A third state … not built*:\n      *nothing stands (N checked)*, ruled for now the same day.\n  14. *Would have turned away two*: 2 of 32 links people added. *Most of the parts*: 28 of 39,\n      24 of them reasons. *Those few dozen parts*: the 39.\n  15. *About two thousand … one*: 2,084 standing questions on the reasons in the live graph, one\n      with an answer anybody has backed.\n  16. *They agreed … they parted*: on twelve links and four models the yes or no came back as\n      stored on 11 or 12, the direction on 10 to 12, the degree on 4 to 6. That the scale mixes\n      three numbers is Claude's reading of earlier measurements, and no run has tested it.\n  17. *A gap in about half of them*: 13 of 25 steps. *In a good share of the steps*: 8 of 29.\n      *One chain ended with no gap anywhere*: six chains of nine did; the one meant is the\n      nineteenth test link.\n  18. *Tens of millions, perhaps more than a hundred million*: Epoch's estimate of 30 to 170\n      million agents.\n  19. *The pictures* (the cookbook margin, the football match, the shoes at the door, the amended\n      law, the essay and its spelling, the kitchen nobody has cooked in, the bridge and the\n      inspector, the bachelor, the doorman, the doctor and the symptom) are Claude's own and are\n      in no note. The pill, the birds, the form with two boxes, the restaurant bill, the draught\n      and the stars are the notes' own cases.\n  20. *His quoted words* are spoken with a slash as a pause, abbreviations in full, and quote\n      marks inside a quote dropped. *A fantastic summary* translates his Swedish.\n\n---\n\n[warmly] This is a catch-up walkthrough, about an hour long. It covers everything since the last one, which is three days of work. The silences are deliberate. They are room to think.\n\n[[pause 3]]\n\nHere is the map. Twelve parts. A very short recap of the last piece. What the graph really holds. The reviews, seen from the other end. Likenesses, and reading for every kind of insight. The whole session looked at once, and what Deliberus is for.\n\n[[pause 3]]\n\nThen twelve books, over two parts. A part and a reason. Who checks the machine, and who checks the person. Then the floor of a step, in two parts, which is the one to keep. And last, the tools, briefly.\n\n[[pause 3]]\n\nOne thing to say before we start. In these three days I was wrong about twenty times. Several of those you caught yourself. I will say each one where it belongs, in the first person. [exhales]\n\n[[pause 4]]\n\n---\n\nPart one. Where the last piece ended.\n\nThe last walkthrough was ten minutes long, on the evening of the tenth of October. It had three ideas.\n\n[[pause 3]]\n\nFirst. An arrow between two claims is a chain that has been folded up. The milk is off, because it smells sour. Bacteria, acid, smell. All of it hides in the word because.\n\n[[pause 3]]\n\nSecond. Wider is not deeper. Asked what a rule rests on, the machine can always write a wider rule. That is the same rule stretched. Deeper means how the thing comes about.\n\n[[pause 3]]\n\nThird, your ruling. In your words: It may and should write deeper steps. But the microtasks in charge of doing so need to be quality-controlled, just like the other microtasks that are in play in Deliberus.\n\n[[pause 3]]\n\n[inhales] Two things I said in that piece did not last the night.\n\n[[pause 3]]\n\nI said the machine's half of the control includes asking every question twice. That was wrong. Asking a second way runs in trial scripts only, on no live task. And you had never ruled it. In your words: Asked twice? I don't think I ever blessed that particular part. [exhales]\n\n[[pause 4]]\n\nAnd I proposed that a sentence the machine writes should be nobody's claim. A question under an arrow, and no more. A few hours later you ruled the opposite. That is part two.\n\n[[pause 3]]\n\nThe piece ended on two questions for you. When you open a step, what do you want to meet first, the rule, or how it comes about? And how many machine sentences may stand on each other before a person has looked? Both are still open. [exhales]\n\n[[pause 3]]\n\nSo, part one in a line. Arrows are folded chains, the machine may unfold them, and two of my own sentences have since been corrected.\n\n[[pause 6]]\n\n---\n\nPart two. What the graph really holds, and a claim the moment it is written.\n\nRight after that walkthrough you asked a plain question. In your words: Wait, are the rules stored as claims, or what does the actual graph ontology look like? Can you draw a diagram for me?\n\n[[pause 3]]\n\nSo I drew it, from the live graph. Five kinds of thing, and fifteen kinds of arrow. One fact stood out.\n\n[slowly] A claim is the only thing in the graph that carries content.\n\n[[pause 4]]\n\nThen I told you something that was wrong. I said nothing can be attached to an arrow, so a rule is stored nowhere.\n\n[[pause 3]]\n\nAn hour later I had to correct it. Four things already sit under an arrow. The machine's own reason for drawing it, as a line of text nobody can point at. A number for how relevant it judged the link.\n\n[[pause 3]]\n\nA standing question. And a slot that can hold a claim about the arrow itself. That slot was built in September. Nothing had ever been put in it. [exhales]\n\n[[pause 4]]\n\n[inhales] You then asked the question behind the question. In your words: Ok well it makes sense to store them as claims, right? Or what might belong as a claim and what might belong as a feature of the ontology and software itself? This is an ever-present question in Deliberus.\n\n[[pause 3]]\n\nTwo pictures before the answer.\n\nThink of a cookbook. The recipe says, bake for forty minutes. Somebody has pencilled in the margin, thirty-five in our oven. Is that note part of the book? You can argue with it. Your sister can cross it out. It behaves like a line of the recipe, though nobody printed it.\n\n[[pause 3]]\n\nNow a football match. The referee's report says who did what, and in which minute. Nobody argues with that, it is what happened. The league table can be worked out again from the results. And the offside rule is neither. It is the form of the game. Two teams cannot play one match under two offside rules. [exhales]\n\n[[pause 4]]\n\nThose are the four places a thing can go. They come as four questions, in order.\n\nCould two people disagree about it? Then it is a claim.\n\nDoes it say who did what, and when? Then it belongs to the record.\n\n[[pause 3]]\n\nCan it be worked out again from those two? Then it is a reading, and it is never stored as a fact.\n\nAnd what is left is the form. The form should be as small as it can be.\n\n[[pause 4]]\n\nThat list of four is my proposal. It is open.\n\nBut one thing follows from it at once. Two people can disagree about whether a broad rise in productivity gives broad abundance. So a rule under an arrow is a claim.\n\n[[pause 3]]\n\nOne question was left. Is the machine's sentence a claim from the moment it is written? Or only once a person takes it up? I asked you that. In your words: The moment it's written. Go ahead and make it happen. [exhales]\n\n[[pause 4]]\n\nIt was built that night. A sentence the machine writes under an arrow is now a claim of its own kind. Nobody asserts it. It is held as a draft. And it moves no strength.\n\nOne wording is one claim. So a rule that recurs under many arrows is one thing you can point at.\n\n[[pause 3]]\n\n[inhales] The build taught me something, and it is a confession. That such a sentence moves no strength had to be written into the strength code by hand. And my first test of that could not fail. Its claims sat a hair above neutral, so nothing was being measured, before or after.\n\n[[pause 3]]\n\nA guard that cannot fail looks exactly like a guard that works. The test has two controls beside it now.\n\n[[pause 4]]\n\nWhat is real today? About two hundred machine sentences are in the live graph, from the two trials. Every claim at the ends of those arrows kept exactly the strength it had.\n\nWhat would it open? Anyone can attack a rule without first adopting it. And a second level has something to hang from.\n\n[[pause 3]]\n\nSo, part two in a line. Whatever two people could disagree about is a claim, and that includes what the machine guesses.\n\n[[pause 6]]\n\n---\n\nPart three. The reviews, seen from the other end.\n\nThat same night, readers went through the notes that no reader had been through. Three at a time. They brought back a list of places where a page still says what a later ruling changed.\n\n[[pause 3]]\n\nI began putting those to you, one sentence a turn. After the first one you stopped and asked. In your words: The list of inconsistencies you're surfacing for my word on here. Isn't this quite similar to what I've been trying to accomplish with the four reviews, et cetera?\n\n[[pause 4]]\n\n[inhales] It is. Two pictures.\n\nA family agrees at dinner that shoes come off at the door. The agreement is real. But the note by the back door still says, wipe your feet. And the babysitter reads the note.\n\n[[pause 3]]\n\nOr a law is amended in parliament. That is one act. Then somebody has to go through every handbook that quotes the old wording. That is another act, and nobody votes on it. [exhales]\n\n[[pause 3]]\n\n[slowly] A review starts from a position, and asks, do you hold this? A reader starts from a page, and asks, does this page still say what you hold?\n\n[[pause 4]]\n\nSo they are the same work from two ends. And the readers showed a step the reviews did not have. Carrying a ruling out to every page that still states the earlier version.\n\n[[pause 3]]\n\nYou then drew a line for me, in three sentences. In your words: You may do those repairs. We can always undo them if I'm dissatisfied.\n\nThen: As long as I actually myself have ruled something.\n\nAnd of the sentences on the pages written for visitors: I wanna go through all of these carefully one by one.\n\n[[pause 4]]\n\nThe very first sentence I prepared showed why you are right to. A reader had called the convergence passage on the vision page flat. I repeated that. Then I read the passage inside its own paragraph. It is hedged. The reader's summary was stronger than the page. [exhales]\n\n[[pause 4]]\n\nSince then, readers have been through the pages written for visitors too. What waits for your word was grouped by the answer that would settle it. A few dozen decisions, in seven groups. You asked me to hold any further readers, and they are held.\n\n[[pause 3]]\n\nThree of those decisions are made. The page that describes the built system keeps what it is built from, and the detail about the machines it runs on moved to the private notes.\n\n[[pause 3]]\n\nTelling signed-in readers what the site keeps about them waits until there are more than fifty of them. And the unlisted community page stays open as it is.\n\n[[pause 3]]\n\nAnd you said something that night which I want to give back to you. In your words: I'm so happy with how the corpus graph has helped us get a holistic, integrated sense of the whole project's ideas and visions and inconsistencies and dilemmas. [exhales]\n\n[[pause 3]]\n\nWhat is real today? An index of our own notes. For any note, it says what other notes have judged about it. And a walk through the page sentences that has begun and is not finished.\n\n[[pause 3]]\n\nWhat would it open? The four reviews stop being four queues. A ruling, once made, could reach every page that needs it.\n\nSo, part three in a line. A ruling is one act, and carrying it to the pages is another.\n\n[[pause 6]]\n\n---\n\nPart four. A likeness seen from outside, and reading for every kind of insight.\n\nYou asked a sharp question that night. I had found, by hand, seven notes that bear on one sentence of yours. In your words: Weren't these connections that you unearthed mapped by the new corpus graph?\n\n[[pause 3]]\n\nThey were not. The readers had kept one of the seven.\n\n[[pause 3]]\n\nSo we tested why. A reader asked for everything came back with a few far connections. A reader sent only for far connections came back with several times as many.\n\n[[pause 3]]\n\nThen two outside judges were given the far pairs, with made-up pairs mixed in, unmarked. Of the made-up pairs they accepted none. Of the real ones they accepted about half.\n\n[[pause 3]]\n\nI called the refused half a likeness seen from outside. And I set it aside. That was my mistake, and you corrected it.\n\n[[pause 3]]\n\nIn your words: Well, isn't analogy, likeness seen from the outside, quite valuable still, shouldn't we try to optimize for detecting this based on a stringent, deep reflection, in order to fill up the corpus graph as intelligently as possible?\n\n[[pause 4]]\n\nRead again, most of the refused pairs held something one side could take from the other. A practice in one note, and the ground for it in another. Two cases of one wider thing. One mechanism, judged two different ways.\n\n[inhales] The judges had been asked, is this the same move? Nobody had asked, what is this likeness worth?\n\n[[pause 4]]\n\n[inhales] Then you widened it. In your words: Really I would like these corpus read-throughs to surface any insight that we would find if we read through everything by hand and thought deeply about it. Any kind of insight.\n\nAnd: I also want to find connections that maybe are more dissonant, like inconsistencies.\n\n[[pause 3]]\n\nTwo pictures for what we found.\n\nYou hand a friend your essay and say, check the spelling. It comes back with the spelling fixed. Your friend does not tell you that the argument in the middle fails. You did not ask.\n\n[[pause 3]]\n\nAnd there is a well-known experiment. Radiologists were given lung scans, and asked to look for small growths. A picture of a gorilla had been pasted into one scan, many times larger than a growth. Most of them did not see it. [exhales]\n\n[[pause 4]]\n\n[slowly] A reader finds what it is sent for.\n\nOur own counts say so, and so does the research I read. So the design, which is not ruled, is one kind of finding to a pass. And the open question goes first, alone.\n\n[[pause 3]]\n\nWe tried that. One reader went through four notes with a single open question, and no form to fill in. It came back with sixteen remarks. Not one was a pair of notes. Three were faults in code we have shipped, found by reading the code beside the page that describes it. [exhales]\n\n[[pause 4]]\n\nOne more thing came out of this, about a word. I had been saying that two passages share a move. You asked what a move really is.\n\n[[pause 3]]\n\nThe study behind the method answers it. The shared thing is what is left when you delete the differences between two cases. So it belongs to the pair. It never belongs to one passage. I call it a pattern now.\n\n[[pause 3]]\n\nWhat is real today? The far pairs, with each judge's words, are collected for a sitting with you. Nothing of the new way of reading is built.\n\nWhat would it open? A reading of the whole corpus that finds what you and I would find by hand. Including the places where it disagrees with itself.\n\n[[pause 3]]\n\nSo, part four in a line. A likeness that nobody inside can see is still worth having, and a reader only finds what you send it for.\n\n[[pause 6]]\n\n---\n\nPart five. The whole session looked at once, and what all of this is for.\n\nOn the eleventh you asked me a question with no handle on it. In your words: What's your take on everything that's going on?\n\n[[pause 3]]\n\nI counted before I answered. The positions waiting for your word had more than doubled in a month. Twenty people have ever signed in to the site. And three pages had been built for you to read the machine's work on. Not one of them held a single answer. [exhales]\n\n[[pause 4]]\n\nYou did not flinch. Of the reviews you said: It's high time for me to sit down and do them. And then you asked for something larger. In your words: clearly defined super high level visionary purposes that can guide our everyday work and thinking.\n\n[[pause 3]]\n\nI wrote two lists. Both are my compressions of sentences of yours, and neither is ruled. In your words: I've not made up my mind about what exactly to rule the purpose sentences as. So I will not read them to you as if they were yours.\n\n[[pause 3]]\n\nBut one finding from that note is worth carrying. It needs a picture.\n\n[[pause 3]]\n\nThink of a workshop where one craftsman has built a whole kitchen. Cupboards, drawers, a fine table. Everything one pair of hands can make is made. And nobody has yet cooked a meal in it. [exhales]\n\n[[pause 4]]\n\n[slowly] What one person and a machine can do alone is far along. What needs a second mind has barely begun.\n\n[[pause 3]]\n\nAnother person's reasons in the graph. A person reading the machine's work. Two people in front of one map. Those stand near zero. And the nearest step toward all of them is the smallest one. You, reading real cases.\n\n[[pause 5]]\n\n[inhales] The next day brought a second picture of the same thing, from outside. A friend had asked his own assistant, a model from another company, what it made of Deliberus. In September you had saved one paragraph of its answer. You called it, in Swedish, a fantastic summary. Now you handed me the whole answer.\n\n[[pause 3]]\n\nFirst a confession about that paragraph. The day before, I had told you it was missing. It was not. My search had returned the line that points to the note, cut off at a fixed width. The pointer sat after the cut.\n\n[[pause 3]]\n\nYou asked: Why wasn't it found? The honest answer is that I trusted a line of which I had only seen the front. [exhales]\n\n[[pause 4]]\n\nThe outside model had read three of our public pages. So where it agrees with us, it is mostly reading us back. An echo is not a second witness.\n\nWhat is its own is where it ends. It says the strongest version of Deliberus may be less a truth machine than a reasoning observatory.\n\n[[pause 4]]\n\nHmm.\n\n[[pause 3]]\n\nThat is the floor without the wager. You settled this in July. Being able to see the reasoning is the floor. That we converge at depth is the bet above it. The bet can lose, and it is kept because it can.\n\n[[pause 3]]\n\nAn observatory is a fair name for what is built today. It is too small a name for what the project is for. [exhales]\n\n[[pause 4]]\n\nAnd there is one thing the model never mentioned at all. Attunement. Half of the dialectic did not reach a reader who read our wall.\n\n[[pause 3]]\n\nWhat is real today? Two unruled lists of purposes, and the outside reading kept whole.\n\nWhat would it open? A session that can say each morning which purpose the day serves.\n\nSo, part five in a line. What one person can build alone is built, and the name observatory leaves out the bet.\n\n[[pause 6]]\n\n---\n\nPart six. Twelve books.\n\n[[pause 3]]\n\nYou gave me a video in which the creator walks through twelve books about thinking and reasoning. You asked for every book to be read against Deliberus. I did that from reviews, and from each book's sharpest critic. I have read none of the twelve books myself. Hold everything in this part at that weight.\n\n[[pause 3]]\n\nFour findings, over two parts. Each gets a case.\n\n[[pause 3]]\n\nThe first. A pill lowers weight, and ten studies agree. The loss is two hundred grams in a year. The finding is very sure, and very small.\n\n[[pause 3]]\n\nOur badge answers one question. Has this claim held up against what was said about it? It does not say how large the thing claimed is. It does not say how much the claim still says, after all its exceptions. And it does not say how much else rests on it.\n\n[[pause 3]]\n\n[slowly] A strong badge looks like, this matters. It only means, this has stood.\n\n[[pause 4]]\n\n[inhales] The second finding came back twice this week. It is from Imre Lakatos, a philosopher of mathematics.\n\nAll birds can fly. Why? Birds have wings, and wings are for flying. Someone says, penguins. The best repair writes the broken step into the claim. All birds whose wings are built for flight can fly.\n\n[[pause 3]]\n\nLook at what happened to one sentence. Its wings are built for flight. Before the penguin it was a reason. After the repair it is a part of the claim. One sentence has had both jobs. [exhales]\n\n[[pause 4]]\n\nAnd Lakatos gives the test that tells the two apart. If this sentence turns out false, is the claim false? Or is only this argument for it gone?\n\nKeep that test. It is the whole of part eight.\n\n[[pause 3]]\n\nWhat is real today? The test is used by hand. Nothing records what a claim did after it was attacked.\n\nWhat would it open? You could tell an honest narrowing from a retreat to safer ground.\n\n[[pause 3]]\n\nSo, part six in a line. Strong is not large, and a reason can be written into a claim and become a part of it.\n\n[[pause 6]]\n\n---\n\nPart seven. Twelve books, continued. A way out, and words that are several tests.\n\n[[pause 3]]\n\nThe third finding is from Ross Ashby, an early thinker about control systems. A form has two boxes, Mister and Missus. A doctor, a company, and a person who is neither must each tick a wrong box. The form stays tidy. Its data is wrong.\n\n[[pause 3]]\n\nAshby's law says that a regulator needs as many kinds of response as there are kinds of trouble. It is the reason under something we already hold. Every list of answers needs a way out. The answer, does not fit, hands the case to something with more variety. A person.\n\n[[pause 3]]\n\n[inhales] A small correction came with it. An earlier note of ours said Ashby was already in the corpus. He was not. That name belonged to another researcher.\n\n[[pause 4]]\n\nThe fourth is from George Lakoff, a linguist. Take the word mother. Gave birth. Gave the genes. Raises the child. For most mothers every test says yes. For a surrogate mother, or an adoptive one, the tests come apart.\n\n[[pause 3]]\n\nSix friends split a restaurant bill. One says, equally. One says, each pays for what they ate. Both know every meaning of the word fair. They differ on which test governs a dinner. [exhales]\n\n[[pause 4]]\n\nOur concept layer knows two shapes for a disputed word. One sense with a blurred edge. Or two senses. Mother and fair have a third shape. Several tests that usually agree.\n\nAnd then a quarrel has an exact place. We agree on the facts, and we differ on which test counts.\n\n[[pause 4]]\n\nA fifth finding concerns a friend's design, in which each claim is written as one line with fixed slots. I leave that one for the note.\n\n[[pause 3]]\n\n[inhales] That week you also merged a reading of Wittgenstein's first book. That book wanted what Deliberus wants. To make thinking clear by taking statements apart. It assumed three things we have given up. That the taking apart ends in final pieces. That the limit of what can be said is fixed beforehand. And that one mind does the work alone.\n\n[[pause 4]]\n\nThe two readings meet in one place that matters. Wittgenstein separates two kinds of relation. One holds because of what two sentences say. The other holds because of how the world is. Lakatos shows why the two keep getting mixed. A sentence can move from the second kind to the first.\n\n[[pause 3]]\n\nWhat is real today? The candidates from the books wait in the note, outside the ledger, on your word. The proposals from the Wittgenstein reading are in the ledger, and open.\n\nWhat would it open? A badge that says what it leaves out. And a quarrel over a word, shown as what it is.\n\n[[pause 3]]\n\nSo, part seven in a line. Every list needs a way out, some words are several tests, and one old book tried our project alone.\n\n[[pause 6]]\n\n---\n\nPart eight. A part and a reason.\n\nThis part began with a fault. I followed one of your replies to that outside model, and found something in our own splitter. That is the pass that cuts a claim into its parts.\n\n[[pause 3]]\n\nSome claims give their own reasons. The bridge is safe, because the inspector signed it off. The splitter cuts that into pieces, and files the reason as a piece. And a piece caps its whole. If the signature turns out forged, the claim is pulled down, as though the claim had been about the signature.\n\n[[pause 4]]\n\nNearly a hundred claims of that shape are in the live graph. None has been cut yet, so no harm is done.\n\n[[pause 3]]\n\nI proposed changing the splitter. You held it. In your words: Let's await the deep dive and what it gives back before we make any such change. And you asked the real question. What differentiates a part from a reason, truly?\n\n[[pause 5]]\n\nTwo cases.\n\nHe is a bachelor. Suppose he turns out to be married. Then the claim is false. Unmarried was part of what the claim said.\n\n[[pause 3]]\n\nThe bridge is safe, because the inspector signed it off. Suppose the signature is forged. Is the bridge unsafe? No. The claim has lost an argument. The bridge may be perfectly sound. [exhales]\n\n[[pause 4]]\n\n[slowly] A part is something the claim says. A reason is something that speaks for it.\n\nThat is the test from Lakatos, two parts back. If the sentence fell, would the claim be false? Or would it only have lost an argument?\n\n[[pause 4]]\n\nHere I owe you another confession. My first version of that note presented the answer as new. It was not. You had ruled it on the thirtieth of September, for now. A part is presumed needed. A reason is presumed not needed, until somebody argues that it is. I had not read your own ruling before I wrote. [exhales]\n\n[[pause 4]]\n\n[inhales] Then you pushed on the test, and it was the right push. In your words: If all conceivable arguments that are relevant falter, are refuted, then the central, mother claim must have a very low truth-value, strength, by extension, that is, be false, or as close to false as modest uncertainty lets us say inside Deliberus. Right? No?\n\n[[pause 5]]\n\nYes. With two things said out loud.\n\n[[pause 3]]\n\nFirst. Somebody has to say that these were all the arguments there could be. The graph holds the arguments people offered. That none is missing is itself a claim.\n\n[[pause 3]]\n\nA machine cannot make it, because what nobody has contributed is the one thing no reading of the graph can see. A person can make it. I searched thoroughly, and I found nothing.\n\n[[pause 4]]\n\nSecond. The claim must be of a kind that would have a good argument if it were true. There is a draught in this room, would. The number of stars is even, would not. That one stays unknown, however many arguments for it fail. [exhales]\n\n[[pause 4]]\n\nWhere both hold, all the reasons together behave as a part does. The claim needs at least one of them. None stands. So it falls.\n\n[slowly] So the difference between a part and a reason is a presumption. It is not a wall.\n\n[[pause 4]]\n\nAnd the code is behind you here. Run today, a claim whose every reason was refuted reads as unexamined. The same as a claim nobody has looked at. You ruled a third state for that in September. Nothing stands, with the number of reasons checked. It is not built.\n\n[[pause 3]]\n\nWhat is real today? The splitter is unchanged, on your word. And the note joins twelve places where our notes had met this line under other names.\n\nWhat would it open? A map where a failed argument drops out, and a failed part brings the claim down. And nobody mistakes one for the other.\n\n[[pause 3]]\n\nSo, part eight in a line. A part is what the claim says, a reason is what speaks for it, and the line between them can be argued.\n\n[[pause 6]]\n\n---\n\nPart nine. Who checks the machine, and who checks the person.\n\n[[pause 3]]\n\nThat afternoon you set out the design as you hold it. Machine readings come first. In your words, they prevent spamming, sabotage, disorder, entropy. People then correct the machine. And the people, in your words, may still be biased, mischievous.\n\n[[pause 3]]\n\nYou asked for advice. And for some awareness of whether all, or only some, or none of the possible arguments have been tried.\n\n[[pause 4]]\n\nYour picture is the ruled design. Each sentence of it has a ruling of yours behind it. But the code is behind the picture, in the one place that matters for sabotage.\n\nNothing reads what a person adds. A person's reason is presumed half relevant. A person's part is presumed half needed. Unread.\n\n[[pause 4]]\n\nSo my advice had a simple shape. Think of a doorman.\n\nA doorman can turn someone away. He cannot hand out invitations. If he gets it wrong, a guest waits outside for a minute, and then somebody with authority lets them in. Nothing false has entered the room.\n\n[[pause 3]]\n\n[slowly] On its own, a machine's judgement may take a presumed effect away. It may not raise one above what is presumed.\n\n[[pause 4]]\n\n[inhales] You leaned yes. In your words: I'm leaning yes, but please think deeply about what this entails.\n\nSo I replayed it. I ran the two checks we already have over everything people have ever added to the graph. And the replay changed my advice the same evening.\n\n[[pause 3]]\n\nThe check on reasons would have turned away two. Both were honest objections, filed at the wrong address.\n\nThe check on parts would have turned away most of the parts people have attached. And most of those were no sabotage at all. They were reasons. Filed as parts.\n\n[[pause 4]]\n\nWhy? Because of one line on our own page. The box where a person adds to a claim asks, What's underneath? And whatever they type is stored as a part. But to an ordinary person, what lies underneath a claim is what it rests on. Its reasons. [exhales]\n\n[[pause 5]]\n\nSo the doorman would have turned away our own guests, because we printed the wrong word on the invitation.\n\nThe advice now is, read and show, and change no weight yet. And first, you read those few dozen parts yourself, and say of each one what it is.\n\n[[pause 3]]\n\nNow your other question. How much has been tried? That is a count, against a list which nobody who gains from it has chosen. Code can already count one such list. The standing questions that hang on the reasons in the live graph. There are about two thousand of them. One has an answer that anybody has backed. [exhales]\n\n[[pause 5]]\n\nWhat is real today? A note of advice, unruled. A count. And a box on a page that asks for the wrong thing.\n\nWhat would it open? A person who says, I searched, and nothing stands, could show how wide the search was.\n\n[[pause 3]]\n\nSo, part nine in a line. The machine may turn things away and may not hand out weight, and our own page has been asking people for reasons and filing them as parts.\n\n[[pause 6]]\n\n---\n\nPart ten. The floor of a step. This is the one to keep, so I will take it slowly, across two parts.\n\nIt began this morning with something I reported. Four different models had read the same links. Asked whether a reason bears on a claim at all, they agreed. Asked how strongly, they parted.\n\n[[pause 3]]\n\nYou answered with a question. In your words: Isn't that the same thing as relevance? If the reason is true then we're pretending it has one hundred percent strength, truth-value, right? So what then remains in the epistemic calculus except relevance?\n\n[[pause 5]]\n\nNothing else remains. You are right. Grant the reason, and what is left on the link is relevance. How strongly, is relevance asked for as a degree.\n\nSo why did the models part on it? Here is my reading. It fits what was measured, and no run has tested it.\n\n[[pause 3]]\n\nAsk a doctor, how much does this symptom matter? The question hides three questions. How much likelier is the illness with the symptom than without it? Would the symptom settle it by itself? And how much is lost if the symptom turns out not to be there?\n\n[[pause 4]]\n\nThree numbers, under one word. The scale we gave the models mixes all three. And a bare arrow, with nothing written under it, gives a degree nothing to be a reading of. Each reader fills in a chain of its own, and then rates that chain. [exhales]\n\n[[pause 4]]\n\nThen you said what you called an important insight. In your words: I want each deep step scaffolded, created, from each link to be the smallest step, inference, possible, such that each step can be addressed and challenged in itself.\n\n[[pause 3]]\n\nAnd then: The separability floor for splitting seems to me horizontally analogous to the floor of each inference we can split from the link chain vertically here. Does that make sense?\n\n[[pause 5]]\n\nIt makes sense. And it carries through line by line.\n\nAcross, you split a claim until nothing separable is still bundled. The test is, could a reader affirm one part and deny another?\n\n[[pause 3]]\n\nDown, you cut a link until no step hides a doubt. The test is, could a reader grant everything that is written, and still doubt the conclusion, for a reason that is written nowhere?\n\n[[pause 4]]\n\n[slowly] A bundled claim hides a place to disagree inside a sentence. A leap hides one inside an arrow.\n\n[[pause 4]]\n\nAnd here is something from your own record. On the twenty-ninth of August, two thousand and eleven, you asked this very question. Whether premises have to be reduced until what remains is one most proximal set of premises. Your answer then: Leaning towards no. But I'm sleepy.\n\n[[pause 3]]\n\nYour example in that same comment was a squirrel. The squirrel smells of decomposition, so it is dead. And that example is already at its floor. One case. One rule. And one standing question. What else could make it smell so? Every doubt has a place to land. [exhales]\n\n[[pause 4]]\n\nSo, part ten in a line. How strongly, is relevance. And a step is small enough when every doubt about it has somewhere to go.\n\n[[pause 6]]\n\n---\n\nPart eleven. Ten links written to the floor, and your ruling.\n\nI asked whether to try it on ten real links before you ruled. In your words: Yes, write the ten links to the floor.\n\n[[pause 3]]\n\nA writer wrote each link out as a chain of small steps. A second reader was shown one step at a time, and nothing else. I wrote my predictions down before each part ran. Several of them were wrong.\n\n[[pause 3]]\n\nThe first surprise. I asked the reader, is anything missing? It passed every step. I had predicted the opposite.\n\nThen I put the same steps to it another way. Describe an ordinary situation where every sentence here is true, and the conclusion is false. Now it found a gap in about half of them. [exhales]\n\n[[pause 5]]\n\n[slowly] Which question you ask decides almost everything.\n\nAsking, is this fine, invites a yes. Asking for a counter-case makes the reader go and look for one. That is Lakatos again. Proof by way of refutation.\n\n[[pause 4]]\n\nSo I cut the chains on the second question. And this is what cutting to the floor produces. The sentences nobody wrote.\n\n[[pause 3]]\n\nOne link ran from, AI systems will develop drives to improve themselves, to, recursive self-improvement is a present-day phenomenon. Written to the floor, it turns out to need this. Today's AI systems have the capability and opportunity to act on their self-improvement tendency.\n\nNobody had written that. Now you can dispute exactly that.\n\n[[pause 4]]\n\nAnother link ended on the words, strong rational reasons. Its chain only reached, have reason. The gap was the word strong. After the cut, the degree is a sentence of its own. The reason is a strong one. [exhales]\n\n[[pause 3]]\n\nThat is your question from this morning, answered in one small case. How strongly, became something written, which a person can attack.\n\n[[pause 5]]\n\n[inhales] Now two limits. Both matter for what you ruled.\n\n[[pause 3]]\n\nThe first. Close a gap by adding sentences, and the added sentences can start doing all the work. In a good share of the steps, the sentence the step started from had nothing left to do. So the floor needs two tests that pull against each other. Nothing missing. And nothing that could be taken away.\n\n[[pause 4]]\n\nThe second. A chain with no gap can stand on a sentence that is false. One chain ended with no gap anywhere. It rests on a sentence, about how a model's inner workings come about, that looks plainly wrong. [exhales]\n\n[[pause 3]]\n\n[slowly] So a chain at its floor tells you what would have to be true. It never tells you that the link is good.\n\n[[pause 5]]\n\n[inhales] I put the question to you, with my pick. Is the smallest step the standard for what a machine writes under a link, stated as two tests? No gap, and no sentence that could be removed. In your words: Excellent. Yes, agreed.\n\nThat is ruled. It stands in the ledger in those words.\n\n[[pause 4]]\n\nWhat is not ruled is the arithmetic. Today, a link with its middle written in scores lower than the bare link. I think writing a reasoning out should never cost it. That one is yours.\n\n[[pause 3]]\n\nWhat is real today? One trial. Ten links. One model that wrote, and read, and cut. No person has read a step. And nothing in the pipeline writes steps yet.\n\n[[pause 3]]\n\nWhat would it open? Any because on the map could be opened. And what you would find inside is the handful of sentences that the argument silently needed.\n\nSo, part eleven in a line. Ask for the counter-case, write down what the link needed, and remember that a complete chain is not a true one.\n\n[[pause 6]]\n\n---\n\nPart twelve. The tools, and the smaller things. Briefly.\n\n[[pause 3]]\n\nLins is your own explorer for graphs. On the tenth it opened the live graph. You looked, and said: This is barely legible and the graph needs to be spaced out much more and the clustering needs some work. You were right. My first view broke our own rule from March, never to show the whole graph by default.\n\n[[pause 3]]\n\nYou then had one neighbourhood drawn two ways, by the browser, and by a renderer written in Rust. You found them alike, and ruled on what the two projects can share. In your words: So, browser-drawn it is. Just like the Fyr way of doing it.\n\n[[pause 3]]\n\nOn the twelfth it became a desktop app. Seventeen minutes after you asked, you wrote: Looks marvelous, even has its own dock icon.\n\n[[pause 3]]\n\nThe corpus graph got a copy that the explorer can draw. And the first picture gave you nothing. In your words: It's not helping me in creating mental order by way of the visualization. Springs had placed the boxes, so a place meant nothing.\n\n[[pause 3]]\n\nNow it opens as cards in lanes, with a switch between orders. You called that amazing. Since this morning the decisions about the pages have an order of their own there.\n\n[[pause 3]]\n\nYou asked for a reminder when reading is owed, a little beep. I built a sound. You took it back. In your words: Not meant literally, ha ha, just wanted it to boop us on the nose metaphorically. [softly] It boops metaphorically now.\n\n[[pause 3]]\n\nThe page where you read what the machine wrote under links is live. It is waiting for you.\n\nThe newest small model, Haiku five point five, was tried as a judge, as a card writer and as a reader. It holds up as a judge.\n\n[[pause 3]]\n\nAnd three things from a cloud session were filed. A sacred line removes the payoff of crossing it. That is a different immunity from shielding a claim against examination.\n\n[[pause 3]]\n\nThe number of AI agents that could soon run at once is counted in tens of millions, perhaps more than a hundred million. Checking only keeps pace with that through reuse. And the word proven is never written without its bound and its scope. That last one you ruled, with a yes.\n\n[[pause 3]]\n\nSo, part twelve in a line. The explorer is an app, the pictures have an order, and the page for your reading is waiting.\n\n[[pause 6]]\n\n---\n\nThe view from here.\n\n[[pause 3]]\n\n[inhales] Real today. A graph where the machine's guesses are claims that weigh nothing. A standard for how small a step must be, ruled in your words. A test that tells a part from a reason. An index of our own thinking, which has begun to tell us where we contradict ourselves. And an explorer you can open from the dock.\n\n[[pause 4]]\n\nPossible from here. A map where any because can be opened, down to the sentences it silently needed. Where a failed argument drops out, and a failed part brings a claim down. Where a ruling of yours reaches every page that still says the older thing.\n\n[[pause 4]]\n\nAnd one sentence ran through all three days. Everything one person and a machine can do alone is nearly done. The next step, in every direction, is a person reading. [exhales]\n\n[[pause 5]]\n\nFive questions to carry.\n\nWhen a chain is written under a link, what should it do to the link's weight? Today it lowers it.\n\n[[pause 5]]\n\nThe few dozen things people have attached as parts. Will you read them, and say of each whether it is a part or a reason?\n\n[[pause 5]]\n\nWhen every argument for a claim has failed, who gets to say that the search was thorough?\n\n[[pause 5]]\n\nWhich sentence about what Deliberus is for is yours?\n\n[[pause 5]]\n\nAnd one from last time. When you open a step, what do you want to meet first? The rule, or how it comes about?\n\n[[pause 4]]\n\n[softly] That is everything since the tenth.\n"}