{"path":"listening/2026-10-14-parts-reasons-and-what-the-number-means.md","content":"# Parts, reasons, and what the number means: a listening walkthrough (2026-10-14)\n\n- **What it is**: the seventeenth listening piece, on one outside note rather than on days of work:\n  [`claim-parts-and-inferential-reasons-2026-10-14.md`](../research/claim-parts-and-inferential-reasons-2026-10-14.md),\n  written by Codex on the founder's side that morning (he called it Astra's), after two session\n  answers to his question about parts, reasons, chain depth and the completeness check. His\n  request: *\"Read. Ponder it deeply and what it means for us. Then make an audio walkthrough. It can\n  be an hour long.\"*\n- **Rendered**: 5,887 words, 50 minutes 17 seconds, twelve parts and a closing view; 160 marked pauses,\n  14 breaths, 8 slowed sentences. Every factual line was checked against the note it names before\n  the render; the four corrections it caused are filed (the weakest-link note § 1f, the chain note\n  under § 11.8, the April ontology note under its August update, the part-and-reason note § 4) and\n  the session record has the account (`session49-what-goes-under-a-link.md` § 7n).\n- **Covers**: where the question stood · three operations on three objects · part and reason as\n  roles · what the false-without-it test measures (need, not content) · where the confusion entered\n  (the Dr Smith example of April) · depth, and the invariance rule in two halves · one reason and\n  fourteen · three senses of completeness · three ways to attack a link · the Toulouse method\n  (principles before operators) · what the number on a claim means, with the note's eight test\n  cases · the view from here and four questions, the first of which comes before the others.\n- **Status**: a snapshot of 14 October. Nothing in it is ruled; the four questions at its close\n  are session49 § 7m rows 1a to 1c and row 1. How these pieces are written:\n  [how-to-write-for-listening.md](how-to-write-for-listening.md).\n- **Where the spoken text is rounder than the notes**, so nobody reads more into the audio than\n  the notes hold.\n  1. *Four of my claims too strong*: the note's audit table (§ 7) has eight rows; the four named in\n     part twelve are the session's own list (a reason passes in whole; min the only invariant rule;\n     fourteen from one source; the bicycle's brakes), with the bicycle standing where the table has\n     *AG14's tests prove completeness*.\n  2. *A bare arrow reads a touch better than the same arrow with its middle written in*: 0.519\n     against 0.500, the squirrel run of 2026-10-08, reproduced by the note's author.\n  3. *Two things each at nine in ten … eighty-one in a hundred*: the one number spoken as the idea.\n  4. *Since the twenty-eighth of September a link can carry a relevance claim … empty until the\n     tenth*: the slot since 28 September (ledger AD22); no link held one on 8 October; 29 relevance\n     claims from the trials of 10 October.\n  5. *Stored the copper-box shape backwards for six months … repaired on the seventh*: 261 part-of\n     links from April to 2026-10-07.\n  6. *A discount for shared evidence is built and switched off*: `evidence_routing.py`, built and\n     tested and not adopted.\n  7. *Our test for a part, since the eleventh*; *our note of the twelfth*: the part-and-reason note,\n     dated 2026-10-12, which says the notes have held the test since 11 October.\n  8. *The ruling of yesterday*: AG14, 2026-10-13.\n  9. *Three papers from Toulouse*: Amgoud and David 2018; Amgoud, Ben-Naim, Doder and Vesic 2017;\n     Amgoud, Doder and Lagasquie-Schiex 2026, read by the note's author in abstract and selected text.\n  10. *His April sentence* is spoken with its parentheses dropped and *QEM/QBAF* spelled letter by\n      letter with a pause for the slash; *his 30 September words* with *probability/Bayes* as\n      *probability, Bayes* and the trailing dots dropped.\n  11. *Our strength calculus was chosen in March and has been patched since*: the March spec; the\n      repairs of 17 and 22 to 23 August, 17 and 30 September.\n  12. *The corpus has said since August that strength is standing and not probability*: the\n      weakest-link note § 1f (31 August), ledger P5 superseded, and the participation-dependence\n      line in the Bengio note.\n  13. *A code reading on the eighth found that an undercutter has no arrow type of its own*: a\n      helper's report of 2026-10-08, as a reader returned it on the 10th.\n  14. *The thorough search … unbuilt*: the four pieces of ledger Q34, all unbuilt.\n  15. *The pictures*: the cold damp room and its thermometer, the sealed copper box, the kitchen\n      light and Anna, the machine and its fuel are the note's and the chat's; the bicycle was the\n      session's and is owned as muddled; the recipe, the suitcase, the detective, the weather\n      forecast, the doctor's test, the baking contest and the bridge with four questions are\n      Claude's own and are in no note.\n  16. *His quoted words* are spoken with a slash as a pause, abbreviations spelled, and quote marks\n      inside a quote dropped.\n\n---\n\n[warmly] This is a walkthrough of one note, about an hour long. The note is the one you called Astra's. Its own header says it was written by Codex, this morning, after my last two answers to you about parts and reasons. The silences are deliberate. They are room to think.\n\n[[pause 3]]\n\nHere is the map. Twelve parts. Where we were, briefly. Three operations on three different things. Why part and reason are roles and not kinds. What the false-without-it test really measures. Where the confusion entered our own history, in April.\n\n[[pause 3]]\n\nThen depth, and the rule you were reaching for, stated right. One reason, and fourteen. Three different things that the word completeness can mean. Three ways to attack a link. What the Toulouse school adds. What a number on a claim is supposed to mean. And the view from here, with the questions to carry.\n\n[[pause 3]]\n\nOne thing to say before we start. The note audits my last answer line by line and finds four of my claims too strong. It also finds one plain error in a note of ours from August. I will say each one where it belongs, in the first person. [exhales]\n\n[[pause 4]]\n\n---\n\nPart one. Where we were.\n\n[[pause 3]]\n\nYou wrote that you were confused. There is a difference between parts and reasons, you said, right? And a reason that is true and relevant passes its strength on to the claim it argues for, whether there is one such reason or fourteen?\n\n[[pause 3]]\n\nAnd splitting a chain into more steps should not change what reaches the top, if everything true and relevant is passed along? And what was that completeness check again?\n\n[[pause 4]]\n\nI gave you two answers. The first was short. The second was grounded in eight of our notes and ten ledger rows. Both said: yes to the difference, yes to the depth intuition, and the way forward is the weakest rule with full passage, gated on completeness.\n\n[[pause 3]]\n\nThen the note arrived. Its finding, in one sentence, is this. [slowly] Your confusion was not yours. It was the corpus's, and it is older than this conversation.\n\n[[pause 4]]\n\nHere is the shape of what it found. The project has three different operations, and it has been moving between them without building the bridges. Splitting what a claim says. Exposing why anyone should believe it. And weighing everything that bears on it.\n\n[[pause 3]]\n\nEach is a different question about a different object, and each has its own arithmetic. My answers to you crossed from one to the next without saying so.\n\n[[pause 4]]\n\nSo this hour is less about parts and reasons than about keeping three things apart. And the note is kind about one thing. It says our October notes had already named the three wholes, and that it would be unfair to say I never noticed the distinction.\n\n[[pause 3]]\n\n[inhales] The trouble is that the arithmetic I recommended did not respect it.\n\n[[pause 3]]\n\nWhat is real today: your question, the note on disk, and our own notes, which hold most of the answer in pieces. What this opens: a vocabulary clean enough to rule on. Part one in a line. The confusion has a history, and we are about to walk it.\n\n[[pause 6]]\n\n---\n\nPart two. Three operations, three objects.\n\n[[pause 3]]\n\nTake a plain sentence. The room is cold and damp at noon. Now ask three questions about it.\n\n[[pause 3]]\n\nFirst. What does it say? It says two things about one room at one time. That it is cold, and that it is damp. You can agree with one and deny the other. That is splitting content. It stays inside the sentence.\n\n[[pause 3]]\n\nSecond. Why believe it? Because the thermometer by the window reads twelve degrees. That reading is not a piece of what the sentence says. It is a different claim, used to establish the first. That is exposing an argument. It goes outside the sentence and brings something in.\n\n[[pause 4]]\n\nThird. What survives, once everything is on the table? The thermometer reading, but also a note that the thermometer was last calibrated years ago, and a second thermometer, and a disagreement about what counts as cold. That is weighing the case. It takes in every argument and every objection at once.\n\n[[pause 4]]\n\nHere is the same thing in a kitchen. Anna is home and asleep. What it says: two things. Why believe it: her car is in the drive and the lights are off. What survives: well, the car might be her sister's, and she sometimes leaves the lights off when she reads. Three questions, three different objects. [exhales]\n\n[[pause 4]]\n\nOur own notes reached this on the twelfth, two days ago, and called them three wholes. What the claim says. One argument for the claim. The whole case for the claim. The note says the same from the other side: three operations. Split content, expose an argument, assess the case. The nouns and the verbs match.\n\n[[pause 4]]\n\nThe point for us is where each lives in the code. The first operation has a home. The splitter cuts a claim into parts, and the parts cap their whole. The third has a home. The badge on a claim page adds up every link into it.\n\n[[pause 3]]\n\nThe second, the argument, had no home until the tenth of October. Since then a sentence the machine writes under a link is a claim, held under that link's relevance claim, and it moves no strength.\n\n[[pause 4]]\n\nSo what is real today is that the middle operation exists as a place and not yet as arithmetic. What this would open is a ruling that can say, for each of the three, what the number does. Part two in a line. [slowly] Three questions, three objects, and the arithmetic should know which one it is in.\n\n[[pause 6]]\n\n---\n\nPart three. Part and reason are roles, not kinds.\n\n[[pause 3]]\n\nHere is the note's own example, and I think it is the best one in it. A sealed box, and a label on it that says every item inside is copper. And a slip that says item number four is in this box. From those two you can conclude: item four is copper.\n\n[[pause 4]]\n\nNow, those two sentences are parts of the argument. They work together. Take either away and the argument is gone. But they are not parts of the conclusion. Item four is copper says nothing about labels or boxes.\n\n[[pause 3]]\n\nAnd here is the turn. Suppose the label turns out to be wrong, because one item is steel. The argument collapses. Does item four stop being copper? No. It might be copper anyway.\n\n[[pause 3]]\n\n[inhales] A failed premise kills a route. It does not kill the destination.\n\n[[pause 5]]\n\nSecond example, the kitchen again. The car in the drive is my reason to think Anna is home. The car turns out to be her sister's. My reason is gone. Anna is still home, as it happens. The reason was a part of my argument, and never a part of the fact.\n\n[[pause 4]]\n\nSo the note says: part and reason describe relations and roles. They are not two species of sentence. One sentence can be a part of an argument and a reason for a conclusion at the same time. Our note of the twelfth said the same, and the audit confirms it. It adds one demand.\n\n[[pause 3]]\n\nAlways name the whole. A part of what? Of the sentence, of the argument, or of the case?\n\n[[pause 4]]\n\nAnd there are two different ways to sort, and they cross. One asks: is this sentence a piece of what the claim says, or a ground for believing it? The other asks: do these premises work together on one route, or do they give separate routes? Our notes call the second question linked against convergent.\n\n[[pause 3]]\n\nBoth halves of that question are about arguments. Neither half makes a premise into a piece of the conclusion. [exhales]\n\n[[pause 4]]\n\nI will own one thing here. I told you that fourteen reasons from one source count as one. Too wide. Fourteen copies of one observation are one. Fourteen different readings from one thermometer are fourteen, and they share a weakness. Dependence is not identity. We come back to that in part seven.\n\n[[pause 4]]\n\nWhat is real today: the linking step stored the copper-box shape backwards for six months, premise as a part of conclusion, and that was repaired on the seventh. What this opens: the splitter's own gate, which asks whether a claim would be weaker without a sentence, and which a reason passes. Part three in a line.\n\n[[pause 3]]\n\nA premise can fail and leave its conclusion standing, so it was never a piece of it.\n\n[[pause 6]]\n\n---\n\nPart four. What the false-without-it test really measures.\n\n[[pause 3]]\n\nOur test for a part, since the eleventh, goes like this. If this sentence fell, would the claim be false, or would it only have lost an argument? False, a part. Lost an argument, a reason. I used it on you with a bicycle. The bike is roadworthy. Brakes, tyres, lights. No brakes, not roadworthy.\n\n[[pause 3]]\n\nSo brakes are a part.\n\n[[pause 4]]\n\nThe note takes that apart, carefully. The test finds something real. It finds what the claim needs. But needing is not saying.\n\n[[pause 3]]\n\n[inhales] Here is why.\n\n[[pause 3]]\n\nAny sentence needs, in a strict sense, the sentence or something else. If it is raining, then it is raining or Tuesday. Take away the second sentence and the first falls. So by the test, raining or Tuesday is a part of it is raining. Nobody believes that. The test over-collects.\n\n[[pause 4]]\n\nA plainer example. The machine is running. Under the engineer's model, it cannot run without fuel. So no fuel, not running. Is fuel a part of what the sentence says? No. The sentence says the machine is running.\n\n[[pause 3]]\n\nThe fuel is a condition it needs in the world, and a reason the engineer has, and neither is in the words.\n\n[[pause 4]]\n\nSo back to my bike. Brakes, tyres and lights could be physical components, or practical requirements, or evidence of fitness, or lines on a checklist that defines the word roadworthy. If someone has stipulated the checklist, the lines are parts relative to that stipulation.\n\n[[pause 3]]\n\nOtherwise I was naming parts of a bicycle and calling them parts of a sentence about it. And when I split brakes into cable and pads, I quietly assumed a brake design and turned two nouns into claims. That example was mine, and it was muddled. [exhales]\n\n[[pause 5]]\n\nThere is a name for what the test misses. A philosopher called Stephen Yablo works on aboutness: what a sentence is about, as distinct from what it implies. His point is that implication is too coarse to tell you which content belongs to which subject. The note draws on him for this.\n\n[[pause 3]]\n\nSo the question what does the claim say has to stay beside the question what does the claim need, and neither replaces the other.\n\n[[pause 4]]\n\nNow what this means for our arithmetic, and it is good news in one half. The cap, the rule that a part holds its whole down by how strongly the whole needs it, runs on need. And need is exactly what the false-without-it test measures.\n\n[[pause 3]]\n\nOn the twelfth we checked that the cap's formula is the exact ceiling in probability's terms. So the cap is sound on its own terms, and its dial is the right dial.\n\n[[pause 4]]\n\nThe other half is the lift. The lift is the old open question: when every part of a claim is well established, may the claim rise to meet them? The lift needs the parts to be all of what the claim says. That is a question about content, and the need test cannot answer it.\n\n[[pause 3]]\n\nSo the lift waits on a different check than the cap, and I had been treating them as one.\n\n[[pause 4]]\n\nWhat is real today: the cap is live and its measure is the right one. What this opens: the lift's gate has to be a content check and not a need check. Part four in a line. [slowly] Needing is not saying, and only the cap runs on needing.\n\n[[pause 6]]\n\n---\n\nPart five. Where the confusion entered.\n\n[[pause 3]]\n\nThe note traces the mix-up to one example, and the example sits in our always-loaded file, under a heading marked decided. It is from the first of April. Doctor Smith is credible as an expert in labour economics. And under it, four pieces. She has published forty-seven papers in the field.\n\n[[pause 3]]\n\nA university appointed her to a chair. A publication record in a domain indicates expertise. And labour economics is the relevant domain.\n\n[[pause 4]]\n\nLook at those with part three in mind. The papers and the chair are grounds for calling her an expert. They are why you would believe it. They are not what the sentence says. Someone can be an expert without that chair, and without that many papers.\n\n[[pause 3]]\n\nThe note's words: the later August annotation calls all four jointly required, and that is too strong even as an account of this argument.\n\n[[pause 5]]\n\nTwo everyday cousins of the same slip. The cake is good. Under it, as parts: fresh eggs, real butter, a hot oven. Those are reasons it came out well, not what good means. And the bridge is safe. Under it: the inspection passed, the steel is new. Reasons again.\n\n[[pause 3]]\n\nIn each case the sentence says one thing, and the list under it says why.\n\n[[pause 4]]\n\nNow the fair part, which the note is careful about. Your own words from April were wider and better than the diagram.\n\n[[pause 3]]\n\nIn your words: the mother claim, the more general one, should simply be modeled as being supported by, and receive Q-E-M, Q-B-A-F support transitively, programmatically, all the subclaims that it can and should be decomposed into.\n\n[[pause 3]]\n\n[inhales] Supported by. Your sentence had the arrows running as support. The worked example under it drew them as parts, and the August annotation hardened that into jointly required. The example was the assistant's. The hardening was mine.\n\n[[pause 5]]\n\nWhat it means for us is uncomfortable and clean. The example that teaches every new session what decomposition is, is a reasons-as-parts example. A session that learns from it will file grounds as parts, and the splitter and the arithmetic will follow. That is a root, not a leaf.\n\n[[pause 4]]\n\nWhat is real today: the example stands in the always-loaded file, marked decided, and nothing in it has been changed, because that file changes only on your word. What this opens: a decision for you, which I will state in the last part. Part five in a line.\n\n[[pause 3]]\n\nThe confusion has a birthplace, and it is our canonical example.\n\n[[pause 6]]\n\n---\n\nPart six. Depth, and the rule you were reaching for.\n\n[[pause 3]]\n\nYou said: splitting a chain into more steps should make no difference to what reaches the top, if everything true and relevant is passed along. I said yes, and I said the weakest rule is the only rule that gives you that. Half of that was right.\n\n[[pause 4]]\n\nHere is the rule stated the way the note states it, in two halves that must be kept apart. [slowly] Making existing reasoning explicit should not, by itself, change the verdict. Revealing a new substantive uncertainty may. [exhales]\n\n[[pause 5]]\n\nTwo examples. A recipe says: bake for forty minutes. You rewrite it as: preheat the oven, put the cake in, wait forty minutes. Nothing about whether the cake bakes has changed. You only wrote down what was always there. Now you discover the oven's thermostat is broken. That is new.\n\n[[pause 3]]\n\nIt may change whether the cake bakes. The first move must not shift the number. The second may.\n\n[[pause 5]]\n\nSecond, from our own graph on the eighth. A bare arrow from one claim to another reads a touch better than the same arrow with its middle claim written in between. Nothing was learned. The number moved down. That is the first half failing. Our calculus charges you for writing down what was already there.\n\n[[pause 4]]\n\nNow the part I had wrong. I said the weakest rule is the only rule where splitting changes nothing. The note says: true under one particular algebraic requirement, false as a general claim about representation. Probability, done honestly, is also unmoved by redescription. The chance of A and A is the chance of A.\n\n[[pause 3]]\n\nDrawing an extra box between two claims does not multiply anything by nine tenths. Multiplying is what you do when you have decided two things are independent, and that is a decision about the world, not about the drawing.\n\n[[pause 5]]\n\nSo the invariance you want is not a property of one operator. It is a demand on any operator we choose: that redescription is free and new information is not. The weakest rule satisfies it for jointly needed things. So do others, if used faithfully.\n\n[[pause 3]]\n\nWhat makes the weakest rule special is narrower: among the ways of combining jointly needed grades, it is the only one where a thing combined with itself is itself. That is what makes padding free, and it is still true. It is just not the same claim.\n\n[[pause 5]]\n\nAnd now the plain error, which I want to own because it is in a note of ours and I stood on it. Our August note on the weakest rule says that taking the weakest part is the one estimate of a conjunction's chance that never overstates it, under any dependence.\n\n[[pause 3]]\n\n[inhales] It is the other way round. The weakest part is the ceiling. The chance that two things both hold can never be more than the chance of the less likely one. It can be less.\n\n[[pause 3]]\n\nTwo things each at nine in ten: both hold somewhere between eight in ten and nine in ten, and at eighty-one in a hundred if they are independent. The weakest rule reports the top of that range. As a probability it is the generous reading, not the cautious one.\n\n[[pause 3]]\n\nThe note two paragraphs earlier had the bound the right way. The sentence that followed reversed it, and nobody caught it until today.\n\n[[pause 6]]\n\nDoes that sink the weakest rule? No. It sinks one justification for it. The rule can still be right as a measure of standing, of how far a case has been made, where the axioms of probability are not being claimed.\n\n[[pause 3]]\n\nBut that is a choice about what the number means, and we have not made it. That is part eleven.\n\n[[pause 4]]\n\nWhat is real today: the two halves of the rule are now written down, and the code fails the first half on a written-in middle. What this opens: a test for any new arithmetic. Draw the same argument with and without a redundant middle step and demand the same number. Part six in a line.\n\n[[pause 3]]\n\nRedescription must be free, new doubt may cost, and no operator gets that for nothing.\n\n[[pause 6]]\n\n---\n\nPart seven. One reason, and fourteen.\n\n[[pause 3]]\n\nYou asked whether a true, relevant reason passes its strength on in whole. I said yes in principle. The note says: too strong, and here is the gap. Truth, relevance and sufficiency are three different things.\n\n[[pause 4]]\n\nThe kitchen light is on. True, I can see it. Relevant, it bears on whether Anna is home. And weak. Lights get left on. So a true relevant reason can pass on very little, and that is not a defect. It is what the reason is worth.\n\n[[pause 4]]\n\nOur rule from the twenty-eighth of September already has the two factors. In your words, blessing it: about the decision, your proposal is blessed, make it so. A link carries its supporter's own standing, times the link's relevance.\n\n[[pause 3]]\n\nSo in whole only when relevance is full, which is the deductive case, the copper box with a true label. The landline, I said, would settle it. The note says even that is quick: call forwarding, a wrong number. Not to demand endless doubt, but to keep from hiding the assumptions the system exists to show. [exhales]\n\n[[pause 5]]\n\nNow fourteen. Suppose fourteen reasons, each relevant. Do they add? The note says: only with appropriate event definitions and dependence assumptions. Overlapping routes cannot simply be added. Two everyday cases. Fourteen friends say the restaurant is good, and they were all at the same dinner. That is one dinner.\n\n[[pause 3]]\n\nFourteen readings from one thermometer over a week: fourteen readings, and if the thermometer is off, all fourteen are off together. Dependence, not identity.\n\n[[pause 5]]\n\nOur corpus has the question open under two names. Whether reasons add at all is contested. The philosopher John Pollock, whose weakest-link rule we borrowed for parts, rejects adding outright: two arguments for one conclusion give you the better of the two, not the sum. The accrual school says they add.\n\n[[pause 3]]\n\nProbability says independent routes add in its own way, and overlapping ones do not. Our August note chose summing, deliberately, and said so. And the September review says a stored relevance number means a different thing under each of those structures, so the structure has to be recorded somewhere.\n\n[[pause 5]]\n\nWhat is real today: reasons add, nothing discounts a shared origin, and a discount for shared evidence is built and switched off. What this opens: recording, on each claim, which shape its reasons have, before we pick how they combine. Part seven in a line.\n\n[[pause 3]]\n\n[slowly] A reason passes on what it is worth, and worth has three ingredients, not one.\n\n[[pause 6]]\n\n---\n\nPart eight. Three things that completeness can mean.\n\n[[pause 3]]\n\nYou asked what the completeness check was. I gave you one answer. The note gives three, and all three are in our system under different names.\n\n[[pause 4]]\n\nThe suitcase. Did I list everything that is in it, and nothing that is not? That is content completeness. A faithful split keeps all of what the claim said and adds no commitment. Our check for that is the rebuild: can a reader write the claim back from its parts alone, without having seen it?\n\n[[pause 3]]\n\nIt is built, it runs on the splitter's cuts, and the lift waits on it.\n\n[[pause 5]]\n\nThe recipe. Given these steps and this oven, is any step missing, and is every step used? That is argument completeness. It is relative to a rule. It says nothing about whether the eggs are fresh.\n\n[[pause 3]]\n\nOur check for that is the ruling of yesterday, the two tests: no ordinary situation where the step's sentences hold and its end fails, and no sentence that could be taken away.\n\n[[pause 5]]\n\nThe detective. Have I considered every suspect, every alibi, every alternative story? That is evidence completeness. No split and no argument gives it.\n\n[[pause 3]]\n\nIn our system it is the thorough search that found nothing, the one contribution a person can make that the machine cannot, since what nobody has contributed is the one thing no reading of the graph sees. It is in the ruling of the thirtieth of September, and unbuilt.\n\n[[pause 5]]\n\nSo when I told you the lift's gate is the rebuild and the chain's gate is yesterday's two tests, I was right about which check goes where. What I had not said is that they answer different questions, and that passing one is not passing another.\n\n[[pause 3]]\n\nA chain can pass the two tests with a false sentence in it. The note says this plainly: the tests concern adequacy and irredundancy of a reconstruction, not premise truth, not all relevant evidence, not a transmission rule.\n\n[[pause 5]]\n\nAnd one more thing the note finds about yesterday's ruling. Sufficiency plus nothing-removable does not give you one smallest step. Take: it rained, if it rained the ground is wet, if the ground is wet the path is slippery. So the path is slippery. No sentence can be removed. No ordinary situation breaks it.\n\n[[pause 3]]\n\nAnd it contains two steps, not one. The same rule, applied twice. To get smallest, you have to say which moves count as one move.\n\n[[pause 5]]\n\nOur chain note already has that clause, as my reading and not as the ruling. A step is at its floor when it applies one stated rule, once, to one stated case. Your words on the thirteenth were: the smallest step, such that each step can be addressed and challenged in itself.\n\n[[pause 3]]\n\nYour ruling blessed the two tests. The one-rule-once clause is what makes the two tests reach your words. [exhales]\n\n[[pause 4]]\n\nWhat is real today: three checks, each a home for one sense of complete, two built and one not. What this opens: a third clause for the floor, for you to bless or not. Part eight in a line. Complete about the words, complete about the steps, complete about the world: three jobs, and no check does two.\n\n[[pause 6]]\n\n---\n\nPart nine. Three ways to attack a link.\n\n[[pause 3]]\n\nHere is a distinction the note brings from Pollock and from a formal framework called ASPIC plus, by Modgil and Prakken. When you challenge an argument, you can do one of three things. You can attack a premise. You can attack the conclusion. Or you can attack the connection between them.\n\n[[pause 4]]\n\nThe weather forecast says rain tomorrow. First attack: the data it used was from the wrong station. That undermines a premise. Second: look, the sky is clear and the pressure is rising. That rebuts the conclusion with a reason of your own. Third: that model is useless in October, it always says rain.\n\n[[pause 3]]\n\nThat undercuts the link. The premises may all be true, the conclusion may even come true, and still the step from one to the other was no good.\n\n[[pause 5]]\n\nThe doctor. The test came back positive, so you have the condition. Undermine: the sample was mislabelled. Rebut: your other markers are all normal. Undercut: this test gives a false positive in one case in five among people your age. Three different conversations.\n\n[[pause 4]]\n\nWhy this matters to us. The note says: this is a better model for separately addressable reasoning than treating every challenge as a content split. When I was confused in part four, part of the confusion was that an objection to how an inference runs looked like it should become a part of something.\n\n[[pause 3]]\n\nIt is not a part of anything. It is an undercut, and it has a home of its own.\n\n[[pause 5]]\n\nAnd we have that home. Since the twenty-eighth of September, a link can carry a relevance claim, a claim about whether this reason bears on that conclusion, and that claim can be supported and attacked like any other.\n\n[[pause 3]]\n\nA code reading on the eighth found that an undercutter has no arrow type of its own in our graph; it acts as an answered connection question on the link. So the three attacks have three homes: a premise's own claim, the conclusion's own claim, and the relevance claim under the link.\n\n[[pause 4]]\n\nWhat is real today: the homes exist, and the one under the link was empty on every link in the graph until the tenth. What this opens: sorting challenges by what they attack before deciding what they do to a number. Part nine in a line.\n\n[[pause 3]]\n\n[slowly] A challenge to the step is not a piece of the claim, and it has a place of its own.\n\n[[pause 6]]\n\n---\n\nPart ten. What the Toulouse school adds.\n\n[[pause 3]]\n\nThe note reads three papers from the argumentation group in Toulouse, around Leila Amgoud, which our corpus already calls the IRIT harvest. Three things to take from them.\n\n[[pause 4]]\n\nFirst. An argument is a minimal set of sentences that together entail a conclusion, paired with that conclusion. Minimal means no sentence can be dropped. The conclusion and its support are different objects.\n\n[[pause 3]]\n\nTwo arguments can share a conclusion and rest on different grounds, and exposing those grounds is not discovering two new sets of parts of the conclusion. That is part three again, from the formal side.\n\n[[pause 5]]\n\nSecond, and this is the method. When these researchers choose a way of computing strengths, they first write down principles. Should quality matter more than quantity? Should several weak attacks add up to one strong one? Should two copies of one attack count twice?\n\n[[pause 3]]\n\nThey state the properties, they check which candidate rules satisfy which, they find the conflicts, and then they choose. They do not pick an operator because it is in a paper.\n\n[[pause 5]]\n\nAn everyday version. You are choosing a scoring rule for a baking contest. Before you pick one, you say what it must do: a cake entered twice must not win twice, a tie must be possible, a single judge's grudge must not sink a cake. Then you look at candidate rules and see which ones obey.\n\n[[pause 3]]\n\nOur strength calculus was chosen in March and has been patched since. The principles it must obey have been named in pieces, in several notes, and never in one place.\n\n[[pause 5]]\n\nThird. In their newest work, the relations themselves can be attacked. An attack on an attack. That is close to our relevance claim under a link, and the note is careful: an attack on an attack is not automatically the same as a sentence stating that a support link is relevant. Close, and not identical.\n\n[[pause 4]]\n\nAnd the note's honest limit about all three. These papers start from formalized structures. They do not solve our upstream problem, which is finding, in ordinary language, which sentences are content and which are grounds. That part stays ours.\n\n[[pause 4]]\n\nWhat is real today: the harvest is in the corpus, the principles are scattered. What this opens: one page of principles, each as a test the eight examples in part eleven can run. Part ten in a line. Say what the arithmetic must obey before you pick the arithmetic.\n\n[[pause 6]]\n\n---\n\nPart eleven. What the number means.\n\n[[pause 3]]\n\nThis is the part to keep. The note ends on it, and I think it is the question under your Question One, the one we were about to put to you yesterday: may a chain written under a link raise the link?\n\n[[pause 4]]\n\nBefore changing any arithmetic, the note says, clarify what a displayed number is meant to express. It names four readings. The probability that the claim is true. The confidence warranted by the case that has been examined. An ordinal degree of acceptance under a declared set of rules.\n\n[[pause 3]]\n\nOr a bottleneck measure of one particular argument, how strong its weakest link is.\n\n[[pause 5]]\n\nThese are not the same, and they combine differently. Here is the cleanest everyday picture I have. A bridge. You can ask how likely it is to carry a given lorry. That is a probability, and it takes in everything, including the possibility that the plans are wrong.\n\n[[pause 3]]\n\nYou can ask how much of the inspection has been done and passed. That is confidence from the examined case. You can rank bridges from fine to unsafe with no numbers at all. That is an ordinal degree. Or you can ask: of this one route across it, which plank is weakest? That is the bottleneck.\n\n[[pause 3]]\n\nFour questions, four different numbers, for one bridge.\n\n[[pause 6]]\n\nNow our situation. The corpus has said since August that strength is standing and not probability, that a claim can be strong because nobody has attacked it yet.\n\n[[pause 3]]\n\nAnd on the thirtieth of September you said, in your words: every claim in Deliberus will ultimately have a subgraph which is equivalent to the already-guess in the probability, Bayes perspective, no? This has always been my intuition for Deliberus. [exhales]\n\n[[pause 5]]\n\nThe note's sentence on this is the one I would carry home. Calling a value standing makes the axioms of probability optional. It does not make any chosen arithmetic self-justifying. So both roads need the same thing first: a declared meaning, and principles the arithmetic must obey, tested on cases.\n\n[[pause 5]]\n\nThe note gives eight such cases, and I think they are the specification. In plain words. A claim with two independent arguments for it. Two linked premises, one false, with the conclusion still true. A conjunction cut faithfully into its two halves. An either-or, or an if-then, where the inner sentences are not both asserted.\n\n[[pause 3]]\n\nThe same evidence copied into fourteen arguments, against fourteen distinct readings from one instrument. The same argument with and without a redundant middle step. A newly exposed assumption that should lower confidence. And a very credible statement of a rule that still supports its instance only defeasibly.\n\n[[pause 6]]\n\nAny arithmetic we choose must say what it does on each of those, and we can check it against what we meant. Today's code, run on the eighth, gives a different number with and without the middle step, which is already one failure.\n\n[[pause 4]]\n\nWhat is real today: four possible meanings, none declared, and a founder intuition on record for the first. What this opens: the decision that comes before Question One. Part eleven in a line. [slowly] Say what the number means, then let the eight cases choose the arithmetic.\n\n[[pause 6]]\n\n---\n\nPart twelve. The view from here, and the questions to carry.\n\n[[pause 3]]\n\nFirst what is real today, in one breath. We have three operations with three homes, and arithmetic for two of them. The cap is sound and runs on the right dial. The lift waits on a content check. The chain's floor has two ruled tests and one unruled clause.\n\n[[pause 3]]\n\nReasons add, shared sources are not discounted, and nothing has declared what the number means.\n\n[[pause 5]]\n\nThen what I owe. Four claims of mine were too strong, and I have said each where it belonged: a reason passes in whole, the weakest rule is the only invariant one, fourteen from one source are one, and the bike's brakes are parts.\n\n[[pause 3]]\n\nOne sentence in the August note runs the wrong way and I will correct it today, with a dated line. And one example in the always-loaded file is the root of the confusion, which is yours to change or keep.\n\n[[pause 5]]\n\nThen what the note opens that we did not have. A clean separation of content, argument and case. The two-halved invariance rule, free redescription and costly new doubt. Three senses of complete, each with its check. Three attacks with three homes. A method, principles before operators. And eight cases as the specification.\n\n[[pause 5]]\n\nThe note also leaves a listening list, in case the walk is long. First on it is a conversation with the philosopher of probability Branden Fitelson, where he separates conditional probability from what it means for evidence to be relevant. That is our two factors, from outside, and it is the one to hear next. [exhales]\n\n[[pause 5]]\n\nNow the questions, as questions to turn over. I will give four, and say which comes first.\n\n[[pause 3]]\n\nOne. What does the number on a claim mean? A probability, a confidence from what has been examined, an ordinal degree, or a bottleneck of one argument. This comes first, because every other arithmetic question is downstream of it.\n\n[[pause 5]]\n\nTwo. The canonical example. Should Doctor Smith's four pieces be redrawn as support, which is what your April sentence said, with the principle kept and the example corrected? That is a change to the always-loaded file, so it waits on your word.\n\n[[pause 5]]\n\nThree. The floor of a step. Should the ruling's reading carry the third clause, one stated rule applied once, so that the two tests reach your words about the smallest step?\n\n[[pause 5]]\n\nFour, and only after one. May a chain written under a link raise the link? Yesterday I was going to put this to you with my pick. Today I think it cannot be answered until the first question is, because what raising means depends on what the number means.\n\n[[pause 5]]\n\nSo one decision, if you take one from this hour. [slowly] What is the number for? Everything else I have said today is in service of that one.\n\n[[pause 4]]\n\nThat is the walkthrough. The note is on disk under today's date, the corrections will be too, and the eight cases are waiting to be used. Thank you for listening.\n"}