{"path":"listening/2026-10-22-what-an-argument-is-made-of.md","content":"# What an argument is made of, and what waits to be built: three days since the morning walk of 20 October\n\n- **What it is**: a listening walkthrough of about forty minutes, made the evening of 2026-10-22. His\n  request: *\"Read up on everything that happened in Deliberus recently. Generate a walkthrough for\n  everything that happened and was built and was reflected and theorized about since the last\n  substantial walkthrough. Put a little extra emphasis on helping me move forward with actionable\n  decisions, look at the different clustering in the new ledger page on the site for inspiration\n  perhaps.\"* It covers what was committed between the morning walk of 20 October (persisted 10:48) and\n  the afternoon of 22 October, in three groups: what an argument is made of (the relevance claim, the\n  weakest premise, two kinds of rule), trying rules out, and the machine and its record. The Codex walk\n  and its companion are recalled in a line. Each part closes on one decision, picked the way the ledger\n  page at `/ledger` sorts its rows: the open rows most other rows name (AD22 and Q4 lead the open rows,\n  eight namers each) and the ruled rows nobody has built.\n- **Covers**: [../research/what-relevance-is.md](../research/what-relevance-is.md) § 3b to § 3f;\n  [../research/retrofitting-the-reasons-channel.md](../research/retrofitting-the-reasons-channel.md) § 9h to § 9j;\n  [../research/normic-rules-normative-reasons-and-trying-rules-out.md](../research/normic-rules-normative-reasons-and-trying-rules-out.md);\n  [../research/theorists-on-normic-rules-exceptions-numbers-and-many-minds.md](../research/theorists-on-normic-rules-exceptions-numbers-and-many-minds.md);\n  [../research/synthesis-build-plan.md](../research/synthesis-build-plan.md) § 10;\n  [../research/pass-quality-and-the-choice-of-model.md](../research/pass-quality-and-the-choice-of-model.md) § 10;\n  [../research/what-he-held-in-the-group-2011-2013.md](../research/what-he-held-in-the-group-2011-2013.md);\n  [../research/what-the-group-held-2011-2017.md](../research/what-the-group-held-2011-2017.md) (part eight; written the same evening by another session, read before it was committed);\n  ledger AD22, Q4, Q38, AT1 to AT10, AS2, AP2, R1, AG12, Q34, AG14, R8 ([../research/session26-open-rulings.md](../research/session26-open-rulings.md));\n  `deliberus/ledger_map.py` and `specs/map-view/design.md` § 3c (the wheel's four ways in).\n- **Chapter list for noting** (rule 33):\n  0. Opening: since the last walks, and the map\n  1. The application is the link's relevance claim: rule AD22?\n  2. The weakest premise, and when arguments add: multiply inside one argument?\n  3. Two kinds of rule: does an unexplained failure count against the rule?\n  4. Trying rules out: the normative trial first, theory before it?\n  5. A reading for one claim, and cheaper models: which map to test on?\n  6. What you held in the group, 2011 to 2013: does the arrow need a vote?\n  7. Ruled, and not yet built: what is built first?\n  8. The group, read again, appended the same evening: who said it first, and the what-if view as the first way in?\n- **Registers used** (rule 57): Lucy the spine; his clone on his own words and seven predictions;\n  Elliot on Pollock, ASPIC+ and Toulmin (`what-relevance-is.md` § 3b), on the Fréchet bounds, Pollock's\n  weakest link and the law's element standard (`retrofitting-the-reasons-channel.md` § 9j), on Schurz\n  and on Bear and Knobe (`normic-rules-normative-reasons-and-trying-rules-out.md` § 1, § 2); Scotty as a\n  rule-keeper who explains away every failure (Popper's conventionalist stratagem, the normic note § 1),\n  as an economist of the seventies (the Lucas critique, the normic note § 4b), and as a tired engineer\n  (ledger R1); Anna on Horty, OpenFisca and LexImpact (the normic note § 3, § 4a); Hope on *independent\n  arguments add* (retrofit § 9h's ruling), on *the death penalty deters* (the graph's two claims), and on\n  a claim reading's unstated premise (`synthesis-build-plan.md` § 10); Myrddin in every part, the\n  signposts drawn from RAND's planning (the normic note § 4c). Part eight: Anna on Mercier and Sperber's\n  argumentative theory, *truth wins* (`../academic-foundations.md` § 7); Elliot on Merton's Matthew effect and\n  Stigler's law of eponymy (from memory; the corpus holds neither) and on Grice's implicature (from memory);\n  Scotty mocking a group about argument maps that argued in a feed (the group note § 1.16, his 2012 line on\n  the linear history); Hope on *science is a subset of consensus tech* (the group note § 1.14); Myrddin on\n  giving the names back.\n- **Correction spoken** (rule 37): the stress test of 22 October read a link's relevance as staying near\n  one half whatever exceptions stand; an outside read found the script hung the rule as a required part\n  of the application, so the flat reading measured the parts discount on the wrong object (commit\n  940c6f6). Not spoken: the § 3f sentence *the relevance of a link in a case is that credence*, refined\n  by AT8; the read-back's corrections of 20 October, told in the Codex walk.\n- **Met again while preparing, one sentence spoken**: the live link from the ACLU's murder-rate\n  statistic (`claim_9184c6060cc0`) to *The death penalty does not effectively deter violent crime or\n  murder* (`claim_eb8a71030abb`) is stored as ATTACKS, with `provenance: extraction`, while its own\n  explanation says the statistic refutes *the death penalty deters crime*: the polarity is inverted.\n  Already recorded in `TODO.md` on 2026-10-17 (the sameness sift, *one link, unrepaired*); still unrepaired.\n- **Numbers kept**: dates; *a third* (the red-sky proportion, the idea of a statistical rule); *half*\n  (the cheap models' price against Gemini Flash); *a few dozen* and *fewer than twenty* (the ledger's\n  ruled rows without a build); *eight* (the namers of AD22). The extraction table, the wheel's counts\n  and the stress-test figures stay in their notes.\n- **His words quoted**: 21 October on the invention (`what-relevance-is.md` § 3b); 22 October, the\n  ruling of AT1, the retired names (retrofit § 9h), his proposal on normative reasons (the normic note\n  § 0), his standing rule on theory first (`what-relevance-is.md` § 3f, cut after *spend any tokens*);\n  20 October on filling the graph and on thinking (`synthesis-build-plan.md` § 10,\n  `pass-quality-and-the-choice-of-model.md` § 10); 14 October to Codex (retrofit § 9h's table; typed\n  in Codex, so the record lookup of rule 56 may not find it); two passages from the Facebook group,\n  2011–12 (`what-he-held-in-the-group-2011-2013.md` § 1.1, § 1.6; written in the group, not typed in\n  a session, so the lookup will not find them). A bare approval (*\"Yes rule both!\"*) is said in a word.\n- **Predictions**, from `listening_context.py --prepare --said` (seven queries) and the notes: part 1\n  from his words of 19 September (relevance judged by machine, challenged by people) and 17 October\n  (a support link carrying a claim about relevance); part 2 from his words of 15 October (the number\n  says the strength in the world, ledger Q41) and the outside read's rerun; part 3 from AT3's guard and\n  his 21 October question on exceptions; part 4 from his standing rule of 22 October; part 5 from his\n  objection of 20 October and his answer of 22 October (*fill up more of the graph*); part 6 from AG12\n  and the ruling of 31 August that bare judgments are telemetry; part 7 from the correction commit's\n  rerun and his ask for the ledger page.\n- **Not covered, no decision raised**: `careful-and-broad-minds.md` (what the research says about\n  exactness and gist); `one-challenge-one-home.md` § 4b (negated attacks, positive wording on display\n  only); the wheel map's error fixes (live); the map-view mockup and AP1 (a question a map holds is its\n  own node, ruled); the outside check voiding its own reading; the group archive's new search kind and\n  its people in the network register; AT4 (rules about groups of people start statistical), AT6, AT7.\n- **The dry run's notes, read and left**: six quotes not found by the record lookup; each was found\n  as typed in his own messages (14 October, session 745c9abd; 20 October, session a236b4e4, filed under\n  another project's folder; the two group passages are from 2011–12 and outside the record). Four\n  sentences in his voice flagged as not him speaking are his own generic *you* (*maybe you are working\n  on a new invention*, *you could imagine*). Number words run over three a part, as in earlier pieces:\n  they are *one argument*, *two kinds* and the like, none a tally. Part four and part seven carry one\n  example each of the kind the counter recognises; part four's outside cases (OpenFisca, LexImpact) and\n  the tax under the Lucas critique are theory, per rule 51.\n- **Part eight, appended the same evening** on his request: *\"Append 10 minutes about all this please, deep\n  reflections on it and commentary by the cast.\"* (he was reading `what-he-held-in-the-group-2011-2013.md` at\n  its summary paragraph). It draws on the companion note `what-the-group-held-2011-2017.md`: four 2026\n  rulings first said by Alex Vance and Tomas Kronvall and a philosopher in 2012, three warnings the group let\n  pass, eighteen ways in left unbuilt. His 2012 group lines (*\"I... don't understand, at least not right\n  now\"*, the linear history, the rough diamond) were checked word for word against\n  `data/fb_group_archive/2026-10-22_archive.json`; the record lookup of rule 56 cannot find them, since they\n  were typed in the group. Other members are named only where `docs/history.md` names them; the warnings are\n  paraphrased, not quoted. The decision is the first of the eighteen ways in; the credit to the 2012 members,\n  the other candidate, was already being written into the ledger rows by the session that wrote the note.\n  Prediction from his words of 17 October (views that show the ontology, *\"Where do I click?\"*) and 20\n  September (click and drill through the map like a game).\n- **Status**: snapshot of 22 October. Of what it describes as ruled, only the scheme questions of AT9\n  and the practical-reasoning question are built (committed 22 October).\n- Method: [how-to-write-for-listening.md](how-to-write-for-listening.md).\n\n---\n# What an argument is made of, and what waits to be built\n\n[[scene Lucy opens an evening walk for Fredrik on 22 October. Three days of work lie behind it. She is calm and warm, a colleague laying out the map before a long talk.]]\n\nWhich of this week's decisions would unlock the most?\n\n[[pause 5]]\n\nThis walk covers three days of work in Deliberus. From the morning of 20 October to this afternoon.\n\nThe silences are deliberate. They are room to think.\n\nNo chapter notes came back from the last walks. So there is no report to answer.\n\n[[pause 4]]\n\nThe morning walk of 20 October ended on four groups of decisions. One has moved since. When a person keeps a link a machine wrote, the keep now holds only until the machine's own check of that link is proven reliable. You ruled that in the small hours today.\n\n[[pause 4]]\n\nThe walk on Codex's review found one argument counted twice. That fix is still unbuilt. It waits on how a joint argument is drawn.\n\n[[pause 4]]\n\nHere is the map, in three groups.\n\nFirst, what an argument is made of. The link's relevance claim. The weakest premise. And two kinds of rule.\n\nSecond, trying rules out. What the graph could do with rules held in supposition.\n\n[[pause 4]]\n\nThird, the machine and its record. A machine's reading of one claim. What you held in a Facebook group, long ago. And what is ruled but not yet built.\n\n[[pause 4]]\n\nEach part ends on one decision. I picked them the way the new ledger page sorts its rows. The open rows that many other rows name. And the ruled rows that nobody has built.\n\n---\n\n# Part one. The application is the link's relevance claim\n\n[[scene Lucy opens the part that decides the others: what a link's relevance is. Elliot places it in three traditions. Fredrik's own voice gives a case he raised, and later the prediction. Myrddin closes with a few sentences of proportion. Mood: clear and focused.]]\n\nWhen a reason supports a claim, what exactly is relevant? The reason, or something about the reason?\n\n[[pause 5]]\n\nTake a link from the graph on the site. Anthropic has stated that it grows hard to tell whether its models crossed certain thresholds. That claim supports another. Frontier AI companies struggle to measure how much research their models do.\n\nThe link says something of its own. This one company's case bears on all of them.\n\n[[pause 4]]\n\nThat sentence is what we now call the application. The case bears on the conclusion. And the finding of 21 October is simple. [slowly] The application is the link's relevance claim.\n\n[[pause 4]]\n\nThe slot has existed since September. A link's relevance claim is the link made into a claim. People can attack it. What was new was seeing what goes in it.\n\n[[pause 4]]\n\nNow look at what the graph stores for that link today. Two machine claims, from the extraction in March. One says Anthropic's difficulty is representative of the industry. The other says it is not. Both point at the conclusion.\n\n[[pause 4]]\n\nUnder the new reading, the second one is an exception. It attacks the application, not the conclusion. Anthropic may be unusual, and the industry may still struggle.\n\n[[pause 5]]\n\n[[voice elliot]]\n\nThree traditions put the exception in that same place.\n\nJohn Pollock calls it an undercutter. It attacks the step and leaves the conclusion alone.\n\nASPIC plus gives every defeasible rule a name. The name is a sentence saying the rule applies. An undercutter argues against that sentence.\n\n[[pause 4]]\n\nToulmin wrote it as the rebuttal. Unless.\n\nWhat the three share is this. An exception does not deny the conclusion. It denies that this reason reaches it.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nUnder the application sits a rule. Here, something like: what one frontier company finds hard, the others find hard too. Early this morning the rule found its place. It is a reason for the application. It is not one of its parts.\n\n[[pause 4]]\n\nThat brings my one correction. Last night you saw a stress test. It said a link's relevance barely moves, whatever exceptions stand. An outside reading found the test had hung the rule as a required part. So it measured the discount for parts, on the wrong object. The test has not been run again.\n\n[[pause 5]]\n\nOn 21 October you gave a case that sharpened all this.\n\n[[voice fredrik]]\n\n[thoughtful] Like... maybe you are working on a new invention but don't realize until after a decade it can't work ever because of some previously unknown quirk of physics. So then it was in a way incredibly relevant but you never knew the extent.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nThe premise there was: no law of physics forbids the mechanism. It was relevant from the first day. What changed over the decade was its doubt. And nobody had written it down. An unwritten premise has no score at all.\n\n[[pause 4]]\n\nSo a machine will ask, of each application: describe a situation where the rule holds and the conclusion fails. You chose that question as standing work. Work the machine keeps doing across the whole graph.\n\n[[pause 4]]\n\nMy take. This makes relevance a claim like any other. It can be argued, and it can be wrong. That is exactly what a number guessed by a judge could never be.\n\n[[pause 4]]\n\n[[voice myrddin]]\n\nMost of what matters in an argument hides in the step, not in the facts. People argue the facts because the facts are visible. Give the step its own sentence, and it gets a place to be argued.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nHere is the decision. The ledger row is called: a step's rule is the grounds of its link's relevance claim. It is open. Eight other rows name it, and no open row is named by more. Should it be ruled now?\n\n[[pause 5]]\n\nMy prediction of your answer.\n\n[[voice fredrik]]\n\n[firmly] Rule it. The application is the relevance claim, and the rule sits under it as a reason. Then every exception has one place to land.\n\n[[pause 4]]\n\n[[voice lucy]]\n\nToday, the slot exists and nothing writes into it. Ruled and built, every link could carry its own sentence. And someone standing in the case could narrow it with a fact only they know.\n\n---\n\n# Part two. The weakest premise, and when arguments add\n\n[[scene Lucy turns to what was ruled in the small hours: how premises and arguments combine. Fredrik's own voice gives a correction he made, then a stated position. Elliot brings the arithmetic's theory; Hope asks what a sentence means to a newcomer; Myrddin speaks briefly. Mood: precise, a little lighter.]]\n\nA claim has two arguments. Should two count for more than one?\n\n[[pause 5]]\n\nIn 2011 you asked this with a squirrel. The squirrel is dead, because it smells of decomposition, and dead squirrels do. And because it has not eaten for a week, and squirrels need to eat.\n\nThose are two arguments. Each has premises that count only together.\n\n[[pause 4]]\n\nIn the small hours today you ruled both halves. Within one argument, the weakest premise governs. Across arguments, the best one counts until their independence has been checked. Then independent arguments add.\n\n[[pause 4]]\n\nA machine pass checks independence: a step the machine runs over the whole graph. This one is the keeper, which already decides whether two claims say the same thing. Its second job would ask of two reasons: setting the claim aside, would establishing one make the other more likely?\n\nIn the same turn, you retired a word.\n\n[[pause 4]]\n\n[[voice fredrik]]\n\n[impatient] Stop being so goddamn poetic all the time, this is what's causing the muddling I've been sensing and complaining about for some time now.\n\n[[voice lucy]]\n\nSo the object is an argument. Line, rope and argument bundle are gone as names for it.\n\n[[pause 5]]\n\n[[voice hope]]\n\nI heard: independent arguments add. So if I find lots of weak reasons, my claim gets strong? That sounds like winning by volume.\n\n[[voice lucy]]\n\nOnly if they are independent. Many rewordings of one reason are one argument. Two reports from one source count once. And every reading of independence can itself be attacked.\n\n[[pause 4]]\n\nThe deeper row behind that is open. It is called independence as structure. Sources become nodes, and inputs from one origin count once. On 14 October you wrote to Codex:\n\n[[voice fredrik]]\n\n[dryly] Overlapping roots of evidence obviously shall not be added up.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nNow the open question. Inside one argument, should premises multiply instead?\n\n[[voice elliot]]\n\nTake two contacts on an alarm, each very likely closed. The weakest premise reports that likelihood. But if the contacts fail for different reasons, both are closed less often than either one.\n\n[[pause 4]]\n\nAdd more premises, and the gap grows. With enough of them, all hold less often than not. The weakest premise still reports near certainty.\n\n[[pause 4]]\n\nWith unknown dependence, the truth lies between a floor and the weakest premise. So the weakest premise is a ceiling. It can only overstate. Mathematics calls these the Fréchet bounds.\n\n[[pause 4]]\n\nPollock chose the weakest link on purpose. He built it on rejecting probability for reasons. The law's element-by-element standard does the same. Probability asks something else. How likely is it that everything the argument needs is true?\n\n[[voice lucy]]\n\n[[pause 4]]\n\nMy take follows the proposal in the retrofit note. Both defaults should err low. Within one argument, multiply by default. Take the weakest premise only where the keeper reads two premises as leaning on each other.\n\n[[pause 4]]\n\n[slowly] Then reading the record only ever raises a number. An argument nobody has checked never looks stronger than it is.\n\n[[voice myrddin]]\n\nA chain is not as strong as its weakest link. It is weaker, when its links can break for different reasons. The old saying flatters the chain. Your graph should not.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nHere is the decision. Within one argument, do premises multiply unless the keeper reads them as dependent? It would revise the first sentence you ruled today.\n\n[[pause 5]]\n\nMy prediction.\n\n[[voice fredrik]]\n\n[earnest] Yes, multiply. The number should say what holds in the world, and in the world many shaky premises rarely all hold. But rerun the test cases with my expectations written down first.\n\n[[pause 4]]\n\n[[voice lucy]]\n\nToday, the arithmetic on the site still adds every reason, independent or not, and neither ruled sentence is built. Built, a claim's number would rise only when its record is read. Never because nobody looked.\n\n---\n\n# Part three. Two kinds of rule\n\n[[scene Lucy opens the theory of the night: whether a rule's word usually is a hedge. Fredrik's own voice gives his ruling. Scotty plays a rule-keeper who explains away every failure; Elliot brings Schurz and the psychology of normal; Hope reacts to a sentence; Myrddin speaks briefly. Mood: curious, a little playful.]]\n\nA rule says usually. Is usually a hedge to strip, or part of what the rule says?\n\n[[pause 5]]\n\nYou asked exactly that, on 22 October. The answer is that usually is a scope word. It says how many cases the rule covers. Strip it, and the claim changes.\n\n[[pause 4]]\n\nOne question decides how to make it exact. When the rule's condition holds and its outcome fails, does the failure call for an explanation?\n\nIf yes, the rule is normic. That is Martin Smith's term. It becomes a universal, with named exceptions. The gate is open every Tuesday. It is closed on public holidays.\n\n[[pause 4]]\n\nIf no, the rule is statistical. The proportion goes into its words. About a third of red-sky mornings bring rain.\n\nEither way, the rule claim keeps one number. How sure we are that the rule as written holds. You ruled that the same night.\n\n[[voice fredrik]]\n\n[delighted] This is what I was after. The second number falls. Interesting.\n\n[[pause 4]]\n\n[[voice lucy]]\n\nAnd you ruled that the pass writing a rule must ask the explanation question.\n\n[[pause 5]]\n\n[inhales] Now a real disagreement from the graph. In the death-penalty debate, the ACLU claims the death penalty does not deter murder better than other punishments. Under it sits a statistic. States with the death penalty had higher murder rates.\n\n[[pause 4]]\n\nThe rule under that link is statistical. Where a punishment deters, places with it should see fewer murders.\n\n[[pause 4]]\n\nErnest van den Haag argues from somewhere else. A major reason the public views murder with horror, he writes, is that murderers are executed. Read that way, deterrence works through what people take as normal. Each murder is a failure that calls for an explanation.\n\n[[pause 4]]\n\nSo the two sides may disagree about the rule's kind. Not only about the cases.\n\nOne aside, which I came across while preparing. That stored link from the statistic is an attack. Its own explanation says it refutes the opposite claim. So its direction is wrong. It was first noticed on 17 October, and it is still unrepaired.\n\n[[pause 5]]\n\n[[voice hope]]\n\nSo when one side says the death penalty deters, they could mean two different kinds of rule? And argue past each other for decades?\n\n[[pause 4]]\n\n[[voice lucy]]\n\nYes. And that is the moment the lens page wants to show. Two people agree on what happened. One sees an exception that needs a reason. The other sees the ordinary run of things.\n\n[[pause 4]]\n\n[[voice scotty]]\n\nThese are my words, in the manner of a rule-keeper who never loses. A murder where the penalty stands? The man was drunk. Another? He never expected to be caught. Another? Mad. My rule has never failed once!\n\n[[pause 4]]\n\n[[voice lucy]]\n\nThat is what Popper called a conventionalist stratagem. The guard follows from the form. An exception is itself a rule. It must name a kind that applies to other cases, and can fail on its own. An exception that fits one case only describes the failure.\n\n[[pause 4]]\n\n[[voice elliot]]\n\nGerhard Schurz adds a caution. A normic law still has a frequency of exceptions. Naming them does not abolish it. It moves it into the exception claims, and into how often their conditions hold.\n\n[[pause 4]]\n\nAnd Bear and Knobe found that the ordinary word normal blends two things. What usually happens, and what ought to happen. So whether a failure needs explaining can carry a value inside it.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nOne more open row follows. In a case, a link's relevance is the rule's strength, times how often its outcome follows. For a statistical rule, its words state how often. For a normic rule, it is the chance that no named exception holds here.\n\n[[pause 4]]\n\nMy take. The explanation question is the most useful single question this week produced. It turns a vague word into exceptions anyone can argue, or a proportion anyone can check.\n\n[[pause 4]]\n\n[[voice myrddin]]\n\nWhat surprises a person tells you what they take as normal. Asking what needs explaining is gentler than asking what someone believes. It finds the value without demanding a theory.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nHere is the decision. The row is called: an exception or a disconfirmation. When a rule fails and nobody names a kind of reason, does the failure count against the rule? And must every exception name a kind that could fail elsewhere?\n\n[[pause 5]]\n\nMy prediction.\n\n[[voice fredrik]]\n\n[curious] Yes, on both. An exception that only fits the one case is an excuse. Enough unexplained failures, and the rule was statistical all along, or wrong.\n\n[[pause 4]]\n\n[[voice lucy]]\n\nToday, nothing writes a rule's kind. The step writer that would ask the question is unbuilt. Built, the word usually becomes either a list of exceptions or a stated proportion. Both can be argued.\n\n---\n\n# Part four. Trying rules out\n\n[[scene Lucy opens Fredrik's own proposal of 22 October: running rules in the mind of Deliberus. His voice gives the proposal and his standing rule. Anna brings what already exists elsewhere; Scotty warns in an economist's voice; Myrddin closes with the humility of planning. Mood: visionary but sober.]]\n\nCould the graph run a policy before anyone adopts it, with no model pretending to be anyone?\n\n[[pause 5]]\n\nThe idea is yours, from 22 October.\n\n[[voice fredrik]]\n\n[in wonder] This seems to relate to normative reasons in Deliberus also. The normic rules. You could imagine a normative claim or set of claims being tried out as rules hypothetically, for example.\n\n[[voice lucy]]\n\nYou went on: a way to simulate, in the mind of Deliberus, what complex policies could do.\n\n[[pause 4]]\n\nIt splits in two. Keeping the two apart is most of the design.\n\nThe first is a normative trial. Take a set of rules, held in supposition and asserted by nobody. Add the facts of some cases. Then compute what the rules prescribe.\n\n[[pause 4]]\n\nWhich rules apply. Which exceptions bar them. Where two rules conflict and nothing settles it. And which cases no rule decides.\n\nThat needs no language model. It is exact and cheap. Its best outputs are the conflicts and the gaps. Each is a place where the policy is underspecified.\n\n[[pause 5]]\n\n[[voice anna]]\n\nThis exists in pieces elsewhere. John Horty modelled reasons as default rules with priorities. OpenFisca computes taxes and benefits for test households. And LexImpact lets French deputies see an amendment's effects before they vote.\n\n[[pause 4]]\n\nHere, each rule would also be a claim people can attack. And a gap the trial finds is offered as a move beside it.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nThe second is a consequence trial. What would happen if the rules were followed? That needs statistical rules about how people behave.\n\n[[pause 4]]\n\n[[voice scotty]]\n\nThese are my words, in the manner of an economist from the seventies. You measured how people behave under the old tax. Lovely. Now you change the tax. Why would they behave the same?\n\n[[pause 4]]\n\n[[voice lucy]]\n\nThat is the Lucas critique. So every rule a trial uses is written a second time, scoped to the policy. Under the new policy, does this still hold?\n\n[[pause 4]]\n\nHere it meets the live graph. Anthropic's framework, as mapped, raises safeguards as assessed capability rises. Beside it sits Anthropic's own statement. It grows hard to tell when a threshold is crossed.\n\n[[pause 4]]\n\nScope the rule to the policy. Once safeguards hang on a measurement, does the measurement still track capability? That is one of seven questions the ledger now lists for such a rule.\n\n[[pause 5]]\n\nToday you chose two additions to the argument schemes. A causal link is now asked whether it still holds under the action. And the scheme for practical reasoning now asks whether the action achieves the goal at all. Before, it asked only whether the means were best, feasible or harmless.\n\n[[pause 4]]\n\nMy take. Simulate the rule set, never the person. The rules are people's own words, and the trial derives from them by computation.\n\n[[pause 4]]\n\n[[voice myrddin]]\n\nA plan names what it hopes. A good plan also names what it assumes, and watches for the first sign that an assumption is failing. The planners at RAND called those signposts. The humility is in the watching.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nIn short, your standing rule, from the same day:\n\n[[voice fredrik]]\n\n[earnest] If there's one thing you should know about me it's that I always wanna settle the theory and whatever we can know from first principles and philosophy before we do any empirical testing or eval or spend any tokens\n\n[[voice lucy]]\n\n[[pause 4]]\n\nSo here is the decision. The row is called: trying out what a rule set prescribes. Should the first trial be the normative one, with no model, its theory settled first? What counts as a case, how priorities are written, and what the trial reports.\n\n[[pause 5]]\n\nMy prediction.\n\n[[voice fredrik]]\n\n[firmly] Theory first, always. The normative trial needs no model, so it's the honest place to begin. Settle what a case is and how priorities are written. Then run it on one policy we've already mapped.\n\n[[pause 4]]\n\n[[voice lucy]]\n\nToday, the two new scheme questions are committed, and the trials are unbuilt, their theory too. Built, a policy's gaps could be read before anyone votes on it.\n\n---\n\n# Part five. A reading for one claim, and cheaper models\n\n[[scene Lucy reports what was built for the map's hover card and what a comparison of models found. Hope reacts to one sentence of a machine's reading. Fredrik's own voice gives two objections he made, then the prediction. Myrddin speaks briefly. Mood: practical and grounded.]]\n\nWhen a machine writes three sentences about a claim, how would you know it got the relations right?\n\n[[pause 5]]\n\nOn 20 October you asked for a reading on your new map, the one that turns with the scroll wheel. I will call it the map view. The reading shows when you hover over a card. A short reading of one claim and the claims around it. It is built. It stays a draft until a person publishes it.\n\n[[pause 4]]\n\nThe first live run hid nearly all the claims the focus rests on. A budget meant for casual depth did that. Now each side of the claim has its own limit.\n\n[[pause 4]]\n\nThen you blessed three repairs. Each sentence now names its claims and the relation it states. Code checks every relation against the stored link, and refuses to publish a mismatch. A few rare claims get marked places. And the pass is handed the rulings that bear on it.\n\n[[pause 4]]\n\nOn two readings of the custody claim, every stated relation matched. Both named a premise nobody states: provider custody risks are left unmodeled.\n\n[[voice hope]]\n\nI read that sentence. Is it the author's claim, or the machine's?\n\n[[pause 4]]\n\n[[voice lucy]]\n\nIt is marked: nobody said this out loud. Whether a stranger hears that mark is what your reading of the cards will tell.\n\n[[pause 5]]\n\nThen you objected.\n\n[[voice fredrik]]\n\n[thoughtful] This is a good plan but I feel like the claim graph has to be allowed to fill up a bit more under whatever topic or claims we wanna test this out on\n\n[[pause 4]]\n\n[[voice lucy]]\n\nMeasured that evening, every claim on the custody map comes from one source. Daniel Lindqvist's paper, mapped by hand. A reading would summarise one critic's view, and no check above could see it.\n\n[[pause 4]]\n\nThen the models. You asked whether something cheaper than Gemini could carry every pass, extraction included. One article was extracted by several models, against a reading made by hand.\n\n[[pause 4]]\n\nOpenAI's Luna and Anthropic's Haiku reached far more of the article than Gemini Flash. They cost about half as much. They did it by copying nearly every sentence.\n\n[[pause 4]]\n\nCopying brought its own fault. Luna wrote: states attacked by non-state actors will not be able to defend themselves. The author says that would follow from a view he rejects. The claim was put in his mouth.\n\nYou also objected to switching thinking off.\n\n[[voice fredrik]]\n\n[dryly] Disabling thinking sounds drastic.\n\n[[pause 4]]\n\n[[voice lucy]]\n\nSo both settings were run. For one model, the thinking setting changed what kind of extraction it was. The router now lets each run cap how long a model may think. No model is admitted yet. Your blind reading of a sample comes next.\n\n[[pause 4]]\n\nMy take. The relation check is the right kind of repair. A model cannot be told to be careful. A written relation can be counted.\n\n[[voice myrddin]]\n\nA summary of one voice, however fair, is still one voice. Wait until the other side has spoken in its own words. Then read what the machine says about both.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nHere is the decision. The row is called: claim readings in bulk. Before that, which map are the first readings tested on? The death-penalty map, with two opposed sources already in it. Or the custody map, after its other side's own texts are mapped by hand.\n\n[[pause 5]]\n\nMy prediction.\n\n[[voice fredrik]]\n\n[earnest] Map the custody side's own texts first, by hand. That's the debate I care about, and a reading of one critic tells us nothing. The death-penalty map can be the control.\n\n[[pause 4]]\n\n[[voice lucy]]\n\nToday, the reading is built and checked, and no map's cards have been generated. With a map of two sides, the hover card could say what holds a claim up, in sentences a stranger can trust.\n\n---\n\n# Part six. What you held in the group, 2011 to 2013\n\n[[scene Lucy reads Fredrik's own words from a Facebook group of 2011 to 2013 and matches them to today's rulings. His voice reads the old quotes, then the prediction. Myrddin speaks of continuity. Mood: warm, a little wondering.]]\n\nWhat do you think you got right back then, and what did you later reverse?\n\n[[pause 5]]\n\nEarly this morning, your posts and comments in the group Collaborative Argumentation Analysis were read whole. Most of what the project holds today was said there, often in a single sentence.\n\n[[pause 4]]\n\nThe first: strength is judged on the arrow, not on the claim. A member asked if people should vote on arguments rather than conclusions.\n\n[[voice fredrik]]\n\n[amused] That is actually what I'm planning. I've been adamant in this respect when people suggest voting on the claims themselves.\n\n[[pause 4]]\n\n[[voice lucy]]\n\nWithin the year you dropped relevance as redundant. A member argued it belongs to the arrow, and you took it back. The two factors on a link, ruled in September, are that design.\n\n[[pause 4]]\n\nOthers. One argument with many wordings, the losing wordings one click away. Built this week as one strength for a group of same claims, and switched off. A concept defined by the statements that use it. Ruled on 19 October: a use is a vote on the meaning.\n\n[[pause 4]]\n\nThree positions you have since reversed. That the inference needs no statement of its own. Today a machine's sentence under a link is a claim the moment it is written.\n\nThat evidence stays outside the system, referred to. In March you ruled that evidence decomposes.\n\n[[pause 4]]\n\nAnd a ladder of validators, deciding what is valid. Today there is no ratify act, and no expert gate.\n\n[[pause 5]]\n\n[takes a slow breath] And the squirrel. You asked whether many reasons must be reduced to one set of premises. Your answer then:\n\n[[voice fredrik]]\n\n[chuckles] Leaning towards no. But I'm sleepy.\n\n[[pause 4]]\n\n[[voice lucy]]\n\nThe archive itself had a fault. Dozens of posts carried text from Facebook's own feed of this year, some under your name. Found, and re-read from the live group, the same day.\n\n[[pause 4]]\n\nMy take. The reversals are the valuable part. Each one is a ruling with its reasons written down. So the record shows not just what you thought, but how you changed it.\n\n[[pause 4]]\n\n[[voice myrddin]]\n\nIt is rare to meet your younger self and find it mostly right. Rarer still to find where it was wrong, and see you already changed course. The thread held. The knots are where you learned.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nHere is a decision the old posts raise. You were adamant that the vote belongs on the arrow. Today the agree-or-disagree vote sits on the claim. Does the arrow need a vote of its own?\n\n[[pause 5]]\n\nMy prediction.\n\n[[voice fredrik]]\n\n[warmly] No. The arrow already got what I wanted: its relevance is a claim anyone can argue. A bare vote on it would just be telemetry.\n\n[[pause 4]]\n\n[[voice lucy]]\n\nToday, the group is searchable beside your other writing, and its people are in your network register. What it opens: any early position can now be checked against your own words, not a summary of them.\n\n---\n\n# Part seven. Ruled, and not yet built\n\n[[scene Lucy reads the new ledger page's own sorting to find what is ruled but waits on a build, and proposes an order. Scotty plays a tired engineer who says the project rules more than it builds. Fredrik's own voice gives the prediction. Myrddin closes. Mood: decisive.]]\n\nOf everything ruled, what would make this week's rulings visible on the map soonest?\n\n[[pause 5]]\n\nThe new ledger page sorts its rows four ways. By state. By arc. By what a row is mostly about. And by how many other rows name it.\n\nBy state, most rows are open. A few dozen are ruled with no build written. Fewer than twenty are ruled and known to be unbuilt.\n\n[[pause 4]]\n\nThe most named of the ruled-but-unbuilt rows form a cluster. A machine's sentence under a link is a claim the moment it is written: partly built. A refuted reason adds nothing: four pieces unbuilt.\n\n[[pause 4]]\n\nThe smallest step, as the standard for what a machine writes: tried in a script, never in a pass. And this week's two: the second number falls, and the pass asks whether a failure needs explaining.\n\n[[pause 5]]\n\n[[voice scotty]]\n\nThese are my words, in the voice of a tired engineer. You rule on Monday. You refine on Tuesday. On Wednesday you correct the refinement. Meanwhile the arithmetic on the site reads every relevance claim the same. Who is the ledger for?\n\n[[pause 4]]\n\n[[voice lucy]]\n\nThat critique has a row of its own. It says the project's bottleneck is integration, not ideas. It is open.\n\n[[pause 4]]\n\n[inhales] My proposal for an order. Begin with the rerun of the test cases. The rule hangs as a reason, and your expectations are written before the run. That settles whether premises multiply.\n\n[[pause 4]]\n\nThen build the weakest-premise reading on the relevance claim. Take the link from Anthropic's statement to the industry claim. Today its width on the map is a presumption. After that build, an exception that stands would narrow it.\n\n[[pause 4]]\n\nThen the explanation question in the step writer. Then the counterexample question as standing work, held to a few calls per pass at first, as your brake says.\n\n[[pause 4]]\n\nSome things did get built since the morning walk. Every stored property on a link or a claim is now defined where it is written. A test fails on one that is not. An old field that only repeated who wrote a link was retired.\n\n[[pause 4]]\n\nAnd two builds wait on your look at the local preview. This ledger page. And signing in that returns you to the page you were on.\n\n[[voice myrddin]]\n\nA ruling that never reaches the code is a promise. Promises are easy to make one at a time, and heavy to carry in bulk. Keep them soon.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nHere is the decision. Which ruled thing is built first?\n\n[[pause 5]]\n\nMy prediction.\n\n[[voice fredrik]]\n\n[excited] The rerun first, then the weakest-premise reading, so this week's rulings show up on the map. And deploy the ledger page so I can steer from it.\n\n[[pause 4]]\n\n[[voice lucy]]\n\nToday, this week's rulings live in the ledger and the notes, and almost none in the arithmetic. Built in that order, the next walk could show you a band on a link that narrows when an exception stands.\n\n---\n\n# Part eight. The group, read again: who said it first\n\n[[scene Lucy returns to the Facebook group of 2011 to 2017 after a second, fuller reading of it. Fredrik's own voice reads four of his old lines. Anna brings the study that explains one thread; Elliot names what happened to the credit; Scotty mocks the medium in a borrowed manner; Hope reacts to one old sentence; Myrddin speaks of giving names back. Mood: reflective, warm, a little wry.]]\n\n[[pause 4]]\n\nWhen an idea you thought was yours turns out to be someone else's first, what changes?\n\n[[pause 5]]\n\nSince this walk was written, a second reading came in. Another session read the whole group this time. Everyone's posts, not only yours.\n\nIts finding sharpens part six. Four rulings of this autumn were first said in the group by other members, in 2012.\n\n[[pause 4]]\n\nAlex Vance said that relevance belongs to the arrow, and validity to the claim the arrow comes from. He also said that every link is an if-then.\n\n[[pause 4]]\n\nTomas Kronvall said that the inference is a statement of its own, with its own confidence. And a philosopher in the group said: track how people use a word, and never try to regulate it.\n\n[[pause 4]]\n\nPart six said you took relevance back within the year. Here is how it happened. Alex Vance argued for it, and you answered him.\n\n[[voice fredrik]]\n\n[curious] I... don't understand, at least not right now. Maybe you could upload a simple sketch here to elucidate.\n\n[[voice lucy]]\n\nYour next sketch had both columns.\n\n[[pause 5]]\n\n[[voice anna]]\n\nThere is a study for exactly this moment. Mercier and Sperber argue that reasoning evolved for arguing with others, not for thinking alone. Alone, people mostly confirm what they already hold.\n\n[[pause 4]]\n\nIn a group with a real disagreement, it goes the other way. As soon as one person has the answer and can argue for it, the answer spreads. They call it truth wins.\n\nThat thread is a small case of it. One member had the distinction. He argued for it. You changed your sketch.\n\n[[voice lucy]]\n\n[[pause 4]]\n\n[[voice elliot]]\n\nSociology has a name for what happened next. Robert Merton called it the Matthew effect. Credit flows toward the best-known name in the room.\n\nStephen Stigler later made a law of it. No discovery is named after the person who found it first. And Stigler said he got the law from Merton.\n\n[[pause 4]]\n\nThe project's record did the same thing. It filed these four ideas under your name, or under a research note. Nobody did that on purpose. A record keeps the name it heard last.\n\n[[voice lucy]]\n\nThose rows are getting their earlier sources tonight, from the other session.\n\n[[pause 4]]\n\n[[voice scotty]]\n\nThese are my words, in the manner of the group's most stubborn member. A group about mapping arguments. It argued for six years in a Facebook feed. A feed! The one format where every argument scrolls away by Thursday.\n\n[[voice lucy]]\n\nYou said it yourself, back then.\n\n[[pause 4]]\n\n[[voice fredrik]]\n\n[wryly] It's a problem that everything just vanishes in the linear history.\n\n[[voice lucy]]\n\n[[pause 5]]\n\n[takes a slow breath] The feed swallowed more than order. It let three warnings pass.\n\nA member asked whether implicature would be a problem. That is what an argument means without saying it. Two members answered: never. Today the premise nobody says out loud is the sharpest test of extraction.\n\n[[pause 4]]\n\n[[voice elliot]]\n\nPaul Grice gave it the name. Implicature is what a speaker means beyond what the words say. Conversation runs on it, and so does every real debate. The crux is usually the part neither side says.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nAnother member warned that the tool would become the ultimate instrument of rationalization. People would pick premises to reach the conclusions they already wanted. That is now a hazard the project watches, and nothing yet measures.\n\n[[pause 4]]\n\nAnd later, the author of a rival mapping site named two things whole-tree analysis lacks. First, how representative the mapped arguments are of everything known on the subject.\n\n[[pause 4]]\n\nSecond, how far two subtrees in different words say the same thing. Those are the two problems the corpus still calls its hardest. What nobody thought to bring in. And when two wordings are one claim.\n\n[[pause 5]]\n\n[[voice hope]]\n\nI read one line of yours from that group. Science is a subset of consensus tech. Does that mean science is decided by a vote?\n\n[[pause 4]]\n\n[[voice lucy]]\n\nNo. Read the whole post. You wrote that every inferential step would be mapped, and every fallacy found and corrected kept in the version history. Anyone could check every step. Then this.\n\n[[pause 4]]\n\n[[voice fredrik]]\n\n[earnest] Science is a subset of consensus tech, and in its current form it's a very rough diamond. It's about time we sharpened it a bit.\n\n[[pause 4]]\n\n[[voice lucy]]\n\nConsensus there means agreement earned by steps anyone can check. Not a headcount. In the same group, Alex Vance said that polls carry no reasoning. The project ruled the same this year: votes are telemetry, and reasons are contributions.\n\n[[pause 4]]\n\nAnd the archive kept your rule, against itself. Facebook's own feed text had slipped into it, some under your name. The repair was not silent. It is written down beside the error, with how it was found. Part of it came from setting each post against the group's own preview of it, which no feed text could match.\n\n[[pause 4]]\n\nThat is the rough diamond, sharpened once. The error stays in the record, next to its correction.\n\n[[pause 5]]\n\nMy take. This reading is the project's thesis, shown in the project's own past. One mind could not see the arrow. Another mind could, and said so. [slowly] Then the medium buried the name for fourteen years.\n\n[[pause 4]]\n\nThe ontology the group argued toward is built. What it left unbuilt is the ways in. A map of the page you are reading. A notice when someone attacks your claim. A companion that follows a conversation as it runs. Each was asked for in that group. None of them is ontology. They are doors.\n\n[[pause 4]]\n\n[[voice myrddin]]\n\nAn idea that comes back with someone else's name on it is not a loss. It is proof that you were not alone in the room. Give the name back, and keep the idea.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nHere is the decision. The second reading lists the ways in that the group asked for and nobody has built. The first is Tomas Kronvall's. Grant a few premises, and see what follows from them.\n\n[[pause 4]]\n\nThe arithmetic already runs inside the graph. The hinge score clamps a claim true or false and measures how far its conclusions move. Nothing lets a reader use it. Should this what-if view be the first way in?\n\n[[pause 5]]\n\nMy prediction.\n\n[[voice fredrik]]\n\n[excited] Yes. The arithmetic is already there, it just needs a door. Put it on the wheel: tap a premise, grant it or deny it, and watch the bands move.\n\n[[pause 4]]\n\n[[voice lucy]]\n\nToday, the group lives in two notes, searchable beside your other writing. A what-if view would let a reader do, for the first time, what Tomas Kronvall asked for in 2012.\n\n---\n\n[[scene Lucy closes the walk, brief and warm.]]\n\nThat is the walk. One word a chapter, if you have one: muddled, agreed, or disagreed and why.\n"}