{"path":"listening/2026-10-23-where-dependence-hides.md","content":"# Where dependence hides, and what the site asks of a reader\n\n- **What it is**: a listening walkthrough of about thirty-five minutes, made the afternoon of 2026-10-23.\n  His request: *\"Create a walkthrough that covers all ground since the last walkthrough produced.\"* It\n  covers what was committed after the catch-up walk of 22 October (whose last part was the group\n  archive) up to the afternoon of 23 October, except the two pieces on why independent reasons add in\n  log-odds, which are recalled in a line. Three groups: how reasons combine and where their dependence\n  hides, with the logic underneath; what the site shows a reader and asks of one; how the project\n  steers itself and pays for its machine work. The first piece whose predictions draw on everything he\n  has said (rule 52 as widened the same day: his prompts in every folder, Messenger, the 2011–2013\n  group and the archived AI chats, each question searched with variants, Jev keeping the hits that\n  state his view).\n- **Covers**: [../research/how-independent-reasons-add.md](../research/how-independent-reasons-add.md) § 10;\n  [../research/how-independent-reasons-add/his-readings.md](../research/how-independent-reasons-add/his-readings.md);\n  [../research/folded-routes-and-shared-premises.md](../research/folded-routes-and-shared-premises.md);\n  [../research/derivation-sciencepedia-and-the-logic-core.md](../research/derivation-sciencepedia-and-the-logic-core.md);\n  [../research/what-if-and-lenses.md](../research/what-if-and-lenses.md) § 1 to § 3;\n  [../research/the-feed-reconsidered.md](../research/the-feed-reconsidered.md) § 1;\n  [../research/the-canon-distillation-pass.md](../research/the-canon-distillation-pass.md) § 4d, § 4e;\n  [../research/flywheels-in-building-deliberus.md](../research/flywheels-in-building-deliberus.md) § 1, § 5 and its first check-in;\n  [../research/signposts-for-the-whole-project.md](../research/signposts-for-the-whole-project.md) § 4;\n  [../research/whose-tokens-pay-for-mapping.md](../research/whose-tokens-pay-for-mapping.md) § 3h;\n  [../research/what-he-held-in-the-group-2011-2013.md](../research/what-he-held-in-the-group-2011-2013.md);\n  ledger Q40, Q44, Q46, AS3, AW1 to AW3, AU1 to AU8, AV2, AX1 to AX7 ([../research/session26-open-rulings.md](../research/session26-open-rulings.md)).\n- **Chapter list for noting** (rule 33):\n  0. Opening: where a reason's strength hides\n  1. The independence reading, built and tested: a second model before reasons add?\n  2. Markets and experts, one level down: relabel the case as partly?\n  3. Folded routes: fold long routes on the map, never into a link?\n  4. The logic core: a prover first, or quantities first?\n  5. What-if and lenses: start with one branch seen from the other side?\n  6. The feed, and what a vote is: never ask for a stance vote?\n  7. Flywheels and the ladder: a trigger when rows outrun rulings?\n  8. Ready to show, and what pays for it: what should this month's credit buy?\n- **Registers used** (rule 57): Lucy the spine; his clone on five of his sentences and eight predictions;\n  Elliot on Satopää, Pemantle and Ungar (`how-independent-reasons-add.md` § 10), on Hesse and Shogenji\n  (`folded-routes-and-shared-premises.md` § 5c), on Goldman's test (§ 3 there) and on the framework\n  thesis (`derivation-sciencepedia-and-the-logic-core.md` § 3); Anna on the folded band (§ 5 there)\n  and on Eyal, Steffel and Epley (`what-if-and-lenses.md` § 3a); Scotty as the critic of folding and of\n  the check-in; Hope on *one level down*; Myrddin in every part.\n- **Correction spoken** (rule 37): relevance as a claim on the link was his own idea first, July 2011\n  (commit `8fbf5a6`); the walk of 22 October had told it as the group's.\n- **Objects from the graph** (rule 39): the silicon-and-feeling pair under *Claude's processing is no\n  likelier to be conscious than a simulation*; the eleven-link route from instrumental drives to\n  beliefs about AI descendants; the custody map's claim and its four required parts; the ACLU\n  death-penalty claim `claim_eb8a71030abb` with its inverted link; the feed's modes as measured on\n  22 October.\n- **No collaborator is named**; the Derivation member's thesis is described without its author or title.\n- **His words quoted** (each found in his record by the quote check, which since this piece also reads his\n  Messenger and group lines): 22 October, the what-if view as the mind-meld (typed); 2015, *a consequentialist\n  filter* (Messenger); 29 July 2011, voting replaced by arguing on a meta level (the group); 24 May 2011,\n  *adamant* about voting on claims (the group); 22 October, passes reliable before money (typed). His\n  readiness words of 22 October were not yet in the record and are paraphrased by Lucy.\n- **Predictions** (rule 52 as widened: each question searched with variants in his typed prompts in every\n  folder, Messenger, the group and the AI chats; Jev keeping the hits that state his view; queries in the\n  prepare call of this session). What each drew on:\n  1. A second model before reasons add: 21 September 2026 (narrow passes trusted over time), 29 September\n     (machine work trusted *if and only if* reliable enough), his Q40 ruling (*standing work done in a heartbeat*),\n     2017 (*votes inform the mods' decisions, they don't determine anything*).\n  2. Relabel markets and experts as partly: his case reading of 23 October; 7 October (*unwritten premises*);\n     January 2012 (*wary of confirmation bias*).\n  3. Folded routes: 2017 (*drill down as far as you want* with low response time), January 2012 (*levels of\n     zoom*), 20 September 2026 (*like a video game … drill down/sideways endlessly*), 10 October (*the dark matter\n     of the intermediate chain*).\n  4. Logic first: 21 October 2026 (*the rule isn't certain for a deductive argument either, there are\n     exceptions*), 2021 (*other people judge the validity*), 2015 (*we need the logical structure*).\n  5. What-if: 31 December 2011 (*a filtered view should be visually made obvious … never far away from\n     deactivating it*), 2015 (*a consequentialist filter*).\n  6. No stance votes: 24 May 2011 (*adamant*), 2017, December 2019 (an *I believe this to be true* button,\n     tucked away), 31 August 2026 (*should there even be a voting/ratification beyond the stance vote*).\n  7. Pacing rulings: 22 October 2026 (*always give me a brief overview before … a question*), 20 October (*1-2\n     short sentences*), 10 October and 22 March (*one at a time … the most crucial first*), Q46 (*rule all of\n     this blessed*).\n  8. The credit: 22 October (*passes reliable enough first*), 15 October (*inte helt redo för primetime*),\n     19 August (*I wish the landing page was more exciting*).\n- **Rendered**: 35.1 minutes (forecast 32:21 to 36:23; 149 words a minute while speaking, 106 overall), ten chapter marks; audio in `.private/local-only/listening-audio/`.\n- **Status**: snapshot of 23 October.\n- Method: [how-to-write-for-listening.md](how-to-write-for-listening.md).\n\n---\n# Where dependence hides, and what the site asks of a reader\n\n[[scene Lucy opens an afternoon walk for Fredrik on 23 October. He asked to cover everything since the last walk. She is calm and warm, a guide who enjoys showing where things hide.]]\n\nWhere does a reason's strength come from, and where does it hide?\n\n[[pause 5]]\n\nHere is the short answer. A reason's strength comes from what it rests on. And what two reasons share is often one level down, out of sight.\n\nThis walk follows that thread. Into the map, the feed, and how the work steers itself.\n\nThe silences are deliberate. They are room to think.\n\n[[pause 4]]\n\nSince the walk of the twenty-second, two pieces explained why independent reasons add. This one covers everything else since that evening.\n\n[[pause 4]]\n\nThe map has three groups. First, how reasons combine, where their dependence hides, and the logic underneath. Second, what the site shows a reader, and what it asks of one. Third, how the project steers itself, and pays for its machine work.\n\n---\n\n# Part one. The independence reading, built and tested\n\n[[scene Lucy takes the topic that decides the others: a machine pass that tells two witnesses from one. Elliot brings the forecasters. Myrddin speaks briefly. Mood: curious, precise.]]\n\nTwo reasons stand under one claim. How would a machine tell two witnesses from one?\n\n[[pause 5]]\n\nSince last night, a machine pass asks exactly that. It is called the independence reading.\n\nIt begins by asking whether the two are premises of one argument. If they are, they are not two reasons at all.\n\n[[pause 4]]\n\nThen it supposes the claim is false. For each reason, it asks what most likely explains that reason anyway.\n\nLast, it compares the two explanations. One story, and the reasons count once. Different stories, and they add.\n\n[[pause 4]]\n\nWhy suppose the claim false? Because every reason shares one thing by design: the claim itself. Asked about the claim directly, the machine saw that shared thing everywhere. It called independent pairs dependent.\n\nSupposing the claim false takes that away. What remains is what the reasons share apart from the truth.\n\n[[pause 4]]\n\nThe first live pair showed why the opening question matters. Take the claim that Claude's processing is no likelier to be conscious than a simulation. Under it, two links stood side by side. One said Claude runs on silicon. The other said consciousness is about feeling.\n\n[[pause 4]]\n\nNeither moves the claim alone. Together they make one argument. The pass as first built called them independent, so they would have added.\n\n[[pause 4]]\n\nThen the swap check. The machine reads each pair twice, once in each order. When the two readings disagree, the less independent answer stands.\n\nOn one test, two forecasters who had watched different coin flips were read as one story. A miss, but the safe kind.\n\n[[pause 4]]\n\nThat leaning is the design. A wrong verdict of different counts one story twice, and inflates a claim. A wrong verdict of one story only leaves some strength unused. So every doubt leans toward counting once.\n\n[[pause 4]]\n\nOn twenty-five cases copied from real life, the stronger model's misses all leaned that way. The cheap model read five dependent pairs as different. That is the dangerous kind, so the cheap model does not run this pass.\n\n[[pause 4]]\n\n[[voice elliot]]\n\nForecasters have known half of this for years. Satopää, Pemantle and Ungar showed it for averaged forecasts. How far you may push the average depends on how much the forecasters' information overlaps.\n\nWhat nobody had done is read that overlap from the reasons' own words, pair by pair. That is the new part. It is also the risky part.\n\n[[voice lucy]]\n\n[[pause 4]]\n\n[[voice myrddin]]\n\nA crowd of witnesses is worth what their separate eyes were worth. Ask where each one stood before you count them.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nSo the question to carry. Before two reasons add, should a second reading, by a model of another family, have to agree?\n\n[[pause 5]]\n\nMy prediction.\n\n[[voice fredrik]]\n\n[earnest] Yes, for now, while its error rate is unmeasured. Two readers working apart beat one. Once it is measured reliable enough, drop the second, and let it run across the graph in a heartbeat.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nToday, the reading runs on test cases, and on no live pair yet. Built across the graph, every claim would show which reasons count. And the stories behind each pair.\n\n---\n\n# Part two. Markets and experts, one level down\n\n[[scene Lucy turns to the first case he read himself this morning. Elliot brings Goldman's test. Hope asks what one level down means. Myrddin closes. Mood: warm, appreciative of his reading.]]\n\nPrediction markets favour the incumbent. So do political scientists. One reason, or two?\n\n[[pause 5]]\n\nThis was the first case you read, this morning. The test set said two. You were not so sure.\n\nYou wrote that the experts are part of the market, more or less. And you caught your own belief about who trades, turning into a premise of your reading.\n\n[[pause 4]]\n\nThe record sides with you. Market prices follow the polls closely. And in three famous misses, they erred together: a British election, Brexit, and Trump's first win.\n\n[[pause 4]]\n\nHere is why ordinary years fooled the test set. When the polls are right, both groups are right, and their small errors look unrelated.\n\nBut independence only matters when the claim is false. And when the incumbent loses, the likeliest story for both is the same. The polls were off.\n\n[[pause 4]]\n\nThat is exactly the question the reading asks. So the machine's answers, partly, or one story, were closer than the label.\n\nWritten down to what each belief rests on, both chains pass through one claim. The polls show the incumbent ahead. Over the bare reasons, nobody can see it. One level down, anyone can.\n\n[[pause 4]]\n\n[[voice hope]]\n\nWhen I heard one level down, I pictured a basement. You mean what the reason itself stands on?\n\n[[voice lucy]]\n\nYes. The floor under the floor.\n\n[[pause 4]]\n\n[[voice elliot]]\n\nThe philosopher Alvin Goldman put it as a test. Agreement among believers counts only where they reached it by separate routes. Write the route down, and you can check it.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nSo, a proposal. Take a reason that reports what a group believes: a market, a panel, a poll. It is written down to what that belief rests on. A basis both share then stands under both, and counts once.\n\n[[pause 4]]\n\nThe graph already knows part of this without any model. Among sibling reasons in the graph, about a thousand pairs have one supporting the other. Each within a few links. Code can mark those before a model is asked.\n\n[[pause 4]]\n\nAnd the premise you caught, about who trades? The machine's answer never lists its premises. A second proposal says it should. Name each way two reasons could share an error. And name the premise the verdict hangs on.\n\n[[pause 4]]\n\nYou said people miss this too, especially under motivated reasoning. The remedy is the same for both. The premise is written where someone can attack it.\n\n[[pause 4]]\n\n[[voice myrddin]]\n\nTwo rivers can share one spring. You find it only by walking upstream.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nThe question to carry. Should the case be relabelled as partly, with the polls written in under both reasons?\n\n[[pause 5]]\n\nMy prediction.\n\n[[voice fredrik]]\n\n[thoughtful] Yes, partly. Put the polls in as a claim of their own, under both. And my belief about who trades goes in too, as a premise anyone can attack.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nToday, the case still says different. No reason in the graph is written down to its basis. Built, a shared basis would stand visibly under both reasons, and count once.\n\n---\n\n# Part three. Folded routes\n\n[[scene Lucy takes his idea of collapsible inference leaps. Elliot brings the philosophy of chains. Anna plays with the interface. Scotty objects. Myrddin closes. Mood: playful, exploratory.]]\n\nTwo claims sit eleven links apart. How do you see both at once, without lying about the path between them?\n\n[[pause 5]]\n\nYour idea this morning was collapsible inference leaps. Fold the hops, and see both ends in one view.\n\n[[pause 4]]\n\nThe graph has such routes. The longest runs from the claim that advanced AI systems will exhibit instrumental drives. It passes through capability, agency and self-improvement. It ends at a report about beliefs. Some people believe the consciousness of AI descendants would persist for an extremely long time.\n\n[[pause 4]]\n\nRead whole, the two ends barely bear on each other. Some of its links run from a claim into a report of what people believe.\n\nSo here is the rule underneath. Support does not carry along a chain.\n\n[[pause 4]]\n\n[[voice elliot]]\n\nPhilosophers settled this long ago. A can raise B, and B can raise C, while A does nothing for C. Mary Hesse saw it. Tomoji Shogenji gave the condition.\n\nSupport carries along a chain only where each middle claim screens off what came before. Given the middle, the start tells you nothing more about the end.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nSo a folded route can never become an arrow with a strength. As one arrow, those eleven links would claim what nothing in them establishes.\n\nWhat a fold can show is its health. How many links it folds. Its weakest link. How many links have nothing written under them, or are contested. And which claims it shares with neighbouring routes.\n\n[[pause 4]]\n\n[[voice anna]]\n\nPicture the map zoomed out. Between two claims runs a folded band, thinnest where it is weakest. Tap it, and it opens one level. Tap again, and another.\n\nAnd where a side attack would vanish inside the fold, the band splits. The attack stays in view.\n\n[[voice lucy]]\n\n[[pause 4]]\n\n[[voice scotty]]\n\nMarked as a critic's voice. A folded band is a lie of compression. Anyone can draw a chain from anything to anything, then fold it into something that looks solid.\n\n[[voice lucy]]\n\nThat is why the band wears its weakest link on its face. It carries no strength of its own.\n\n[[pause 4]]\n\n[[voice myrddin]]\n\nA long road is honest when every bend is still on the map.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nThe question to carry. Should the map fold long routes this way? Never into a link, and opened one level at a time?\n\n[[pause 5]]\n\nMy prediction.\n\n[[voice fredrik]]\n\n[curious] Yes. Fold the view, never the data. Let me drill down bend by bend, fast, like a game. And colour the band by its weakest link, so the dark matter between claims is never hidden.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nToday, the map shows only claims near the one in focus, and nothing folds. Built, a reader could hold a conclusion and its far grounds in one view. The gap between them marked honestly.\n\n---\n\n# Part four. The logic core\n\n[[scene Lucy takes the deck he sent and the logic underneath Deliberus. Elliot brings a thesis on frameworks. Myrddin closes. Mood: steady, a little in awe of the scope.]]\n\nSome inferences cannot fail if their premises hold. Should a machine check those, like a calculator?\n\n[[pause 5]]\n\nYou sent the Derivation deck. Its group wants knowledge joined by mathematics and logic. Arguments checked, in their words, like with a calculator.\n\nAnd SciencePedia writes out the reasoning between known results. It calls what is missing the dark matter of knowledge. Your sentence joined them: the inferences in the leap, in the dark.\n\n[[pause 4]]\n\nSo three projects share one diagnosis. Written knowledge keeps conclusions, and drops the reasoning between them.\n\n[[pause 4]]\n\nThey differ in what checks the recovered reasoning. SciencePedia checks only the end, against an answer a machine can verify. Derivation wants formal logic for every inference. Deliberus checks the middle, between two given ends, with small tasks and people. That is where no answer key exists.\n\n[[pause 4]]\n\nSo logic in Deliberus has layers. A deductive layer, where a rule admits no exception and a prover can check it. A computational layer, where a quantity is computed from sourced inputs by a formula. And the gradual layer, where reasons can be defeated, and strength comes from relevance and independence.\n\nOnly the gradual layer is built.\n\n[[pause 4]]\n\nTake your alarm example. The alarm sounds only if contact A and contact B are both closed. Both are closed. So the alarm sounds.\n\n[[pause 4]]\n\nNothing breaks that inference except a false premise. A prover can confirm it, and confirm that no premise could be dropped. Those are already the tests ruled for the smallest inference. That it holds, and that nothing in it is idle.\n\n[[pause 4]]\n\nBut a prover checks form, not faithfulness. In one recorded pipeline, a model filling formal proofs kept slipping the conclusion in as an axiom. More than half its invented axioms were the conclusion itself. Valid, and worthless.\n\nSo the prover would be a second reader of form. The translation into logic is checked first.\n\n[[pause 4]]\n\nAnd quantities. Derivation's own example is the Moon. It drifts from the Earth a few centimetres a year, computed from observations. In Deliberus today, a claim is text. No field holds a value, a unit, or the formula behind it.\n\n[[pause 4]]\n\n[[voice elliot]]\n\nA member of Derivation wrote a thesis on the other half. Knowledge lives in frameworks. What one framework counts as knowledge, another treats as presumed.\n\n[[pause 4]]\n\nFrameworks live side by side through what they share. The only framework above them all is that shared part.\n\nThat is close to your lenses. A holder would see where their own stances conflict. And several lenses would show the claims they all accept.\n\n[[voice lucy]]\n\n[[pause 4]]\n\n[[voice myrddin]]\n\nA calculator never doubts. That is its gift, and its danger.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nThe question to carry. Which layer comes first: a prover for strict inferences, or quantities with units?\n\n[[pause 5]]\n\nMy prediction.\n\n[[voice fredrik]]\n\n[firmly] Deduction first, as a second reader, never the judge. Even a strict rule has exceptions, argued under it. So the prover only says the form holds if the premises do. Then quantities.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nToday, a claim is text, and nothing checks its form or holds its unit. Built, a strict inference would carry a checked form, and a number its unit and its sources.\n\n---\n\n# Part five. What-if and lenses\n\n[[scene Lucy opens the second group: what the site shows a reader. The what-if view and his worldview lenses. Anna brings the studies. Myrddin closes. Mood: warm, imaginative.]]\n\nIf you granted your opponent's premises, how far would their conclusion move you?\n\n[[pause 5]]\n\nOn the twenty-second you heard about a what-if view, and joined it to your worldview lenses. In your words.\n\n[[voice fredrik]]\n\n[in wonder] That's always been my sense of roughly how to effectively render such a mind-meld from the mind of Deliberus in bridging the perspectives/worldviews of two people who don't see eye to eye.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nThe note's answer: these are one operation. Hold some claims granted or denied. Recompute what follows. Show which ones carry the weight.\n\nOnly the source of the suppositions differs. A reader picks them. A holder's own marks supply them. Or a policy's rules do.\n\n[[pause 4]]\n\nAnd the arithmetic exists. The function that computes a claim's strength already accepts fixed values for any claims. Today it is used one claim at a time.\n\n[[pause 4]]\n\nHere it is measured on the custody map. Its claim: once a model crosses a dangerous capability threshold, stronger custody is the right response.\n\nGrant any single premise under it, and the conclusion moves by the same sliver. The ranking meant to say what the conclusion stands on says: everything, equally, barely.\n\n[[pause 4]]\n\nBut whole positions separate widely. Grant every support and deny every challenge, and the conclusion reads high. Do the reverse, and it reads low.\n\nWhere does the gap live? In the required parts, the four measures of custody. Deny one, and the whole is capped. Reasons only nudge. Required parts hold.\n\n[[pause 4]]\n\nA what-if is also an error detector. On the death-penalty map, granting the ACLU's own statistic lowered the ACLU's claim. The stored link pointed the wrong way. Granting a claim's own evidence should never lower it.\n\n[[pause 4]]\n\n[[voice anna]]\n\nAnd the lens must come from its holder. In a run of experiments, imagining another person's view did not make people more accurate about them. Asking them did.\n\nSo a lens holds only what its holder stated. And it succeeds when each holder recognises their own position, not when anyone changes their mind.\n\n[[voice lucy]]\n\n[[pause 4]]\n\n[[voice myrddin]]\n\nTo stand in another's shoes, first ask how the shoes fit.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nThe question to carry. Should the first what-if view be one branch, seen from the position that disagrees? In that holder's own words?\n\n[[pause 5]]\n\nYears ago, you wrote this to a friend.\n\n[[voice fredrik]]\n\n[curious] So you could view the system through a consequentialist filter, or utilitarian filter.\n\n[[voice lucy]]\n\nMy prediction.\n\n[[voice fredrik]]\n\n[warmly] Yes, start with one branch and real holders. Show plainly that the view is supposed, one tap from the plain graph. Then add a tradition's lens, built from its own texts and marked as such.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nToday, no page lets a reader grant anything. Built, two people could see where they part, premise by premise, each in their own words.\n\n---\n\n# Part six. The feed, and what a vote is\n\n[[scene Lucy takes the feed and the ruling on votes, with one correction from the last walk. Myrddin closes. Mood: brisk, a little wry.]]\n\nWhat should the site ask a newcomer to do first?\n\n[[pause 5]]\n\nThe feed was measured on the evening of the twenty-second. One list for everyone, in eight sort orders.\n\nThe mode called needs votes returned a hundred claims, every one with no votes at all. The mode called recent sorted the claims' identifiers as text, and put a hand-built test map first.\n\n[[pause 4]]\n\nYou blessed the redesign. Two halves. A work half brings a person the open places only a person can fill. A reading half orders the graph through the reader's own stances. And no part asks for a vote.\n\nProposed next: the feed asks a newcomer one question. You know this argument. What's missing?\n\n[[pause 4]]\n\nWhy no votes? Ruled the same night. Agreeing makes you a holder of a claim. Disagreeing makes you a holder of its exact negation. Neither moves its strength.\n\nA count says how many clicked, not why. Strength should rest on what was examined.\n\n[[pause 4]]\n\nRecording that ruling turned up a surprise. A vote weight sat in the code. Votes on a claim's parts moved the whole claim. A dozen disagree votes on one required part pushed a claim from even to low.\n\n[[pause 4]]\n\nIt was written in April, days after the opposite decision. A session reached for counts kept to rank the feed. A value made for display had become a value used for judging. It is removed, with a test, and the site has run without it since last night.\n\n[[pause 4]]\n\nOne correction from the last walk. It told the idea of a link's relevance as a claim of its own as the group's. It was yours first.\n\n[[pause 4]]\n\nIn July of the group's first year, you wrote this.\n\n[[voice fredrik]]\n\n[thoughtful] Perhaps in the long run, simple voting on relevance and validity can be augmented or even replaced by arguing about them on a meta level.\n\n[[voice lucy]]\n\nThat is now the ruled design.\n\n[[pause 4]]\n\n[[voice myrddin]]\n\nA crowd can tell you where people stand. It cannot tell you where the ground is.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nThe question to carry. Should the feed never ask for a stance vote?\n\n[[pause 5]]\n\nEarlier that same year, you wrote this.\n\n[[voice fredrik]]\n\n[firmly] That is actually what I'm planning. I've been adamant in this respect when people suggest voting on the claims themselves.\n\n[[voice lucy]]\n\nMy prediction.\n\n[[voice fredrik]]\n\n[dryly] So yes. Never ask for a vote. Keep the stance as a quiet mark of who holds what, never a lever. Ask what's missing.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nToday, the live feed still asks for votes. Built, a newcomer would meet a question only they can answer.\n\n---\n\n# Part seven. Flywheels and the ladder\n\n[[scene Lucy opens the third group: how the project steers itself. Scotty grumbles. Myrddin closes. Mood: practical, gently amused.]]\n\nWhat makes the next ruling cheaper than the last?\n\n[[pause 5]]\n\nYou asked what a flywheel is in this work. The note proposed a test. A loop counts when it closes. When each turn leaves something outside anyone's head. When the next turn costs less because of it. When part of it runs without a push. And when what it saves is named.\n\nHere, what it saves is your attention.\n\n[[pause 4]]\n\nEvery flywheel also has a shadow. It compounds whatever enters it, errors included. A wrong merge spreads as efficiently as a right one.\n\nSo the governor, already ruled. A machine may build on machine output only as deep as its measured error rate allows.\n\n[[pause 4]]\n\nThen the ladder, which you blessed. A ruling goes into the cheapest durable form that acts without being read. First a check in code. Then a line every session reads. Last, the row alone.\n\n[[pause 4]]\n\nWhy that order? A line can be skimmed. A check cannot.\n\nThe vote weight is the proof. Your September ruling on votes was written down, and still never reached the code. A test written the day of the ruling would have caught it.\n\n[[pause 4]]\n\nThe first weekly check-in ran the same evening. Twenty rows added, none ruled. Its verdict: the constraint is your ruling. Proposals open faster than you close them.\n\n[[pause 4]]\n\n[[voice scotty]]\n\nMarked as a critic's voice. So the machine's answer to too many rulings is a weekly report asking for one more ruling.\n\n[[voice lucy]]\n\nYes. One a week, chosen to settle the most.\n\n[[pause 4]]\n\n[[voice myrddin]]\n\nA mill turns only while water falls on it. Mind what the water carries.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nThe question to carry. When a week adds more rows than it rules, what then? Should the next session stop adding, and offer you a ruling sitting?\n\n[[pause 5]]\n\nMy prediction.\n\n[[voice fredrik]]\n\n[wryly] Yes, set it. Then bring them one at a time, the most crucial first, each in a sentence or two with your pick. Where you are confident, let me bless a cluster in one word.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nToday, no signpost has a trigger, and proposals outrun rulings. Built, each week would steer itself toward closing what it opened.\n\n---\n\n# Part eight. Ready to show, and what pays for it\n\n[[scene Lucy closes the walk's topics: when Deliberus is ready to show, and the money that pays for its machine work. Myrddin speaks last. Mood: grounded, hopeful.]]\n\nWhat would make you ask a stranger to try Deliberus?\n\n[[pause 5]]\n\nThe first draft said: wait until close collaborators report insight, unprompted. You corrected it the same evening. You already tell people widely. What waits is readiness. Few people make it through the landing page, you said, and the platform is not yet ready for primetime.\n\n[[pause 4]]\n\nYou named eight parts of readiness. Most are goals. A couple are also signposts, early signs about an assumption the plan rests on. That a newcomer finds the way in. And that the experience speaks for itself.\n\n[[pause 4]]\n\nIn the PauseAI Sweden chat, you linked the site and two maps. You showed sketches of the wheel map view. And you promised an intuitive demo on AI risk.\n\n[[pause 4]]\n\nThat demo would lean on the custody map. It rests on one source: a paper's reading of a company's framework. So readiness, for that demo, means the companies' own frameworks mapped beside it.\n\n[[pause 4]]\n\nAnd here is what can pay for that. Your plan includes Anthropic API credit each month, and this month's lapses in early November.\n\nSince last night, the engine asks Claude models at Anthropic directly. The fixed part of each prompt is cached, so a repeated pass pays full price for it only once.\n\n[[pause 4]]\n\nThat matters because a standing pass is mostly the same instructions, run over and over. Cached, the repeat is cheap, and the money goes to the reading, not the rules.\n\nExtraction itself is still paused. On the same evening, you wrote this.\n\n[[voice fredrik]]\n\n[thoughtful] We do need to get the passes reliable enough first before spending much money.\n\n[[voice lucy]]\n\n[[pause 4]]\n\n[[voice myrddin]]\n\nOpen the door when it opens easily, not when the paint is dry.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nThe question to carry. What should this month's credit buy, before it lapses?\n\n[[pause 5]]\n\nMy prediction.\n\n[[voice fredrik]]\n\n[excited] Proof first, then pages. Run the passes on cases with known answers. Then map both sides of AI risk, so the demo shows what this can be.\n\n[[voice lucy]]\n\n[[pause 4]]\n\nToday, nothing measures whether a newcomer finds the way in. Extraction is paused, and the credit sits unspent. Built, a newcomer would land on a full map, and find the way in within a minute.\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"}