{"path":"research/causes-of-belief-and-the-three-minds.md","content":"# Causes of belief and the three minds: what the psychology of motivated reasoning should become in Deliberus\n\n**Date**: 2026-10-19 · **Type**: the founder's question, the corpus re-read, an outside research pass, and a\nsynthesis · **State**: every proposal here is unruled (ledger Arc AN). The outside sources are search-grade\n(abstracts, publisher pages and summaries), except two arXiv abstracts read on the page (Costello et al.\n2026, Dash et al. 2025). Verify a figure before it reaches a funder or a public page.\n\n## 0. Plain language first\n\nA claim graph holds reasons: what supports a claim, what attacks it, what it rests on. The psychology the\nfounder asks about is about something else: the causes of a belief. People believe things because of who\nthey are, whom they belong to, what they fear to lose, and what they have rehearsed. Models believe things\nbecause of what pleases the person who asks. The question is what Deliberus does with causes.\n\nThe answer proposed here: causes enter the graph in three ways, and each way has one rule. A cause used as\nan attack on a claim enters as a claim, and only after the attacked claim is shown false (C. S. Lewis\nnamed the opposite order *Bulverism* in 1941). A cause that is what a position protects for its holder\n(an identity, a sacred value, a livelihood) enters as that holder's own testimony, beside the reasons and\nnever as weight. A cause that shows as a pattern (the same evidence judged differently by its conclusion)\nis measured by the mind of Deliberus in its own record, across a debate or a population, and never\nattributed to one person. One thing the machine never does in public: it never says why a person\nbelieves something (ruled 2026-10-19, AN5). Behind the scenes, an informed machine guess at the causes of\na belief, kept in a hidden layer by a separate pass, is the founder's own proposal for mid-November (AN13,\n§ 6).\n\nThe research adds one result that ties this to the founder's three minds. The definition of motivated\nreasoning that the literature can measure is **the same evidence judged by different standards depending\non which side it serves**. People do this, and do it equally on the left and the right (Ditto et al.\n2018). Models given a persona do it, and prompts do not stop it (Dash et al. 2025). A scheme's critical\nquestions cannot do it, because a template does not know which side a claim is on. That is a measurable\nsense of the founder's phrase *\"more honest … than most of us\"*: the mind of Deliberus asks every claim\nof a kind the same questions. § 7 lists where it can lose that property.\n\n## 1. His question (2026-10-19)\n\n> What could/should become of all this more gooey psychological/ideological/groupthink/motivated-reasoning\n> stuff in Deliberus? Also dovetails with the three-minds note/reflections I wrote yesterday. Read up.\n> Reflect deeply. Research this extensively online. Ultrathink\n\nIt came in the same hour as the ruling on the fourth critical question of the bias scheme (ledger E21,\n*\"Välsignat, bygg in det, tack\"*), which came from a post on the premise a bias accusation hides\n([cognitive-bias-codex-and-human-contribution.md](cognitive-bias-codex-and-human-contribution.md) § 4.1).\nThe three-minds note is [three-minds-and-the-leap.md](three-minds-and-the-leap.md). Its § 1 holds his\nreflection whole, including the line this note answers most directly: models as *\"fast-and-furious,\nnot-so-careful, latent-space, motivated-reasoning thinkers? Same issue goes for humans I guess.\"*\n\n## 2. What the corpus already holds\n\nThe corpus has absorbed most of the classic literature already. Each finding became something, and none\nof this is re-derived here.\n\n| Finding | What it became | Home |\n|---|---|---|\n| Mercier and Sperber: reasoning is for argument, and works in groups with real disagreement | the group premise under five conditions; frame diversity is selected for, never awaited | [the-group-premise-evidence.md](the-group-premise-evidence.md), ledger L3, L4 |\n| Kahan: reasoning skill makes polarization worse on contested questions; curiosity reverses it | curiosity is a precondition on the wager | [curiosity-as-growth-fuel.md](curiosity-as-growth-fuel.md) |\n| Stanovich: myside bias does not track intelligence; no general bias factor | no expert gate, and no contributor-rationality score, ever | bias note § 3, ledger E6, E7 |\n| Pronin: the bias blind spot is not cured by awareness | never ask a mind to audit its own frame | bias note § 3 |\n| Belief bias, with the ROC refinement | the rigor axis is calibratable; measurable per rater | bias note § 4.5, Arc G |\n| Walton: bias is a subtype of ad hominem | the `bias` scheme, four questions since today | `schemes.py`, ledger E12, E21 |\n| Sunstein: four failures of deliberating groups | minted claims against hidden profiles; *nobody has said X* against cascades; polarization has no structural answer | ledger E16, E17 |\n| Bail; Kalla and Broockman; Wood and Porter | flag versus reasons; arguments alone null, arguments with narrative durable | ledger I2 to I7 |\n| Ginges and Atran: sacred values | symbolic recognition, not more argument; impact testimony as the concession channel | ledger I8 to I12 |\n| Feinberg and Willer: moral reframing | the warrant for the worldview lens | [../worldview-lenses.md](../worldview-lenses.md) |\n| Fernbach and the three replication states | descent humbles more surely than it moderates | Arc H |\n\nWhat the corpus does not have is a single place that says where causes of belief go in the graph. The\npieces sit in five notes, and two rules that follow from them are written nowhere: the order rule in § 4\nand the refusal in § 6.\n\n## 3. The organising distinction: reasons for a belief and causes of it\n\nA belief has reasons, which bear on whether it is true, and causes, which explain why a person holds it.\nPhilosophy keeps the two apart (Sellars's *space of reasons*, and the genetic fallacy). Deliberus's graph is a\nmap of the first. Every mechanism the founder lists in the three-minds reflection (life path, identity,\nreward habit, groupthink, motivated reasoning) is about the second.\n\nThe two meet in three places, and each meeting needs its own rule.\n\n1. A person offers a cause as an attack on a reason: *they only believe P because they are biased*.\n2. A cause is what a position protects for the person who holds it: *if P is false, my community is\n   wrong about something it built itself on*.\n3. A cause shows as a pattern in what people do: the same kind of evidence accepted on one side and\n   rejected on the other.\n\n## 4. A cause offered as an attack: show it false first\n\nC. S. Lewis coined *Bulverism* in a 1941 essay for the habit of assuming an opponent is wrong and then\nexplaining the error instead of showing it. His rule: *\"you must show that a man is wrong before you start\nexplaining why he is wrong.\"* That is the founder's fourth critical question (E21) in Lewis's words, eighty\nyears earlier. It gives the question a name a stranger can look up.\n\nTwo philosophers sharpen why this move matters in polarized debate. Nguyen (*Echo chambers and epistemic\nbubbles*, Episteme) separates a bubble, where outside voices are missing, from an echo chamber, where\nmembers are taught to distrust them. Exposure bursts a bubble. It does not reach an echo chamber, because\nthe trust map was poisoned in advance. Begby (*Evidential Preemption*, PPR 2021) names the poison: *my\nopponents will tell you that q, but I say p*. A pre-emptive bias accusation is that move. It carries the\nhidden premise E21 now asks for, and in the graph it becomes a claim with four questions instead of a\nfilter on whom to hear.\n\nWhat is not yet checked: whether the extraction passes route accusations against a group or an ideology\nto the same scheme. *That community is in an echo chamber*, *that is just groupthink*, *they only say it\nbecause they are paid by the industry* are all the bias move with a group in the `{person}` slot. Nothing\nin `schemes.py` tells the scheme pass that a group fills the slot. A test on real sentences decides it.\n\n**Proposals.** AN1 (*show it wrong before explaining why*): every causal account of a belief offered as an\nattack, whether against a person, a group or an ideology, enters under the `bias` scheme with its four\nquestions. AN2 (*the group in the person slot*): test on real accusation sentences whether the scheme pass\nassigns `bias` when a group is accused, and if it does not, tell the pass that it does.\n\n## 5. A cause as what a position protects: stakes, in the holder's words\n\nThe attunement pole needs causes. A worldview seen from inside is not only premises. It is also what the\npremises protect: belonging, a sacred value, a livelihood, a fear. Kalla and Broockman's durable arm ran\non exactly this (narrative, non-judgmental), and Ginges and Atran found that recognition of what is\nprotected moves people where argument does not. The corpus holds both (ledger I4, I8, I10).\n\nThe research adds a hard constraint on how this material may be produced.\n\n- **Getting beats taking.** Eyal, Steffel and Epley (JPSP 2018), 25 experiments: people told to imagine\n  another person's perspective became no more accurate, and sometimes less, while sometimes more\n  confident. Asking the other person did increase accuracy. Understanding another mind needs new\n  information, not more imagination.\n- **A model given an identity reasons like one.** Dash, Reymond, Spiro and Caliskan (arXiv 2506.20020,\n  Findings of ACL 2026): eight models with political and demographic personas showed up to 9% lower\n  discernment of misinformation, and political personas were up to 90% more likely to evaluate\n  gun-control evidence correctly when the right answer matched their assigned identity. Prompt-based\n  debiasing was largely ineffective.\n\nTogether these say: a lens rendered by a model told to play a conservative, a Buddhist or a socialist is\nimagined perspective, and it is motivated reasoning with a costume. The corpus already prefers lenses\nthat emerge from what holders actually accept ([../worldview-lenses.md](../worldview-lenses.md), the\nbottom-up proposal and the evidential-priors dial). These two results turn that preference into a\nrequirement.\n\nA refinement for the flag-versus-reasons synthesis (ledger I7). Costello, Pennycook and Rand (*Science*\n2024): about 2,190 believers in a conspiracy theory each talked with GPT-4 Turbo, which answered the\nspecific reasons that person gave. Belief fell by about 20%, and the effect held at two months. In a\nfollow-up the model dropped its rapport and gave only facts, and the effect held, by Rand's report. So\narguments did work here, where Kalla and Broockman's arguments alone were null. One difference stands\nout: Costello's arguments answered *this person's own* reasons, while canvassing arguments are generic.\nThe variable may be less *flag versus reasons* than *reasons that meet yours versus reasons in general*,\nwhich is what a descent to the crux does. It also bears on the corpus's wording of the synthesis, *a\nperson's reasons with a face attached* (`does-descending-change-the-descender.md` § 8): here the reasons\ncame from a machine, with no face, and still moved belief. This is one study on conspiracy beliefs, not\non sacred values, where § 9 of that note finds argument the wrong instrument, so it refines I7 without\nreplacing it.\n\n**Proposals.** AN3 (*a lens is made of holders' words*): no worldview lens, mind-meld or attunement\nsurface is rendered by a model playing a holder. It is built from claims holders accept and testimony\nthey give. AN4 (*reasons that meet yours*): record the Costello result against I7 as a candidate\nrefinement, and treat a descent that answers a person's own stated reason as the case the graph is built\nfor.\n\n## 6. The refusal: the machine never says why a person believes something\n\nEvery rule above lets people make causal claims about belief, as claims with questions. None lets the\nmind of Deliberus make one in public. A pass that inferred *this author holds P because of their identity* from the\nauthor's other positions would be Bulverism automated. It would also be the person-scoring the corpus\nkilled on empirical grounds (E7): the trait it would score does not measure reliably, and inferring a\nmotive from a position is an ad hominem with a statistical finish.\n\nHeadline tests, as the corpus requires. **Over-broad**: the rule must not forbid extraction, because an\nauthor who writes *they say this because they are paid* has made a claim the graph must record. So the\nrule names a use, not a noun: the machine never *asserts* why a person believes something, and it may\n*record* that a person said so. **Stranger**: *the machine never says why a person believes something*\nreads plainly. **Collision**: *why* here means cause, and the corpus uses *reason* for the other sense.\nThat collision is real, so the body keeps the two words apart.\n\n**Proposal.** AN5 (*the machine never says why you believe it*): no machine pass asserts the cause of a\nperson's or a group's belief. Causal claims about belief enter only as attributed claims under the\n`bias` scheme.\n\n**Ruled the same evening, and scoped by his reply** (ledger AN5, AN10, AN13). His ruling: *\"AN5 and AN10\nruled! We'll recalibrate these to be trued up when we get to mid-November.\"* His reply to the reason given\nfor ranking AN5 first (*a published motive guess cannot be withdrawn*) scopes it to what Deliberus shows\npublicly, and widens what the mind of Deliberus should be able to do:\n\n> Jag är inte alls nervös/neurotisk om att publicera mina takes om varför andra tror på nåt, personligen.\n> Men det är klart att det är en annan sak om det skulle stå på Deliberus publikt. Dock är allt detta -\n> alla möjliga orsaker till att tro på något - förstås inte alls irrelevant in the grand scheme of things.\n> Vi ska inte vara alltför naiva och ovetandes om alla dessa orsaker och känslo/skuggmotivationer. The mind\n> of Deliberus can and should be as capable of autistic, careful, meticulous deliberation and logical\n> scaffolding as it is of sly, savvy, spiritual vibe-checks and balances.\n\nHe went on to propose a separate pass that stores an informed guess at belief causes in a mostly hidden\nmetadata or telemetry layer, with as much of the source wording as possible and some background on the\nauthor and the reported speaker, built in mid-November (AN13, his words whole on the row). Three points\nfor that build, the session's: the separate pass he names is what blind passes (AN9) need, since it is the\none pass that sees identity, so its output must never reach a pass that judges a claim; a guess that a\nnamed person believes something because of their politics or religion is special-category personal data\nunder the GDPR (Art. 9) even when inferred (CJEU C-184/20, 2022) and even as a probable estimate (Austrian Federal Administrative Court W258 2217446-1, 2020, on Austrian Post's party-affinity scores, upheld by the Supreme Administrative Court with a 13 million euro fine; search-grade), hidden or not. Art. 9 forbids it unless an exception applies, so it is a design condition on AN13, not a prohibition of it: the journalism exemption needs a purpose of spreading to the public under editorial review (CJEU C-199/24, 9 July 2026, on a Swedish database), which a hidden layer lacks by design while a public layer under editorial review could meet it (his route the same evening, *\"'priset' är ju exakt vad Deliberus handlar om\"*; press-ethics mapping in TODO.md § MID-NOVEMBER), and 'manifestly made public by the data subject' covers what the person said, not a conclusion about why. Three routes stay open: a guess attached to a position rather than a person (no personal data); a guess about a user who explicitly consents, such as about themselves; and causes that are not special-category (an economic interest, a role), which are ordinary personal data under Art. 6; and the error to expect from a model reasoning about a person is\nstereotyping (Dash et al. 2025), so the guess carries the object it rests on (AM1). The section above\nstands for what is public; this note's first draft had stated it as a limit on the mind of Deliberus as\na whole, which his reply corrects.\n\n## 7. A cause as a pattern: priors, standards, and where scrutiny goes\n\nThe founder's Englander post (ledger E21) holds that disagreement is often not bias at all. The research\nbacks this and gives the line between the two a measurable form.\n\n- **Different priors can be rational.** Dorst (*Rational Polarization*, Philosophical Review 2023)\n  argues that predictable polarization is rational exactly when evidence is ambiguous, and that the\n  mechanism is selective scrutiny: people search for flaws in different places. Tappin, Pennycook and\n  Rand (2020) show that the standard experimental designs for \"politically motivated reasoning\" often\n  cannot separate motive from prior belief, which is confounded with partisanship.\n- **Different standards are the motivated part.** Kahan, Peters, Dawson and Slovic's design keeps the\n  numbers in a table and changes only the label (a skin cream, then a gun ban): the more numerate became\n  more polarized, not less. Gampa et al. (SPPS 2019): people judged the logical validity of political\n  syllogisms by whether they liked the conclusion, and logic training did not remove it. Ditto et al.\n  (Perspectives 2018, 51 studies, about 18,000 participants): this partisan bias is the same size on the\n  left and the right.\n- **Part of measured disagreement is not belief.** Bullock, Gerber, Hill and Huber (QJPS 2015): small\n  payments for correct answers shrank partisan gaps on factual questions substantially. Some of what\n  surveys record is cheering. Van Bavel and Pereira (TICS 2018) model it as identity goals competing\n  with accuracy goals.\n\n**The graph can tell priors from standards, without accusing anyone.** Priors belong in the lens: the\nevidential-priors dial treats them as base-rate claims a lens's holders accept, and a conclusion that\nflips between lenses shows where two worldviews part. Standards show as asymmetry, and two instruments\ncan read it.\n\n1. **Scrutiny symmetry, per debate.** Dorst's mechanism at the scale of a debate: count which side's\n   claims had their critical questions raised and answered. A debate where one side's claims carry every\n   challenge and the other side's carry none has been scrutinised selectively, whoever did it. This is a\n   property of the graph, read like the completeness oracle, and it names no person.\n2. **Standards symmetry, per population.** The belief-bias measurement already proposed (ledger G4):\n   compare rigor ratings on arguments of the same scheme whose conclusions point opposite ways. This is\n   Kahan's flip design run on the graph's own data. It is read across raters as a population. The\n   per-person reading stays forbidden as a score (E7). Whether a person may see their own reading\n   privately is a separate question for the founder.\n\n**Consider the opposite, as a move.** Lord, Lepper and Preston (JPSP 1984): instructing people to be\nunbiased did not reduce biased assimilation of evidence. Asking them to consider how they would judge the\nsame study with the opposite result did. In the graph this is not an instruction. When a person raises a\ncritical question on a claim, the surface can show the same question, still unanswered, on the strongest\nopposing claim of the same scheme. That is a gap with a move offered, which the corpus requires of every\ngap it shows, and it is *consider the opposite* built into the structure rather than asked of the mind.\n\n**The honesty this gives a precise sense.** The founder called the mind of Deliberus *\"more honest … than\nmost of us\"*. In the sense the literature can measure, it is honest by construction: a scheme's critical\nquestions are templates that do not know which side a claim serves. People apply different standards by\nconclusion (Ditto), persona-prompted models do too (Dash), templates cannot. The fourth bias question is\nan instance: it asks the accuser the same question whichever side the accuser is on.\n\n**Proposals.** AN6 (*priors in the lens, standards in the symmetry*): disagreement from different priors\nis represented in the lens and treated as legitimate. Disagreement from different standards is what the\nsymmetry instruments read. AN7 (*where scrutiny went*): a per-debate count of critical questions raised\nand answered on each side, shown beside the debate. AN8 (*the same question on the other side*): when a\nperson raises a critical question, offer the same question on the strongest opposing claim of that\nscheme.\n\n## 8. Three minds, three ways to reason by motive\n\nThe three-minds reflection says humans and models are both motivated reasoners and the mind of Deliberus\ncan be something else. Each mind has its own form of the failure, and each has its own guard.\n\n**The human.** Identity goals against accuracy goals (Van Bavel and Pereira). The guard is structure,\nnever exhortation (the harness conviction's law, and Lord et al. on why the instruction fails). A further\nconsequence: an agreement count is partly cheering (Bullock et al.), which strengthens the existing stance\nthat votes are instrument readings, not verdicts.\n\n**The model.** Two forms. Sycophancy: preference training rewards answers that match the user's stated\nview, and models show it consistently (Sharma et al., ICLR 2024). Persona-induced motivated reasoning\n(Dash et al.). The guard is blindness: a pass that judges a claim does not see whose claim it is, which\nside the asker holds, or which side the founder holds. The keeper's present inputs are stranger-rewritten\nsentences, which strip the speaker. AM9, ruled today, hands the keeper each claim's quoted passage, to\nrestore the language game a sense depends on, and his ruling already names the risk that the pass anchors\non the passages. A passage can also restore the side. The two aims need not collide: hand the passage,\nstrip the name and affiliation. That belongs to AM9's build, which is owed, and the nine sample pairs\nread before and after can measure it with the anchoring he named.\n\n**The mind of Deliberus.** It has no motives, but it has a selection: what gets ingested, which passes\nrun, who contributes, and what grounds its passes. Its version of groupthink is correlated input (ledger\nL3: a thousand readers from one community are one frame sampled a thousand times) and grounding. AM8 now\ngrounds the keeper in rulings, and the rendered block holds only method rulings: how sameness is judged,\nthe gate, separability, the object a verdict needs, the smallest step, the two factors, uses as votes.\nThat is safe. If substantive positions entered the grounding block (what the founder concluded about a\ncontested question), every pass would inherit one person's priors, and the mind of Deliberus would become\na groupthink of one, crystallised. The founder named the risk himself: *\"Same issue goes for humans I\nguess.\"*\n\nOne more result bears on all three. Costello, Pelrine, Kowal and colleagues (arXiv 2601.05050, 2026;\n3,996 participants): models instructed to argue for a conspiracy raised belief about as well as models\ninstructed to argue against one lowered it, with no consistent truth advantage, and participants rated\nthe misleading model as more informative. When the model was told to use only accurate information, its\npower to mislead fell dramatically. A fluent mind has no built-in advantage for the truth. The mind of\nDeliberus does not get one from being clever. It gets one from its constraints: every claim checkable,\nboth sides on the page, every attack visible. That is the strongest outside argument yet for why the\nsystem declines to be a persuader. Dan Williams's *marketplace of rationalizations* (Economics and\nPhilosophy 2023) names the matching hazard: where many people want a belief, suppliers compete to provide\nits justification. A rich graph is also a rationalization supply. Inside the graph a rationalization\nnever travels without its defeaters. Copied out of the graph, it can. The corpus holds the same hazard\nunder Kahan's *better weapons factory*.\n\n**Proposals.** AN9 (*blind passes*): no machine pass that judges a claim is given the identity or side of\nits holder, the asker, or the founder. AM9's passages go in with names stripped. AN10 (*method grounds a\npass, substance never*): the grounding blocks of AM8 carry rulings about how to judge, never the founder's\nconclusions on contested questions.\n\n## 9. Groupthink is a word, and the mechanisms are what is real\n\nJanis's groupthink model has weak empirical support. Reviews over three decades (Aldag and Fuller 1993,\nFuller and Aldag 1998, Turner and Pratkanis 1998, Baron 2005) find that the full model is rarely\nsupported, while the symptoms occur far more widely than its antecedent conditions predict. What survives\nis what the corpus already uses: cascades, polarization of like-minded groups, hidden profiles, and\ncorrelated error. So when a contributor writes *groupthink*, it is a use of a folk word (AM7: a use is a\nvote on the meaning), usually as an accusation (§ 4). The graph translates it into the mechanism the\nwriter means, and does not build a groupthink detector for people.\n\n## 10. Where two minds part: an anatomy for the mind-meld\n\nThe founder's mind-meld shows each mind what the world looks like from inside the other. The corpus's\nlens page lists four things that change between lenses: salience, the weight of evidence, definitions,\nand conclusions. This note's research adds two places minds part that the graph can show, which makes six.\n\n| Where they part | What the graph shows | Already built or proposed |\n|---|---|---|\n| A fact | the crux claim and its evidence | built (descent, cruxes) |\n| A prior | the base-rate claims each side accepts | proposed (evidential-priors dial) |\n| A sense | the word pair a disagreement turns on | ruled, partly built (concept layer, AM7) |\n| A weighting | which considerations each side lets count | designed (value weighting) |\n| A standard | the same evidence judged by its conclusion | new here: the symmetry instruments, read across a population |\n| A stake | what the position protects for its holder | designed (impact testimony, voice on the node); new here: from holders only |\n\nThe first four are reasons. The last two are causes, entered by the rules above. A mind-meld that shows\nonly the first four shows two arguments. With the last two it shows two people. Proposal AN11 (*six places\ntwo minds part*): adopt this table as the frame for what a mind-meld surface shows.\n\n## 11. The communication half of attunement\n\nThree results bear on how people talk, not on what they hold. Yeomans, Minson and colleagues (OBHDP 2020)\nbuilt a model of *conversational receptiveness* (acknowledgement, hedging, subjectivity): receptive\nwriters were rated better collaborators and drew fewer personal attacks, and receptiveness tends to be\nreturned. Argyle and colleagues (PNAS 2023): a chat assistant that offered rephrasings improved\nconversation quality and democratic reciprocity without moving anyone's policy view. Hruschka and Appel\n(2026, accepted at JCMC): counterarguing chatbots with higher receptiveness reduced affective polarization\nmore. Two cautions: most such effects fade within a month (Walter 2025; Kim et al. 2026 preprint), and the\nBraver Angels randomized trial (Baron et al. 2021) found workshop effects that partly dissipated by six\nmonths.\n\nThis fits the corpus's guard that understanding is not acceptance, and AM6's rule that a hedge is stance,\nnot strength. Proposal AN12 (*a receptive rewording, offered*): when a person writes an attack, the\nsurface may offer a receptive rewording of the same content. The attack's content and strength do not\nchange. Low priority, and a candidate for the workshop rather than the product.\n\n## 12. What this changes, and what to rule first\n\nNothing new is built by this note. In order of what the rest depends on:\n\n1. **AN5, the machine never says why you believe it.** Ruled 2026-10-19 for what Deliberus shows\n   publicly; recalibration due mid-November, with his hidden-guess proposal (AN13, § 6).\n2. **AN10, method grounds a pass, substance never.** Ruled 2026-10-19; recalibration due mid-November.\n   It decides the lists for the other passes, which AM8 left open.\n3. **AN3 and AN9, holders' words and blind passes.** Both follow from the same two results (Eyal; Dash).\n4. **AN1 and AN2, Lewis's order for every causal attack.** AN2 is a test before anything is built.\n5. **AN6 to AN8, the symmetry instruments.** They make motivated reasoning visible in the one form the\n   graph can show without accusing anyone.\n6. **AN4, AN11, AN12.** Refinements and frames.\n\n## 13. What the outside reader found (2026-10-19)\n\nThe ledger's outside reader compared the twelve rows with what is settled. Ten carry its line on the\nledger, two reach him as written (AN6, AN11). Three findings change how the proposals must be read.\n\n- **AN7 against Q3.** The reader read *a per-debate count of critical questions* as reversing Q3, *weigh\n  critical-question answers, never count them*. AN7 means a coverage reading shown beside the debate,\n  never an input to strength, which is the reading Q3 leaves allowed. The row did not say so, so the\n  question to him is whether that scope is enough to keep AN7 clear of Q3.\n- **AN12 against the rule that a claim's text is written once.** A receptive rewording of an attack that\n  is already stored would be a rephrase, and a rephrase is a new claim. AN12 means an offer to the author\n  before the attack is posted, so the author's claim is written once in the wording the author chooses.\n- **AN9 against AM8.** The grounding block names the founder as the author of its rulings. AN9 is about\n  the claim being judged: its holder, the asker, and which side the founder holds on that claim. The\n  method rulings in a grounding block are not a side, and AN10 keeps them that way.\n\nThe reader could not tell, among other things, whether AN11's *standard* and *weighting* are the same\nkinds of difference as the five the corpus already names for a word (AC2: covering more, covering less,\nanother standard, another threshold, another weighting). They are related: AC2 names differences between\nsenses of a word, AN11 names places two people part across a whole disagreement. The record of what the\nreader found is `outside-checks/records.jsonl`.\n\n## Sources\n\nRead-depth: search-grade unless marked.\n\n- Lewis, C. S., *Bulverism* (1941) · [Wikipedia](https://en.wikipedia.org/wiki/Bulverism)\n- Costello, Pennycook & Rand, *Durably reducing conspiracy beliefs through dialogues with AI*, Science 2024 (DOI 10.1126/science.adq1814) · [MIT IDE](https://ide.mit.edu/insights/debunker-chatbot-makes-a-splash-spurs-debate) · [Nieman Lab on the facts-only follow-up](https://niemanlab.org/wrap/an-ai-chatbot-helped-americans-who-believe-in-conspiracy-theories-exit-the-rabbit-hole)\n- Costello, Pelrine, Kowal, Timm, Arechar, Godbout, Gleave, Rand & Pennycook, *Large language models can effectively convince people to believe conspiracies*, [arXiv 2601.05050](https://arxiv.org/abs/2601.05050) (abstract read on the page)\n- Dash, Reymond, Spiro & Caliskan, *Persona-Assigned Large Language Models Exhibit Human-Like Motivated Reasoning*, [arXiv 2506.20020](https://arxiv.org/abs/2506.20020) (abstract read on the page)\n- Sharma et al., *Towards Understanding Sycophancy in Language Models*, [ICLR 2024](https://arxiv.org/abs/2310.13548v1)\n- Eyal, Steffel & Epley, *Perspective mistaking*, JPSP 2018 · [summary](https://behavioralscientist.org/be-mindwise-perspective-taking-vs-perspective-getting/)\n- Dorst, *Rational Polarization*, Philosophical Review 2023 · [PhilArchive](https://philarchive.org/rec/DORRPM-3) · [MIT News](https://news.mit.edu/2023/how-do-reasonable-people-disagree-1120)\n- Tappin, Pennycook & Rand, *Thinking clearly about causal inferences of politically motivated reasoning*, COBEHA 2020 · [MIT DSpace](https://dspace.mit.edu/handle/1721.1/135504?show=full)\n- Kahan, Peters, Dawson & Slovic, *Motivated numeracy and enlightened self-government* · [record](https://hdl.handle.net/1794/18962)\n- Gampa et al., *(Ideo)Logical Reasoning: Ideology Impairs Sound Reasoning*, SPPS 2019 · [Sage](https://journals.sagepub.com/doi/10.1177/1948550619829059)\n- Ditto et al., *At Least Bias Is Bipartisan*, Perspectives on Psychological Science 2018 · [PDF](https://sites.uci.edu/peterdittolab/files/2018/06/Ditto-et-al.-2018-At-Least-Bias-is-Bipartisan-A-Meta-Analytic-Comparison-of-Partisan-Bias-in-Liberals-Conservatives.pdf)\n- Bullock, Gerber, Hill & Huber, *Partisan Bias in Factual Beliefs about Politics*, QJPS 2015 · [NBER](https://www.nber.org/papers/w19080.pdf)\n- Van Bavel & Pereira, *The Partisan Brain*, TICS 2018 · [PDF](https://pages.ucsd.edu/~mboyle/COGS2/readings/The%20Partisan%20Brain%20An%20Identity-Based%20Model%20of%20Political%20Belief.pdf)\n- Nguyen, *Echo chambers and epistemic bubbles*, Episteme · [PDF](https://cssh.northeastern.edu/pandemic-teaching-initiative/wp-content/uploads/sites/43/2020/10/echo_chambers_and_epistemic_bubbles.pdf)\n- Begby, *Evidential Preemption*, PPR 2021 · [SFU](https://www.sfu.ca/research/research-impact/performance-excellence/scholarly-impact-of-the-week/scholarly-impact-stories/2021/03/evidential-preemption---how-to--groom-ing--an-audience-for-misin.html) · Butzer's reply, [Disputatio 2024](https://reference-global.com/article/10.2478/disp-2024-0002)\n- Williams, *The marketplace of rationalizations*, Economics and Philosophy 39(1) 2023 · [RePEc](https://ideas.repec.org/a/cup/ecnphi/v39y2023i1p99-123_5.html)\n- Lord, Lepper & Preston, *Considering the opposite*, JPSP 1984 · [CAUSE](https://causeweb.org/cause/node/7638)\n- Yeomans, Minson, Collins, Chen & Gino, *Conversational receptiveness*, OBHDP 2020 · [HKS](https://www.hks.harvard.edu/publications/conversational-receptiveness-improving-engagement-opposing-views)\n- Argyle et al., *Leveraging AI for democratic discourse*, PNAS 2023 · [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC10576030)\n- Walter 2025; Kim et al. 2026; Hruschka & Appel 2026 on chatbot depolarization · [Kim et al., arXiv 2606.15901](https://arxiv.org/abs/2606.15901) · [Hruschka & Appel preprint](https://www.mcm.uni-wuerzburg.de/fileadmin/06110000/2026/Hruschka___Appel_2026__Preprint_.pdf)\n- Baron et al. 2021, the Braver Angels trial · [EGAP brief](https://egap.org/resource/couples-therapy-for-a-divided-america-evaluating-the-effects-of-a-depolarization-workshop/)\n- Groupthink reviews (Aldag & Fuller 1993; Fuller & Aldag 1998; Turner & Pratkanis 1998; Baron 2005) · [Aldag & Fuller](https://web.mit.edu/writing/2012/Grad_Summary_Readings/beyond_fiasco_final.pdf)\n- Dalege et al. 2016 and Turner-Zwinkels & Brandt 2022 on belief networks (not used in the argument; kept for the position fingerprint, where change spreads to the attitudes nearest the one targeted) · [Tilburg](https://research.tilburguniversity.edu/en/publications/belief-system-networks-can-be-used-to-predict-where-to-expect-dyn/)\n\n## Cross-references\n\n- [three-minds-and-the-leap.md](three-minds-and-the-leap.md) (the three minds; AM8 grounding; AM9 passages)\n- [cognitive-bias-codex-and-human-contribution.md](cognitive-bias-codex-and-human-contribution.md) (the bias scheme, E21; belief bias)\n- [../worldview-lenses.md](../worldview-lenses.md), [../analysis-and-attunement.md](../analysis-and-attunement.md) (the lens, the poles)\n- [the-group-premise-evidence.md](the-group-premise-evidence.md), [does-descending-change-the-descender.md](does-descending-change-the-descender.md) (groups, exposure, sacred values)\n- [the-harness-conviction.md](the-harness-conviction.md) (structure over exhortation)\n- Ledger: [session26-open-rulings.md](session26-open-rulings.md) Arc AN\n"}