{"path":"research/starting-values-and-the-problem-of-the-prior.md","content":"# Starting values: the problem of the prior, and how Deliberus should meet it\n\n**Date**: 2026-09-30 · **Type**: research (academic literature, forecasting practice, prediction-market\nand platform design) plus a synthesis for Deliberus · **State**: findings and recommendations,\n**nothing ruled**. It answers ledger Q35 and the founder's questions below; the recommendations in § 7\nawait his rulings.\n\n**The question, verbatim** (founder, 2026-09-30, after the first calculus choice was ruled): *\"Ok I get it,\nof course the prior is PRIOR to evaluating evidence. Or in probability theory I guess the prior is prior\nto evaluating NEW evidence? It assumes you've already eaten up some evidence prior to seeing the new\nevidence, right? What's the starting value/prior expectation given an \"entirely fresh\" claim? Or is there\n\"no such thing\"? Hmmm...\"* Then, commissioning this: *\"Research this extensively online. Both academic\npublications and whether any super-forecasting articles or prediction market blogs discuss this in depth\nand how they handle it in integrating epistemic calculations/estimations in software that people use etc.\nAnd think deeply about how all this would/could/should integrate in Deliberus, especially given what is\nunique about Deliberus in the grand scheme of things, the recursive decomposition, evidence-as-subgraph,\nQEM/QBAF, argumentation theory/mapping as foundational to the ontology vs traditional probability/Bayes\ntheory, etc etc etc. Go deep.\"*\n\nWhere this sits: `relevance-and-the-two-factors.md` § 7 (the first calculus ruling, his challenge, and\nthe starting-value path) and ledger Q34–Q35.\n\n**In pictures** (2026-10-03): the page *Where a claim starts*\n(claude.ai/artifact/8eyM83VEoFmJBKutpxqmcj, private) draws §§ 6–10 as one measured example and five\nopen questions, with the rulings of § 9 in dependency order. What it shows and how it was made: § 11.\n\n---\n\n## 0. The answer first\n\n1. **A prior is always prior to some particular evidence, never to all evidence.** Yesterday's result\n   is today's prior. Every working field treats it that way, and every probability anyone can state is\n   relative to a reference class (Hájek 2007: probability is fundamentally a two-place notion).\n2. **An \"entirely fresh\" claim does not exist in practice.** A claim always arrives with a kind (a\n   measurement, a report, a forecast, a value judgement) and an origin (who said it, by what method).\n   Kind and origin already carry information. The search for a prior that owes nothing to any evidence\n   is, in the Stanford Encyclopedia's summary, largely abandoned; indifference (50/50) survives only for\n   a stated set of rival options, and breaks when the options are carved differently.\n3. **Everyone who built software on evidence-to-number reasoning hit this, and the ones that work all\n   converged on one move: make the starting value a separate, visible, contestable object, taken from a\n   reference class, and keep it apart from the evidence.** GRADE starts a body of evidence at a level\n   set by study design. Superforecasters anchor on the outside view, then adjust. Clinicians start from\n   prevalence, then apply likelihood ratios. Subjective logic carries a base rate separately from belief,\n   disbelief and uncertainty. Manifold makes the market creator set the starting price and fund how hard\n   it is to move. The systems that struggled hid the prior, fixed it by a domain-blind formula, or let\n   extreme made-up numbers through (Rootclaim; Laplace's 50% first trial; LMSR's liquidity \"black art\").\n4. **Deliberus already has most of the right shape, and one wrong constant.** QEM has a prior slot\n   (`qem_strength`'s base weight: *\"gravity toward initial weight\"*), the Keynes split already refuses to\n   put a number on an unexamined claim, and links already get presumptions by who wrote them (relevance\n   0.5 or 0; requirement weight by provenance). The wrong constant is that **every claim starts at 0.5**,\n   whatever kind of claim it is.\n5. **The Deliberus-native answer: the starting value is a claim too, and it comes from what kind of claim\n   it is.** Decomposition pushes the prior problem down to the leaves, where claims are simplest and base\n   rates exist. A leaf's start comes from its kind, through a reference-class claim whose own value is\n   computed from its own subgraph (*hypotheses tested in top psychology journals start near one in ten*).\n   **Who asserted it is not part of its start but its first evidence**: the report is not the event. So\n   the founder's intuition becomes literally true: the prior is computed in Deliberus, from the reference\n   class's evidence, and never from the claim's own, which would count it twice.\n6. **Honest starts explain the first calculus ruling rather than retire it.** The live link rule is\n   already Jeffrey's rule with every start at 0.5. With honest starts, a failed reason costs its conclusion\n   its start times its relevance, **but only where the conclusion relied on it**: a reason attached from\n   outside may have been chosen because it fails, and a refutation that was bound to be found says almost\n   nothing about the conclusion. That is what the ruled *contribution that says why* captures, so\n   drop-out stays right wherever reliance is unknown (§ 6.8).\n7. **Worldviews differ largely in their starts.** A conclusion that holds under every reasonable start\n   is robust; one that flips across them is prior-dependent. That is a legal standard (beyond reasonable\n   doubt read as *doubt ruled out from every reasonable prior*, Abrams 2025), a measurable instrument for\n   the convergence wager, and a principled home for the worldview lens.\n8. **How several kinds combine has a 2026 answer: pool them, never pick one.** A claim of several nested\n   kinds gets one start, the narrow kind's rate pulled toward the broad kind's by how little data it has\n   (§ 10.3). Learned models move the remaining choice, which kinds count, rather than remove it, and judge\n   it by checking against outcomes (§ 10.4). LawZero's Scientist AI keeps its prior on whole theories and\n   treats an assertion as evidence about a hidden truth, the same split as § 6.2 (§ 10.5).\n\n---\n\n## 1. What the corpus already had (do not re-derive)\n\n- **The founder's recorded positions** (ledger P2, P3, P4, open): *\"if there's inductive evidence for any\n  claim, that's what determines the probability assessment, we try to break down into smaller parts\n  everything these probability assessments are based on and crowdsource+AI-source the evidence and\n  premises and CQ questions/answers or whatever that this depends on. So that was in my mind not different\n  for different people ideally. But of course maybe it is irreducibly subjective\"* (P2); *\"I have a fear of\n  letting fuzzy gooey subjectivity in to affect the assessments in Deliberus without proper decomposition\n  and deep analysis and evidence/cross checking etc etc.\"* (P3); the methodology ladder (meta-analysis 0.70\n  … expert opinion 0.30) as base weights, invented as a starting proposal, never ruled, read by no badge (P4).\n  **This document's central recommendation is P2 and P3 applied to the prior itself**: a start that is not\n  decomposed and argued is exactly the fuzzy subjectivity P3 fears.\n- **`what-counts-as-evidence/philosophy-of-evidence.md` § 4.3–4.4**: uniqueness versus permissivism;\n  objective Bayesianism's partition problem; merging-of-opinions theorems; and the platform principle that\n  *every divergence from shared evidence must be attributable to a named slot*: a prior, a standard, a\n  model, or a piece of evidence. It also says the platform should display the evidence term and never a\n  posterior, *which would import a prior the platform has no right to choose*. § 6 below refines that rule\n  rather than overturning it.\n- **`what-counts-as-evidence/formal-and-computational-models-of-evidence.md` § 2.3–2.4**: subjective logic's\n  base rate as *what \"no evidence\" should default to per claim kind*, and Walley's imprecise probabilities.\n- **`deliberus/evidential_weight.py`**: weight and direction kept apart after Keynes; an unexamined claim\n  carries no number (*\"an absence that is not a number\"*).\n- **`strength-calculus-literature/bayesian-and-informal-logic.md` § 2.1**: the live link rule is Jeffrey\n  conditioning, moving a conclusion by Δ·(q − q0) where q0 is the reason's reference value, set to 0.5\n  today; the straw-man risk sits in the presumptions (relevance 0.5, q0 = 0.5), not in the sign rule;\n  Potyka's MLP semantics has the log-odds form.\n- **`what-counts-as-evidence/formal-and-computational-models-of-evidence.md` § 3.1**: Schum, *the report\n  is not the event*: an assertion is evidence about a proposition, weighted by its source's credibility.\n  § 6.2 below turns on it.\n- **`relevance-and-the-two-factors.md` § 6–7**: *Standing has to originate somewhere*; the first calculus\n  ruling; the founder's challenge; the starting-value path, **stated there as starts \"by kind and origin\",\n  which § 6.2 refines**. Its *Considered and rejected* list includes subtraction only for reasons the\n  target's own source asserted; § 6.8 bears on that option and § 9 puts it back to the founder as a\n  question rather than reopening it.\n- **Presumption by provenance, already built for links**: `relevance.PRESUMED_RELEVANCE` (0.5 for the\n  source, a person or a split; 0 for a machine's guess) and `support_semantics.presumed_requirement_weight`.\n  The pattern this document proposes for claims already exists one level over.\n\n---\n\n## 2. The academic answer\n\n### 2.1 There is no evidence-free prior\n\nThe probability axioms constrain a hypothesis only to lie between 0 and 1, which leaves the starting\nvalue open. That openness is the *problem of the priors* (SEP, *Formal Epistemology* § 1.4; *Bayesian\nEpistemology* § 1.5 and § 4). Three camps:\n\n- **Subjective Bayesians** allow any coherent prior; rationality lives in updating, not in starting.\n- **Objective Bayesians** add a norm against unsupported commitment: the principle of indifference, or\n  its generalisation, maximum entropy (Jaynes 1968). Williamson (*In Defence of Objective Bayesianism*,\n  2010) gives three norms: degrees of belief should be **probabilities**, **calibrated** to known\n  frequencies and chances, and otherwise **equivocal**. His worked example: knowing only that a cancer\n  recurs in 10–40% of such cases, calibrate to that range and equivocate within it, giving 0.4.\n- **The standing objection** to indifference is partition dependence (Bertrand, Keynes): carve the\n  options differently and the \"neutral\" numbers change. Williamson's own reply is that the partition\n  reflects the language one brings, which is knowledge rather than neutrality.\n\nThe SEP's summary of the history is blunt: the quest for an a priori justification of the prior has been\nlargely abandoned (*The Problem of Induction* § 3.3). What remains is the view that priors reflect\nbackground knowledge, which is where empirical assumptions enter every inductive inference.\n\n### 2.2 Probability is two-place: the reference class problem is everyone's\n\nHájek (\"The reference class problem is your problem too\", *Synthese* 156, 2007) shows that the reference\nclass problem, usually pinned on frequentism, afflicts the classical, logical, propensity and subjectivist\ninterpretations too. His dissolution of the metaphysical half: probability is fundamentally conditional,\nrelative to a class, the way simultaneity is relative to a frame. He concedes that the **epistemological**\nhalf remains: which class to use is still a judgement. Nothing in this document makes that judgement\ndisappear; the recommendation is to make it visible and arguable.\n\n### 2.3 Balance, weight and specificity; imprecise priors and their price\n\nJoyce (\"How probabilities reflect evidence\", *Philosophical Perspectives* 19, 2005) separates three\ndimensions of total evidence: **balance** (which way it points; what a credence reflects), **weight**\n(how much there is; reflected in how stable the credence is), and **specificity** (how sharply it\ndiscriminates; reflected in the spread of a credal set). His conclusion for the fresh case: symmetric\nevidence mandates equal *treatment*, not equal point credences; with no evidence, the honest state is an\ninterval. The price is **belief inertia**: a fully vacuous prior [0, 1] can never be moved by ordinary\nconditioning, an old problem (Levi 1980) that Joyce (2010) accepts for the vacuous case, that Rinard\n(2013) takes to show the vacuous prior is never a legitimate state, and that Vallinder (2018) shows\nsurvives Joyce's proposed remedy; Rinard (2014) adds that narrower common intervals across rival\npartitions become incoherent too.\n\n**For Deliberus**: `evidential_weight` already implements the weight/balance split and the honest\nabsence. The inertia problem does not bite, because Deliberus measures evidence as offsets from a neutral\npoint rather than conditioning a vacuous set; but it warns against ever turning \"unexamined\" into an\ninterval that propagates.\n\n### 2.4 Washing out is asymptotic\n\nMerging-of-opinions theorems (Blackwell & Dubins 1962; Gaifman & Snir 1982) show that agents with\nopen-minded priors converge as shared evidence grows. The SEP's entry on scientific objectivity adds the\ncaveat that matters: these are limit results without direct force for finite datasets, where the prior\nstill matters and may carry bias or values; Sprenger (2018) argues that transparency about priors is the\nobjectivity Bayesian inference can actually deliver.\n\n**For Deliberus**: most claims will stay sparsely examined for a long time, so starts will matter for\nmost of the graph most of the time. Convergence across worldviews is a property to *measure* claim by\nclaim, not assume.\n\n### 2.5 Empirical Bayes: learning the prior from the experience of others\n\nRobbins (1956) showed that a large set of parallel cases can estimate its own prior; James and Stein\n(1961) showed that shrinking each estimate toward the group beats treating each alone; Efron calls the\nidea that a large data set implies its own prior revolutionary (\"Empirical Bayes: Concepts and Methods\",\n2021; *Large-Scale Inference*, 2010). Efron closes on the problem that never goes away: **relevance**,\ndeciding which other cases apply to the case at hand. Efron and Morris's relevance functions let an\nindividual case partially opt out of the group. That is the reference class problem again, in statistical\nclothing.\n\n**For Deliberus**: once the graph holds many claims of a kind whose fate is later settled (replicated,\nretracted, refuted), base rates by kind can be estimated from the graph's own history, and which cases\ncount as the same kind can be argued in the graph.\n\n### 2.6 Measured base rates for claims\n\nThe prior for a scientific claim has been measured, which makes \"starting value by kind\" concrete:\n\n- **Ioannidis** (\"Why most published research findings are false\", *PLOS Medicine* 2005): the probability\n  that a claimed finding is true depends on the field's pre-study odds R, power and bias\n  (PPV = (1 − β)R / (R − βR + α)). A well-powered RCT starting from even pre-study odds (R = 1:1) is true\n  about 85% of the time; well-powered exploratory epidemiology at R = 1:10 about one time in five. His own\n  recommendation: study the range of pre-study odds rather than chase significance.\n- **Replication projects**: psychology, 36% of replications significant and 39% judged to have replicated,\n  with effect sizes halved (Open Science Collaboration, *Science* 2015); experimental economics, 11 of 18\n  (61%) (Camerer et al., *Science* 2016); social science in *Nature* and *Science*, 13 of 21 (62%)\n  (Camerer et al., *Nature Human Behaviour* 2018).\n- **Priors elicited by markets**: prediction markets on the psychology replications predicted 71% of\n  outcomes (29 of 41), and the implied **median prior probability of a hypothesis tested in top psychology\n  journals was about 9%**; one significant finding lifted it to a median of 56%, a successful replication\n  to 98%, a failed one to 6% (Dreber et al., *PNAS* 2015).\n\nDreber's three stages are the structure § 6.2 recommends, measured: a start before any test (the\nhypothesis's kind, 9%), then each study as evidence that moves it (to 56% after a significant finding, to\n98% or 6% after the replication). The rate at which *published* findings hold up (about four in ten) is a\ndifferent quantity: it describes what a significant publication is worth as evidence, which belongs on the\nlink from a study to its claim, not in the claim's start. Both are reference-class claims with evidence, of\nthe kind § 6.1 proposes.\n\n### 2.7 Law: the problem of the prior, and reasonable doubt across priors\n\nLegal Bayesians call it by the same name. Even strong supporters of the Bayesian approach often avoid\npriors and report only the likelihood ratio (Fenton, Lagnado, Dahlman & Neil, \"The opportunity prior\",\n*Law, Probability and Risk* 2019). Their proposal: the defendant's prior is 1/N over the people who had the\nsame or better opportunity (the \"extended crime scene\"), a reference class argued from proven facts.\nDahlman & Kolflaath (2021) argue that 1/N over a whole population conflicts with the presumption of\ninnocence and that a prior fixed by law is less problematic than the alternatives. Laudan (2013) argues\nthat the presumption of innocence is not a prior at all but an instruction about proof. Abrams\n(\"Reasonable doubt and reasonable priors\", *Episteme* 2025) reads *beyond reasonable doubt* as doubt\nruled out **from every reasonable prior**, not merely from one's own; the presumption of innocence then\nrequires the set of reasonable priors to include ones favourable to the defendant.\n\n**For Deliberus**: two ideas transfer directly. The start comes from an argued reference class, and the\nstrongest verdict available is one that survives every reasonable start.\n\n### 2.8 Medicine: starting certainty by design\n\nGRADE rates a body of evidence from a **starting level set by study design**: randomised trials start at\nhigh certainty, non-randomised studies at low (or at high under ROBINS-I, then usually down two levels),\nand then the rating moves down for risk of bias, inconsistency, indirectness, imprecision and publication\nbias, and up (for non-randomised evidence) for large effects, dose-response and confounding that would\nwork against the finding (Cochrane Handbook ch. 14; GRADE Book). Diagnostic reasoning starts from the\npre-test probability, usually the prevalence, and applies likelihood ratios; skipping the base rate is\nthe base rate fallacy, which in court is the prosecutor's fallacy.\n\nGRADE is not a probability calculus, but its shape is the one this document recommends: a stated starting\nlevel, then rated moves, each for a stated reason. Note what its start is a start *of*: the certainty of a\nbody of evidence, set by how that evidence was produced. In Deliberus that is a property of the link from a\nstudy to its claim, which is where the code already applies its methodology tiers (§ 7, item 4).\n\n### 2.9 Argumentation theory: base scores, presumption and audience\n\n- **Base scores in QBAFs.** Every argument in a quantitative bipolar framework has a base score, its\n  intrinsic strength, and gradual semantics move it by its attackers and supporters (Baroni, Rago & Toni\n  2019). Where base scores come from is openly unsettled. Argumentative LLMs (Freedman et al., AAAI 2025)\n  either give the claim a neutral 0.5, so that its final strength comes only from the arguments, or let a\n  model estimate it; the follow-up found plain verbal confidence the most effective estimator (EMNLP\n  2025). Social abstract argumentation (Leite & Martins, IJCAI 2011) derives base scores from votes\n  and treats them as an **upper bound** on final strength. The corpus's computational review adds that\n  retrieval-based base scores did worse than a flat prior in ArgRAG.\n- **0.5 as undecided.** In the epistemic approach to probabilistic argumentation (Hunter & Thimm, *JAIR*\n  2017), P(A) = 0.5 means neither believed nor disbelieved, the \"undecided\" label; epistemic graphs (Hunter,\n  Polberg & Potyka, KR 2018) let different agents carry different constraints on the same graph, which is\n  perspective made formal.\n- **Presumption: the argumentation-native prior.** Freeman (*Acceptable Premises*, 2005) makes a premise\n  acceptable when there is a presumption in its favour, and ties presumption to the belief-generating\n  mechanism behind the claim (perception, memory, testimony, expert authority, intuition, moral sense),\n  under a pragmatic condition: if mistakenly accepting it costs more than getting more evidence, the\n  presumption lapses. His motto is *consider the source*. Carneades (in the corpus) allocates presumption\n  per question by burden.\n- **Audience relativity.** Hahn & Oaksford (*Psychological Review* 2007) give a Bayesian theory of\n  argument strength in which the *force* of an argument (how far it moves belief, a likelihood-ratio\n  quantity) is separate from the *conviction* it produces (the posterior, which depends on the prior). They\n  read prior-dependence as the audience-relativity that argumentation theory has always required\n  (Perelman). Their analysis of the argument from ignorance makes its strength depend on the prior and on\n  the sensitivity of the search, which is the corpus's third home for a failed case (ruling Q34).\n- **Subjective logic.** An opinion is (belief, disbelief, uncertainty, base rate), with the base rate\n  defined as the category probability unconditioned on evidence and the projected probability as\n  belief + base rate × uncertainty (Jøsang 2016). With no evidence, the projection is just the base rate;\n  as evidence accumulates, the base rate washes out. Jøsang notes that ordinary belief functions silently\n  default a vacuous belief to 0.5 by relative atomicity, where practice should use statistics, and that\n  base rates are what keep conditional reasoning clear of the base-rate fallacy. Deduction takes exactly\n  the two conditionals the relevance work keeps returning to, the opinion on y given x and given not-x.\n  Oren, Norman & Preece (\"Subjective logic and arguing with evidence\", *Artificial Intelligence* 171, 2007)\n  build an argumentation framework on it, with schemes, accrual, burden of proof, and **sensors** as\n  sources of evidence: sources as nodes, in 2007.\n\n---\n\n## 3. Practice: forecasting, markets and the software people use\n\n### 3.1 Superforecasting: the outside view first\n\nTetlock and Gardner's third commandment is to balance inside and outside views, asking how often things\nof this sort happen in situations of this sort, and to anchor on the outside view before adjusting. The\nGood Judgment Project's training (CHAMPS KNOW, under an hour) improved accuracy by 6–11% in Brier score\nover control across four years, and of its principles **comparison classes** were the one most associated\nwith better accuracy (Chang et al., *Judgment and Decision Making* 2016). A text analysis of forecasters'\nrationales found the best forecasters invoke comparison classes more often (Karvetski et al.,\n*International Journal of Forecasting*). The process AI Impacts distils from the book: break a question into\nsub-questions (\"Fermi-ize\"), take the outside view on the sub-questions, then aggregate across several\nreference classes and models.\n\n### 3.2 Semi-informative priors: choosing the start in the open\n\nDavidson's report for Open Philanthropy (\"Semi-informative priors over AI timelines\", 2021) is the most\ncareful public treatment of a fresh-question prior. Laplace's rule gives the first trial a 50% chance\nhowever \"trial\" is defined, which he rejects; he generalises it to a beta prior set by a **first-trial\nprobability** chosen from several reference classes (his favoured range 1/1000 to 1/100), a number of\n**virtual successes** that sets how fast the prior yields to evidence, a regime start and a trial\ndefinition, then weights several such rules in a hyper-prior that is itself updated. He is explicit that\nthe inputs rest on subjective judgement calls and that the method deliberately ignores inside-view\nevidence. Two lessons transfer: a start needs a **strength** as well as a value, and competing reference\nclasses can be carried side by side with weights rather than collapsed early.\n\n### 3.3 Conditional trees: value of information is a prior-weighted quantity\n\nThe Forecasting Research Institute's conditional trees (McCaslin, Rosenberg, Karger et al., 2024) ask\nexperts for cruxes that would move their view on an ultimate question, and rank them by **value of\ninformation**: the expected update, which weights how much a crux would move the answer by how likely the\ncrux is. Crux questions generated this way were on average nine times more informative than the most\npopular platform questions on the same topic. Their existential-risk tournament (Karger et al., 2023)\nfound minimal convergence between experts and superforecasters after months of exchange: persistent\ndisagreement that sits, at least partly, in priors.\n\n**For Deliberus**: the hinge score is Deliberus's native value-of-information measure. It currently prices\na question by the change it would make, without the question's own probability; with honest starts it\ncould weight by the chance the answer goes each way, as FRI's VOI does.\n\n### 3.4 How platforms set the start\n\n- **Metaculus** has no platform prior: a question's aggregate is the recency-weighted median of its\n  forecasters. It **hides the community prediction when a question opens**, so that the first forecasts do\n  not ground the later ones, and its guidance to forecasters is to start from historical frequencies. Its\n  performance-weighted \"Metaculus Prediction\" was deprecated in November 2024.\n- **Manifold** makes the market creator choose the initial probability and pay for the liquidity pool; its\n  market maker (Maniswap, a constant-function market maker parametrised by the initial probability) lets\n  the whole subsidy become liquidity at any starting price. The more liquidity, the harder the price is to\n  move. The creator's number is the prior; the subsidy is the prior's strength.\n- **Hanson's logarithmic market scoring rule** starts at uniform prices, and one parameter sets the loss\n  bound, the liquidity and how fast the price adapts, a choice the market-design literature calls something\n  of a black art, which is why Othman, Pennock, Reeves and Sandholm (2010) built a market maker that raises\n  it with trading volume. The replication markets of § 2.6 ran on this rule. Hanson's own analysis: a sponsor minimises expected payments by\n  starting the market at her own prior, and the worst-case cost equals the entropy of that starting\n  distribution. A better start makes information cheaper to buy.\n\n### 3.5 Rootclaim: what explicit Bayes with made-up numbers does\n\nRootclaim scores hypotheses by a prior times one likelihood ratio per piece of evidence. In its $100,000\nCOVID-origins debate (2024), the judges found for zoonosis; Scott Alexander collected six participants'\nand observers' Bayesian tallies and found them spread across twenty-three orders of magnitude, and still\na factor of about fifty among the neutral ones. The failure modes he names: a few extreme likelihoods\ndominate, independence between stages is assumed where correlation is likely (the multiple-stage fallacy),\nand the parties disagreed about the base rates themselves. His fairest line for Rootclaim is that the method makes the disagreements legible rather than\ncreating them.\n\n**For Deliberus**: every number that enters must be argued rather than asserted, contributions must stay\nbounded (QEM's saturation already refuses extreme factors), and shared origin must be handled, or the\ngraph reproduces Rootclaim's spread with more steps.\n\n### 3.6 Decomposed estimation software\n\nQURI's Guesstimate (2016) and Squiggle build an estimate from parts, each a distribution, combined by Monte\nCarlo; its relative-value functions estimate ratios between items instead of absolute values, which\nsidesteps an absolute starting unit. It is the closest existing software to recursive decomposition of an\nestimate; it has no argument structure and no contest mechanism.\n\n### 3.7 AI forecasters\n\nThe first language-model system to approach the human crowd (Halawi, Zhang, Chen & Steinhardt, NeurIPS\n2024) decomposes questions into sub-questions to retrieve evidence, and its best reasoning prompt tells the\nmodel to consider historical base rates before settling. Base rates are part of the recipe for machines\ntoo.\n\n### 3.8 A displayed start anchors\n\nIn a randomised experiment on more than 100,000 comments, one artificial up-vote raised the chance of a\npositive rating by 32% and final scores by 25% on average, while a negative seed was mostly corrected\n(Muchnik, Aral & Taylor, *Science* 2013). Metaculus hides its aggregate at opening for the same reason. A\nstarting value shown as a verdict is a nudge, not a neutral fact.\n\n---\n\n## 4. The convergent pattern\n\nAcross philosophy, statistics, law, medicine, forecasting and platform design, the systems that work share\nfive traits:\n\n1. **The start is separate from the evidence**, and each is shown for what it is (GRADE's starting level\n   versus its rated moves; subjective logic's base rate versus belief and disbelief; the legal\n   likelihood ratio versus the prior).\n2. **The start comes from a reference class**, chosen in the open (outside view, prevalence, opportunity\n   prior, study design, first-trial probability).\n3. **The start has a strength**, not only a value (liquidity, virtual successes, subjective logic's prior\n   weight, GRADE's levels).\n4. **Several reasonable starts can be carried at once**, and the verdict that matters is the one that\n   survives them (hyper-priors, multiple reference classes, reasonable doubt across priors).\n5. **The start is not shown as a verdict before people reason** (Metaculus's hidden aggregate; the anchoring\n   evidence).\n\nThe failures share the inverse traits: a hidden start, one start for everything, extreme unexamined\nfactors, or a start displayed as an answer.\n\n---\n\n## 5. What is unique about Deliberus, and why it matters here\n\nMost of the systems above attach a prior to a question and update on evidence items. Deliberus differs in\nfour ways, and each one changes the answer.\n\n- **Recursive decomposition.** A claim can be split into parts, and the parts into parts. In a Bayesian\n  network only root nodes need priors; everything else is computed. In Deliberus the analogue is that a\n  **decomposed claim's start is derived from its parts'**; only undecomposed leaves need their own.\n  Decomposition pushes the prior problem down to the simplest claims, the ones most likely to belong to a\n  measurable reference class. This is the superforecasters' Fermi move built into the ontology.\n- **Evidence as a subgraph.** Evidence is claims and links, not an opaque input. So a reference class can\n  itself be a claim with evidence, and a claim's membership in a class can be a link that can be attacked.\n- **QEM already has a prior slot.** `qem_strength(base_weight, energy)` pulls a claim toward its base\n  weight when evidence is balanced; the docstring calls it gravity toward the initial weight. Deliberus\n  sets it to 0.5 for every claim. The machinery does not need to change for honest starts; only where the\n  number comes from does.\n- **Argumentation is the foundation, probability a computation on top.** Argumentation theory's native\n  form of a prior is **presumption and burden**: who must argue, and what stands until challenged (Freeman,\n  Carneades, the presumption of innocence). Probability's is the **base rate**: what is usually true of\n  things like this. Deliberus can hold both, and they answer different questions: presumption decides\n  whether a claim stands unargued; the base rate decides where it starts when it is weighed.\n\n---\n\n## 6. Synthesis: how starting values should work in Deliberus\n\n### 6.0 One claim, two numbers, not two versions\n\nThe founder's first question on reading this (2026-10-01): *\"Wait, what, so... there'd be two different\n\"versions\" of a claim - one prior and one posterior? Wouldn't this maybe get messy?\"* No version is added.\nEvery claim already carries two numbers: the live rule starts every claim at 0.5 (`qem_strength`'s base\nweight) and its strength is that start plus what its own evidence moved it; the start is invisible only\nbecause it is the same everywhere. The proposal changes where the start comes from, nothing else. Like an\naccount's opening balance beside its current balance, the start is an input, read through a link to the\nclaim for the kind it belongs to, and only the movement travels onward (§ 6.8). A badge still shows one\nnumber (§ 6.7); the start appears when a reader asks where a best guess came from. The real mess sits\nelsewhere: a claim of several kinds has several candidate starts. **Corrected the same day (§ 10.3):**\nnested kinds pool into one start, and only kinds that cross with no data on their overlap stay a range.\n\n### 6.1 The start is a claim too\n\nEvery non-neutral starting value in the graph should be the computed strength of an ordinary, attackable\n**reference-class claim**, linked to the claims it applies to by a membership link that can itself be\ncontested. Two families of such claims do different jobs (§ 6.2):\n\n- **Base rates set starts**: *\"hypotheses tested in this field turn out true about one time in ten\"*,\n  *\"events of this sort happen about once in N years\"*.\n- **Reliability rates weight evidence**: *\"this outlet's reported figures have held up in N of M checks\"*,\n  *\"a significant finding from a study of this design holds up about four times in ten\"*. These belong on\n  the link from a report to the claim it reports, not in the claim's start.\n\nEither way:\n\n- It makes the founder's intuition literal: the honest prior **is** computed in Deliberus, from the\n  reference-class claim's own subgraph.\n- It cannot double count: the class's evidence is about the class; the claim's evidence is about the\n  claim. A start never comes from the claim's own subgraph.\n- It meets P3 head on: no start enters without decomposition, evidence and the chance to be attacked.\n- It turns the reference class problem from a hidden choice into a visible, arguable link, which is the\n  most anyone has managed (Hájek's epistemological residue stays; it becomes contestable).\n\n### 6.2 Three things hide in the word \"origin\"\n\nWhat the founder was told on 30 September (`relevance-and-the-two-factors.md` § 7; ledger Q35) was that\nhonest starts come *by kind and origin*. Working through the Bayesian structure shows that *origin* names\nthree different things, and only the first belongs in a claim's start. **This refines that record.**\n\n1. **What kind of claim it is sets its start**: the rate at which propositions of this kind turn out true\n   *before anyone asserts this one*. A field's pre-study odds for a tested hypothesis (9% in top psychology\n   journals, § 2.6), the outside view for a forecast, indifference over the stated rival set for an\n   ordinary factual question (§ 6.4), neutral for a value claim (§ 6.5). Most claims will have no reference\n   class for a long time and stay at 0.5, which is honest: it says nothing is known of their kind.\n2. **Who asserted it is its first evidence, not its start.** The report is not the event (Schum; § 1): an\n   assertion moves a claim by as much as its asserter's reliability warrants, through a link from the\n   source once sources are nodes (ledger Q4, `what-counts-as-evidence.md` § 8c). This is Freeman's\n   *consider the source* made explicit, and it matters for propagation: only movement travels to the\n   claims a reason supports (§ 6.8), so a reliability counted *into the start* would never reach them, and\n   a reliable source's confirmed reason would support its conclusion **less** than it does today (a reason\n   confirmed to 0.95 passes on 0.05 of movement from a start of 0.9, against 0.45 from today's 0.5). It\n   keeps the existing convention too: an asserter whose reliability nobody has examined moves nothing,\n   which is how law treats a charge, not evidence of guilt (Laudan's reading of the presumption of\n   innocence, § 2.7).\n3. **Who attached it as a reason decides what its failure means.** A reason the conclusion's own source\n   gave was relied on; a reason attached from outside may have been chosen because it fails. That is a\n   property of the link, not of the claim, and how much a conclusion rested on a reason is exactly what the\n   ruled requirement claim argues (§ 6.8).\n\nThe methodology ladder (P4) therefore belongs to the second family: how much a study of a given design\nshould move the claim it reports, which is where the code already applies it, as the base weight of an\nedge (`qbaf.compute_edge_strength`), never of a claim.\n\n### 6.3 Starts need a strength\n\nStore a start as a rate and a strength (how many cases the class rests on), in the manner of a beta prior's\nmean and virtual observations, Manifold's liquidity or Davidson's virtual successes. A start from a thin\nclass should yield quickly to the claim's own evidence; a start from thousands of cases should not. This\nparallels the proposal to store two conditionals rather than one relevance number: keep the ingredients,\ncompute the number.\n\n### 6.4 Rival claims share one partition\n\nThe objection *three rival explanations cannot all be 50%* dissolves when rival claims are recognised as\nrivals. Where claims are marked as mutually exclusive (the keeper's *opposes* answer, or a stated set of\nalternatives), their starts should be set over the set and sum to at most one. That is the one place\nindifference is legitimate: across a stated partition, not claim by claim (Williamson's equivocation within\nthe constraints).\n\n### 6.5 Normative claims\n\nA value claim has no truth base rate. Its start stays at the neutral point, and any non-neutral start for\nit comes only from a named worldview lens, shown as such. This keeps the evidence reading of a normative\nclaim free of anyone's priors, which is what *maps reasoning, does not endorse conclusions* requires.\n\n### 6.6 Lenses as start sets; robust versus prior-dependent\n\nA worldview lens is, in large part, a set of starts (the lens doc already lists evidential priors among a\nlens's dials). Computing a claim's strength under each reasonable start set gives two readable outcomes:\n\n- **Robust**: the conclusion holds under every reasonable start. The legal model is Abrams's reading of\n  reasonable doubt.\n- **Prior-dependent**: it flips across reasonable starts. That is a new, honest state beside *unexamined*,\n  *unsettled* and *nothing stands*, and it names exactly where the disagreement lives, per the evidence\n  dossier's rule that every divergence must be attributable to a named slot.\n\nThis is also a **measurement of the convergence wager**. Merging theorems say starts wash out as shared\nevidence grows; the fraction of well-examined claims that are still prior-dependent measures how far they\nhave washed out in practice, and a residue that stays prior-dependent however much evidence arrives is a\nhorizon of the permissivism kind, typed rather than hidden.\n\n### 6.7 The display rule, refined\n\nThe dossier's rule (show the evidence term, never a posterior the platform had no right to choose) stays,\nwith one refinement: **a best guess may be shown when its start is a named, attributed claim**, labelled as\nsuch (*\"given the usual rate for findings like this\"*, *\"under the empiricist lens\"*). What must never happen\nis a start sitting on an unexamined claim's badge as if it were a verdict; the anchoring evidence (§ 3.8)\nis why. The evidence reading (direction and weight) stays the first thing shown.\n\n### 6.8 What it does to the first calculus ruling\n\nThe live link rule, `relevance × (acceptability − 0.5)`, is already Jeffrey's rule for a reason whose start\nis 0.5 (`bayesian-and-informal-logic.md` § 2.1). Honest starts change one symbol:\n`relevance × (acceptability − start)`. Three consequences:\n\n- **An unexamined reason still moves nothing**, because it sits at its start. Structure alone still\n  manufactures no strength.\n- **Only movement travels upward, never a start.** A conclusion learns what a reason's evidence changed,\n  not where the reason began: the no-double-counting guarantee in arithmetic form.\n- **A failed reason's loss is its fall from its start times its relevance.** From an honest start that is\n  Bayes-correct **when the conclusion relied on the reason**. A reason attached from outside breaks that\n  condition: a motivated contributor can find, for almost any true conclusion, some relevant-sounding\n  reason that fails, so the appearance of one refuted reason is nearly as likely whether the conclusion is\n  true or false, and says almost nothing about it. A start taken from the claim's kind does not repair this, since a well-chosen\n  straw man is plausible and has a middling start. **Honest starts alone do not make subtraction safe.**\n\nSo honest starts explain the first calculus ruling rather than retire it. *Failure counts against a claim\nonly through a contribution that says why* is the practical form of *subtract in proportion to reliance*:\nreliance is known when a requirement claim argues it, and drop-out is the right treatment wherever it is\nunknown, not a stopgap until starts arrive. What honest starts add is the **size** of the loss where\nreliance is known (start times relevance: the founder's two weights, derived rather than judged) and better\nbest guesses wherever a reference class exists.\n\nOne option the ruling rejected comes back into view here: subtracting for reasons the conclusion's own\nsource gave, rejected because *a link's meaning would depend on who wrote it*. That objection is weaker\nthan it looked, since relevance and requirement weights are already presumed by provenance, and the\nselection argument above gives provenance a Bayesian reason to matter. Whether a source's own *because*\ncounts as a contribution that says why is put to the founder in § 9 rather than reopened here.\n\n### 6.9 The update shape, later\n\nOnce starts are real priors, the shape of the update function matters more. QEM is quadratic near its base\n(a lone weak reason barely moves a claim), which is not the shape of a Bayesian update; Potyka's logistic\nsemantics has the log-odds form in which a start and several independent reasons combine exactly as in\nnaive Bayes. That is a separate calculus choice for later; it is noted here because honest starts make it\nbite.\n\n---\n\n## 7. Recommendations (proposed, unruled)\n\nOrdered by dependency; each names what it would need.\n\n1. **Rule the principle: a starting value is a claim too, taken from the claim's kind.** No non-neutral\n   start without a base-rate claim behind it, contestable like any claim; who asserted a claim enters as\n   evidence, never as its start (§ 6.2). This is the decision everything else rests on.\n2. **Keep 0.5 as \"no evidence either way\"** for display and for every claim with no stated start. Nothing\n   changes in the live graph until starts exist, and most claims will keep 0.5 for a long time.\n3. **Build sources as nodes** (ledger Q4), so a source's reliability claim can weight the link from its\n   reports: that is how an assertion becomes evidence, and it is the half of \"origin\" that propagates.\n4. **Rewrite the methodology ladder as reliability claims on the link side**, where the code already\n   applies it (an edge's base weight), with the measured rates of § 2.6 as first evidence, replacing the\n   invented constants (P4).\n5. **Store starts as rate plus strength**, never a bare number, **with the date and the proposer**, so each\n   start can be scored when its claim resolves.\n6. **Combine a claim's kinds by partial pooling, never by picking one** (§ 10.3). The kinds a claim belongs\n   to are themselves claims, so membership stays arguable while the combination becomes arithmetic.\n7. **Score starts against outcomes** as claims resolve (replications, events, retractions), per kind and\n   per proposer, and demote a kind or proposer whose starts prove miscalibrated: the trust ladder applied to\n   priors (§ 10.4).\n8. **A model may propose a kind and retrieve its rate; it never sets a start**, and it reads an assertion as\n   evidence, never as truth (§ 10.4, § 10.5).\n9. **Normalise starts over rival sets** marked by the keeper's *opposes* answer or a stated set of\n   alternatives.\n10. **Add the robust / prior-dependent reading** by recomputing a claim under each reasonable start set, and\n   count prior-dependent claims among the well-examined as a convergence measurement.\n11. **Then change one symbol in the link rule**, `acceptability − start` in place of `acceptability − 0.5`,\n   and let a failure subtract only where reliance is known (a requirement claim; the source's own\n   *because*, if § 9 question 2 is answered yes). Drop-out stays the default elsewhere.\n12. **Later**: estimate base rates from the graph's own history (empirical Bayes), and choose the update\n   shape (log-odds or QEM) knowing that starts now carry meaning.\n\n---\n\n## 8. Risks, and what would change the recommendation\n\n- **Gaming the reference class.** A strategic contributor will link a claim to the class that suits them.\n  Mitigations: the membership link is attackable; several applicable classes are pooled (§ 10.3), and only\n  classes that cross with no data on their overlap are shown as a range; and the strongest verdict (robust) requires every reasonable start, which removes the payoff for choosing a\n  favourable one.\n- **Anchoring.** A visible start pulls later contributions toward it (§ 3.8). Mitigation: § 6.7.\n- **False precision.** Rootclaim shows how explicit numbers can span orders of magnitude among careful\n  people. Mitigations: starts carry strength, a thinly measured kind is pooled toward neutral, crossing\n  kinds without overlap data appear as a range, contributions stay bounded, and nothing enters unargued.\n- **Calibration starvation.** Most claims never resolve, so a kind's record grows slowly and pooling keeps\n  its starts near neutral. That is honest, and it means honest starts will be rare for a long time; the\n  first kinds with real records are Deliberus's own machine passes (§ 10.1).\n- **Incoherence across claims.** Independent starts for overlapping or rival claims can be jointly\n  impossible. Mitigation: § 6.4, and a coherence check over marked rival sets. A second kind cannot be\n  checked yet: a conclusion's start should already expect its reasons at theirs, and testing that needs a\n  link's two conditionals rather than one relevance number (`relevance-and-the-two-factors.md` § 6).\n- **Reliance by identity.** If a source's own *because* counts as reliance, reliance inherits the\n  undefended sybil problem wherever authorship can be faked. Extracted reasons are safe (the source text\n  fixes who gave them); for person-written claims, argued reliance (the requirement claim) is the sturdier\n  route.\n- **The reference class problem does not go away.** This design makes it visible and contestable, which is\n  the most the literature offers.\n- **Would change the recommendation**: evidence from dogfooding that reference-class claims cannot be\n  kept honest at the rate claims arrive, or that people read a labelled best guess as a verdict despite the\n  labelling. Either would argue for keeping starts research-side and displaying only the evidence reading.\n\n---\n\n## 9. Questions only the founder can decide\n\n1. **The principle** (recommendation 1): is a starting value always a claim in the graph, taken from the\n   claim's kind, with who asserted it counted as evidence? Or may some starts be fixed constants?\n2. **The source's own *because***: when a conclusion's own source gives a reason, does that count as the\n   contribution that says why, so the reason's failure subtracts? It bears on an option the 2026-09-30\n   ruling rejected (§ 6.8). Claude's lean: yes in principle, since that reason was not selected to fail,\n   but only once the reason has an honest start, because from 0.5 it overstates the loss of an unlikely\n   reason.\n3. **What a badge shows**: the evidence reading only (today's rule), or also a labelled best guess when its\n   start is a named claim?\n4. **Normative claims**: neutral start always, with lens-supplied starts only, as proposed?\n5. **Robustness as a state**: should *prior-dependent* become a display state beside *unexamined*,\n   *unsettled* and *nothing stands*?\n6. **Crowd estimates and starts**: the best-measured way to price scientific claims is structured human\n   estimation (§ 10.2), but under the 2026-08-31 ruling bare judgments are telemetry and settle nothing. May\n   an aggregate of estimates set a start, or only raise an alarm when the graph's starts diverge from it?\n   Claude's lean: the alarm, which keeps the ruling intact and still uses the method's measured accuracy.\n\n---\n\n## 10. Follow-up (2026-10-01): kinds of claims, the state of the art, learned models, Scientist AI\n\nThe founder's follow-up, verbatim: *\"Give me a bunch of examples of \"kinds of claims\" to be clarify what\nwe're talking about here, please. Research what the SoTA is about assigning base rates wisely like this, or\nif there are different promising approaches or what's going on here. Can the recent success of LLMs/deep\nlearning illuminate the fog of uncertainty around the theoretical/academic best practice here somewhat,\nor?\"* And, while that research ran: *\"Is there any published research concerning LawZero/Bengio's concept\nof ScientistAI and how it handles these questions?\"*\n\n### 10.1 Kinds of claims, with measured rates where they exist\n\nTwo families, per § 6.2. A **base rate** says where a claim of a kind starts before anyone asserts this\none. A **reliability rate** says how far a report or a method of a kind should move a claim.\n\n| Kind | Family | Measured rate | Source |\n|---|---|---|---|\n| A hypothesis tested in a top psychology journal is true | base rate | median 9% (range 0.7–66%) | Dreber et al. 2015 |\n| A drug entering its first human trials is approved | base rate | 13.8% overall; 3.4% in cancer; about 33% for vaccines | Wong, Siah & Lo 2019 |\n| A large project finishes on time and on budget | base rate | 8.5%; with its promised benefits too, 0.5% (16,000+ projects) | Flyvbjerg's project database |\n| An event of a recurring type happens before a date | base rate | how often such events happened in comparable periods; the class is chosen per question | Tetlock & Gardner 2015 |\n| One of three stated rival explanations is right | base rate | a third each, until evidence arrives | indifference over a stated set (§ 6.4) |\n| A factual claim of a kind nobody has measured | base rate | none yet: 0.5 until a kind is argued | § 7, recommendation 2 |\n| A value claim (\"capital punishment is wrong\") | none | no truth rate exists; neutral, or a lens's start | § 6.5 |\n| A definitional claim (\"cruelty requires intent\") | none | a question about usage, for the concept layer | `core-periphery-and-the-concept-layer.md` |\n| A significant psychology finding replicates | reliability | 36% significant again; 39% judged replicated | Open Science Collaboration 2015 |\n| A social or behavioural science finding (2009–2018) replicates | reliability | 49% of 164; replication effects about 0.10 against the originals' 0.25 | SCORE, *Nature*, 2026-04-01 |\n| A finding that has passed a well-powered replication is true | reliability | median 98% | Dreber et al. 2015 |\n| A randomised trial against an observational study | reliability | starting certainty high against low | GRADE |\n| A decomposition link in Deliberus's own graph is a genuine part-of relation | reliability | 46 of 197 (23%), by a model's reading in two runs, September 2026 | `proving-a-pass-reliable.md` |\n| A replication forecast by structured expert groups is right | reliability | 76% (repliCATS) and 78% (Replication Markets), best metric | SCORE, 2026-04-01 |\n\nThree things the table shows. Most kinds people argue about have no measured rate. The measured ones come\nfrom the few fields where an institution counts outcomes. And Deliberus already measures reliability rates\nfor its own machine passes, which makes those the first kinds it could hold as claims with a real record.\n\n### 10.2 The state of the art: five approaches\n\n1. **Pick one class by rule.** Reichenbach's advice was the narrowest class with adequate statistics.\n   Kyburg's *evidential probability* made the rule precise, yields intervals rather than points, and adds\n   rules for conflicting classes; conflicts without a subset relation between the classes remained a known\n   difficulty for it.\n2. **Reference class forecasting in practice.** Kahneman and Lovallo's outside view, built out by Flyvbjerg\n   into databases of past projects (now over 16,000) used to correct optimistic estimates. People choose the\n   class, project by project.\n3. **Pool rather than pick.** A hierarchical (multilevel) model estimates a narrow class's rate by pulling\n   it toward a broader class, more strongly the less data the narrow class has: *partial pooling*. Sudip\n   Paul (June 2026) argues this answers what philosophers have called the reference class problem, once the\n   grouping structure is fixed. The classical rules of Reichenbach, Carnap, Kyburg and Pollock come out as\n   limits or special cases, and by Stein's result any rule that picks a single level is beaten by shrinkage\n   in total risk. He splits the problem in two: how to combine information across levels, which pooling\n   answers, and which grouping structure is right, which stays open. Gelman, endorsing the argument on\n   2026-09-30, restates his long-standing reading: a prior represents the reference set of relevant\n   problems, and Bayesian inference partially pools toward it.\n4. **Ask structured groups, or markets.** The IDEA protocol (investigate, discuss, estimate, aggregate),\n   run by repliCATS, and prediction markets. In the SCORE programme's papers in *Nature* (2026-04-01), the\n   two predicted which social-science findings would replicate 76% and 78% of the time on their best metric.\n   Dreber et al.'s markets managed 71% in 2015.\n5. **Learn the rate from features.** Models trained on replication outcomes, using the text and reported\n   statistics of papers, scored 14,000 psychology papers (Youyou, Yang & Uzzi, PNAS 2023) and drew replies\n   showing they partly learn writing style. In SCORE, none of three machine approaches predicted\n   replication consistently. Forecasting is the success story: LLM systems with search reached parity with\n   superforecasters on ForecastBench, on dataset questions in May 2026 and with one system above the\n   superforecaster median on market questions by July (Forecasting Research Institute, 2026-07-16; parity\n   rather than outperformance, against superforecasts elicited in 2024). A 2025 study of 38 prompts found\n   that none reliably improved forecasting accuracy, that prompts referring to base rates were among the most\n   promising, and that prompts asking the model to reason in Bayesian terms hurt (Schoenegger, Jones,\n   Tetlock & Mellers). Two neighbouring lines: LLMs as elicitors of priors for statistical models\n   (AutoElicit, Capstick et al. 2024; Selby et al. 2024), and networks trained on millions of synthetic\n   datasets that then predict near-Bayes-optimally on small real ones (TabPFN, *Nature* 2025), whose prior\n   is whatever distribution the synthetic datasets were drawn from.\n\n### 10.3 Pooling answers the \"several kinds\" problem, mostly\n\n§ 6.0 said a claim of several kinds has several candidate starts, shown as a range. **That was too\npessimistic for the common case.** Where kinds nest (a phase-two cancer drug, a cancer drug, a drug),\npooling gives one start: the narrowest kind's rate, pulled toward the broader ones in proportion to how\nlittle data it has. Where kinds cross (a cancer drug, and a drug from a small company), a model with both\ngroupings combines them, and only kinds that cross with no data on their overlap remain a range. In\nDeliberus terms the tree of kinds is itself a set of claims, so which kinds a claim belongs to stays\narguable (Paul's open half), while how their rates combine becomes arithmetic (his answered half). Pooling\nalso explains recommendation 2: a kind with no record pools all the way back to neutral.\n\n### 10.4 What learned models illuminate\n\n1. **They relocate the choice rather than remove it.** A trained network's prior is the data it was trained\n   on. The prior-fitted-network authors say so outright (Müller, Reuter, Hollmann, Rügamer & Hutter 2025),\n   and it is the reference class problem one level up: which data counts as relevant.\n2. **They replace justification with checking.** Nobody argues in advance which reference class a\n   forecasting model should use; it is judged by how its probabilities match outcomes on questions it has\n   not seen. That is Dawid's prequential principle (1984), and it is what made the fog navigable in\n   practice. Bob Carpenter, commenting on the same Gelman post, describes the working version: check the\n   calibration of what the model predicts, and fit a better model if it fails.\n3. **It works where outcomes arrive, not yet where they are scarce.** Forecasting questions resolve in\n   weeks and number in thousands, and AI forecasters caught superforecasters there. Replications are slow,\n   expensive and few, and in SCORE trained humans still beat the machine methods.\n4. **Trained on raw assertions, a model mistakes how often something is said for how likely it is to be\n   true.** LawZero's own example: a corpus flooded with \"the Earth is flat\" teaches a raw model the claim,\n   while a contextualized corpus teaches it only about the people who say so (§ 10.5).\n5. **Consequence for Deliberus.** Most claims in the graph will never resolve, so calibration data will be\n   scarce, and pooling then keeps starts near neutral on its own. That is the honest outcome rather than a\n   failure: a start moves away from 0.5 only as far as a kind's resolved record justifies, which is\n   recommendation 7 of § 7 in one sentence.\n\n### 10.5 Scientist AI\n\nFour publications from Bengio's LawZero line bear on this. The overview was read in full for this section,\nthe others through their abstracts and quoted passages; the safety case is read in full in\n`bengio-safety-from-honesty-and-deliberus.md`.\n\n- Bengio et al. (February 2025), *Superintelligent Agents Pose Catastrophic Risks: Can Scientist AI Offer a\n  Safer Path?* A world model holding a Bayesian posterior over explanatory theories, giving simpler theories\n  more prior probability (Occam's razor), and epistemic humility: in doubt, avoid actions that some theory in\n  the posterior says could cause major harm.\n- Bengio, Cohen, Malkin, MacDermott, Fornasiere, Greiner & Kaddar (UAI 2025), *Can a Bayesian Oracle\n  Prevent Harm from an Agent?* Guardrails that bound the probability of harm under the true but unknown\n  theory, by searching for a cautious theory among the plausible ones.\n- Fornasiere, Richardson, Gendron, Serban & Bengio (LawZero, 2026-02-05), *The Scientist AI: Safe by Design,\n  by Not Desiring.* **Contextualization**: training text is rewritten so that \"Newton has an apple\" becomes\n  \"Newton said he has an apple\", and the truth of the original statement becomes a hidden variable inferred\n  from the pattern of evidence.\n- LawZero's safety case (arXiv:2606.29657, July 2026): statements as variables, communication acts kept\n  apart from factual observations, a description-length-biased prior over causal models as its example of\n  a broad prior, calibration, and abstention when uncertainty is large.\n\nHow they meet the questions of this document:\n\n- **Where the prior lives**: on whole theories, as a preference for simple ones, never on individual claims.\n  A claim's probability is an output of the learned model, so its base rate is implicit in what the model\n  learned rather than a stated, arguable number.\n- **Assertion as evidence**: contextualization is the split of § 6.2 (who asserted a claim is evidence about\n  it, not its start), reached independently and for a different reason: to stop a model learning confident\n  falsehoods by repetition.\n- **Uncertainty about the prior**: handled by caution across the plausible theories and by abstaining,\n  which is the guardrail counterpart of the robust-across-reasonable-starts reading of § 6.6.\n- **Values**: the safety case gives no guarantee for contested moral claims, as § 6.5 keeps them neutral.\n- **What none of them addresses**, in the passages read: how to choose or calibrate a base rate for an\n  individual claim against claims that later resolve. Their answer is a well-trained model; Deliberus's\n  would be a stated, arguable kind with a record.\n\nThe two compose. A calibrated predictor of the Scientist AI kind is the strongest candidate proposer of\nstarts Deliberus could have, and Deliberus's attributed claims, scored against what later resolves, are\nthe kind of empirical validation the safety case says it still needs.\n\n### 10.6 What this changes\n\nRecommendations 5 to 8 of § 7 were rewritten from this section (dated and attributed starts, pooling,\nscoring against outcomes, models as proposers), § 6.0 and § 8 were corrected on the several-kinds point,\nand § 9 gained question 6 on crowd estimates.\n\n---\n\n## 11. The open questions in one picture (2026-10-03)\n\nThe founder, two days after § 10: *\"Ok I've still yet to fully absorb all this but the SoTA 2026 paper was\ninteresting. I'd like to learn more about that.\"* Then: *\"Try to visualize the central crucial conundrums we\nstand before here please. Make sure you ground extremely deeply, and think deeply about HOW to visualize\nthis best.\"*\n\nThe result is the page **Where a claim starts** (claude.ai/artifact/8eyM83VEoFmJBKutpxqmcj, private), a\nsequel to *When a reason fails* in the same palette and faces. It adds no research; it redraws § 2.6 and\n§§ 6–10.\n\n**What it shows.**\n\n- **One measured example first.** Dreber's three stages (§ 2.6) on one line from 0 to 100%: 9% before any\n  study, 56% after one significant finding, 98% or 6% after the replication, with Deliberus's 50% drawn\n  beside the first. A ring marks where a claim starts and a dot where it stands, which is § 6.0 as a legend.\n- **Five questions, each drawn on that same line** as a gap between two positions, each with one sentence\n  stating its tension:\n  1. *Where should a claim start?* One half says nothing is known; for some kinds something is. The kinds\n     of § 10.1 as a dot plot, and the start read off a claim about the kind (§ 6.1).\n  2. *Which kind is it?* The same claim starts at 14% or at 33%, depending on where it is filed. A slider\n     pulls the narrow kind's rate toward the broad kind's by how few cases the narrow kind has (§ 10.3).\n     The rates are Wong, Siah and Lo's; the weight given to the broad kind (30 cases) is an example. The\n     slider's two ends are the single-kind rules as limits.\n  3. *Who checks the start?* Only what later turns out true can, and most claims never get that (§ 10.4,\n     and *calibration starvation* in § 8).\n  4. *When a reason fails, does the claim lose?* Only where the claim leaned on it (§ 6.8). The bridge of\n     the earlier page, with a reason its own source gave beside one a sceptic attached, and the ruled\n     drop-out (Q34) stated as what Deliberus does today.\n  5. *What may the badge say?* Three starts (10%, 50%, 90%) under the live `qem_strength` and display\n     bands, converging as checked evidence grows: different badges are *prior-dependent*, one badge is\n     *robust* (§ 6.6–6.7).\n- **The rulings of § 9 in dependency order. Claude's reading, unruled**: question 1 of § 9 is the root, and\n  the other five arise only if its answer is yes. If it is *not yet*, every claim keeps 0.5 and Q34's\n  drop-out stands as it is. The yes branch holds normative claims (§ 9 question 4), crowd estimates\n  (question 6), the labelled best guess (question 3), the prior-dependent state (question 5) and the\n  source's own *because* (question 2).\n\n**Method, for the next visual in this arc.** A conundrum is two things that both seem right, so it is drawn\nas a gap between two marks on one shared scale, and the scale is the same in every figure. Each question\ncarries a one-sentence tension under its heading, so the five headings and five sentences carry the page\non their own. Measured figures name their source; illustrative positions say so. The two interactive\nfigures compute what they show: question 5 with the site's own strength formula and bands (checked against\n`qbaf.qem_strength` and `display_band`), question 2 with an ordinary shrinkage estimate.\n\n**A precision made while drawing.** A first draft said the graph draws a relied-on reason and a planted one\nthe same way. It does not: a link records who wrote it. What § 8 calls *reliance by identity* is the\naccurate statement, and the page carries it: who attached a reason is on record for reasons taken from a\nsource's own text, and can be faked for reasons added by hand.\n\n**Open loop.** He asked to learn more about Paul's partial-pooling paper (§ 10.2, item 3); question 2 of\nthe page is a first pass at it. A fuller walk through the paper is owed: `TODO.md`, the honest starting\nvalues item.\n\n**What the number is (2026-10-04).** The page first said Deliberus *\"computes a number for every claim\"*.\nThe founder: *\"'computes a number' is very vague and overly modest, innit? can the page say what it is?\nIt's a truth estimate, right, this is epistemic calculus we're talking about? No?\"* The vagueness was a\ndodge of an open ruling, ledger P7 (*does Deliberus ever output a probability?*), where the evidence\ndossier recommends that the badge report the evidence term and never a posterior (ledger Q1) and his own\nwords say the strengths are meant to become the graph's best guesses (Q35). The answer given, and what the\npage now says:\n\n- **In aim, yes**: the number is meant to be a truth estimate, the graph's best guess at how likely a\n  claim is true.\n- **Today it is the evidence half of one**: how well the claim has held up under the reasons examined. A\n  truth estimate is that reading plus a start, and every start is a blind 0.5. While the start is 0.5 the\n  two readings give the same number, which is why the constant *does two jobs* and why the name was easy to\n  blur.\n- **So P7 and Q35 are one question from two sides.** *Strength only* is right while starts are neutral; a\n  labelled truth estimate becomes available exactly where a start is a named, attackable claim (§ 6.7,\n  § 9 question 3). Checking against outcomes (§ 10.4; the oracle bridge of ledger P5) is what would let the\n  estimate be trusted.\n\nThe page's opening now reads: *Deliberus runs an epistemic calculus. For every claim it works out how well\nthe claim holds up, from the reasons for and against it and from how well each of those holds up in turn.\nThe aim is a truth estimate: the graph's best guess at how likely the claim is true. What the number lacks\nfor that is an honest start.* **Unruled**: whether public copy may say flatly that the number *is* a truth\nestimate. *Epistemic calculus* is his name for what these docs call the strength calculus.\n\n---\n\n## Sources\n\n**Philosophy of probability and evidence**\n- Stanford Encyclopedia of Philosophy: *Bayesian Epistemology* (2022), §§ 1.5, 4, supplements C–E;\n  *Formal Epistemology* § 1.4; *The Problem of Induction* § 3.3; *Scientific Objectivity* § 4.2.\n- Hájek, A. (2007). The reference class problem is your problem too. *Synthese* 156, 563–585.\n- Williamson, J. (2010). *In Defence of Objective Bayesianism*. Oxford University Press.\n- Joyce, J. M. (2005). How probabilities reflect evidence. *Philosophical Perspectives* 19, 153–178;\n  (2010). A defense of imprecise credences in inference and decision making. *Philosophical Perspectives* 24.\n- Levi, I. (1980). *The Enterprise of Knowledge*. MIT Press, ch. 13.\n- Rinard, S. (2013). Against radical credal imprecision. *Thought* 2; (2014). The principle of\n  indifference and imprecise probability. *Thought* 3.\n- Vallinder, A. (2018). Imprecise Bayesianism and global belief inertia. *British Journal for the\n  Philosophy of Science* 69.\n- Blackwell, D., & Dubins, L. (1962). Merging of opinions with increasing information. *Annals of\n  Mathematical Statistics* 33.\n- Sprenger, J. (2018). The objectivity of subjective Bayesianism. *European Journal for Philosophy of Science* 8.\n\n**Statistics and measured base rates**\n- Efron, B. (2021). Empirical Bayes: concepts and methods; (2010). *Large-Scale Inference*. Cambridge.\n- Robbins, H. (1956). An empirical Bayes approach to statistics. *Proc. Third Berkeley Symposium*.\n- Ioannidis, J. P. A. (2005). Why most published research findings are false. *PLOS Medicine* 2(8), e124.\n- Open Science Collaboration (2015). Estimating the reproducibility of psychological science. *Science* 349.\n- Camerer, C. F., et al. (2016). Evaluating replicability of laboratory experiments in economics. *Science*\n  351; (2018). Evaluating the replicability of social science experiments in *Nature* and *Science*\n  between 2010 and 2015. *Nature Human Behaviour* 2.\n- Dreber, A., et al. (2015). Using prediction markets to estimate the reproducibility of scientific\n  research. *PNAS* 112, 15343–15347.\n\n**Law and medicine**\n- Fenton, N., Lagnado, D., Dahlman, C., & Neil, M. (2019). The opportunity prior. *Law, Probability and Risk*.\n- Dahlman, C., & Kolflaath, E. (2021). The problem of the prior in criminal trials. In *Philosophical\n  Foundations of Evidence Law*, Oxford University Press.\n- Laudan, L. (2013). Why presuming innocence is not a Bayesian prior. *Error Statistics Philosophy* (blog).\n- Abrams (2025). Reasonable doubt and reasonable priors. *Episteme*.\n- Cochrane Handbook, ch. 14 (GRADE); GRADE Book (GRADE Working Group).\n\n**Argumentation**\n- Baroni, P., Rago, A., & Toni, F. (2019). From fine-grained properties to broad principles for gradual\n  argumentation. *International Journal of Approximate Reasoning*.\n- Freedman, G., Dejl, A., Gorur, D., Yin, X., Rago, A., & Toni, F. (2025). Argumentative large language\n  models for explainable and contestable claim verification. *AAAI* 39; follow-up on uncertainty\n  quantification in ArgLLMs, EMNLP 2025.\n- Leite, J., & Martins, J. (2011). Social abstract argumentation. *IJCAI*.\n- Hunter, A., & Thimm, M. (2017). Probabilistic reasoning with abstract argumentation frameworks. *JAIR*;\n  Hunter, A., Polberg, S., & Potyka, N. (2018). Updating belief in arguments in epistemic graphs. *KR*.\n- Freeman, J. B. (2005). *Acceptable Premises: An Epistemic Approach to an Informal Logic Problem*.\n  Cambridge University Press.\n- Hahn, U., & Oaksford, M. (2007). The rationality of informal argumentation: a Bayesian approach to\n  reasoning fallacies. *Psychological Review* 114, 704–732; Oaksford & Hahn (2004) on the argument from\n  ignorance, *Canadian Journal of Experimental Psychology* 58.\n- Jøsang, A. (2016). *Subjective Logic: A Formalism for Reasoning Under Uncertainty*. Springer.\n- Schum, D. A. (1994). *The Evidential Foundations of Probabilistic Reasoning*. Wiley (as summarised in\n  `what-counts-as-evidence/formal-and-computational-models-of-evidence.md` § 3.1).\n- Oren, N., Norman, T. J., & Preece, A. (2007). Subjective logic and arguing with evidence. *Artificial\n  Intelligence* 171, 838–854.\n\n**Forecasting, markets and software**\n- Tetlock, P., & Gardner, D. (2015). *Superforecasting*; Good Judgment, \"Ten commandments for aspiring\n  superforecasters\".\n- Chang, W., et al. (2016). Developing expert political judgment: the impact of training and practice on\n  judgmental accuracy in geopolitical forecasting tournaments. *Judgment and Decision Making* 11.\n- Karvetski, C., et al. What do forecasting rationales reveal about thinking patterns of top geopolitical\n  forecasters? *International Journal of Forecasting*.\n- AI Impacts (2019). Evidence on good forecasting practices from the Good Judgment Project.\n- Davidson, T. (2021). Semi-informative priors over AI timelines. Open Philanthropy.\n- McCaslin, T., Rosenberg, J., Karger, E., et al. (2024). Conditional trees: a method for generating\n  informative questions about complex topics. Forecasting Research Institute; Karger, E., et al. (2023).\n  Forecasting existential risks: evidence from a long-run forecasting tournament.\n- Metaculus FAQ and aggregation documentation; Manifold docs and \"Above the Fold: market mechanics\" (2022).\n- Hanson, R. (2003/2007). Logarithmic market scoring rules for modular combinatorial information\n  aggregation. *Journal of Prediction Markets* 1.\n- Othman, A., Pennock, D. M., Reeves, D. M., & Sandholm, T. (2010). A practical liquidity-sensitive\n  automated market maker. *EC*.\n- Alexander, S. (2024). Practically-a-book review: Rootclaim $100,000 lab leak debate; Rootclaim (2024),\n  debate results and response.\n- QURI: Guesstimate and Squiggle documentation; Gooen, O. (2023). Relative value functions.\n- Halawi, D., Zhang, F., Chen, Y.-H., & Steinhardt, J. (2024). Approaching human-level forecasting with\n  language models. *NeurIPS*.\n- Muchnik, L., Aral, S., & Taylor, S. J. (2013). Social influence bias: a randomized experiment. *Science* 341.\n\n**Follow-up sources (§ 10)**\n- Paul, S. (2026). From selection to shrinkage: partial pooling and the reference class problem. Zenodo,\n  2026-06-27.\n- Gelman, A. (2026-09-30). Prior distributions, partial pooling, and reference sets (Bayesians are\n  frequentists). *Statistical Modeling, Causal Inference, and Social Science* (blog), with comments by\n  B. Carpenter.\n- Kyburg, H. E. (1991). Evidential probability. *IJCAI*.\n- Dawid, A. P. (1984). Statistical theory: the prequential approach. *Journal of the Royal Statistical\n  Society A* 147.\n- Flyvbjerg, B., & Gardner, D. (2023). *How Big Things Get Done*; the Oxford project database (16,000+\n  projects).\n- Wong, C. H., Siah, K. W., & Lo, A. W. (2019). Estimation of clinical trial success rates and related\n  parameters. *Biostatistics* 20(2), 273–286.\n- SCORE collaboration (2026). Three papers in *Nature* on reproducibility, robustness and replicability and\n  its prediction (released 2026-04-01).\n- Hemming, V., Burgman, M., et al. (2018). A practical guide to structured expert elicitation using the IDEA\n  protocol. *Methods in Ecology and Evolution*; the repliCATS process (*PLOS ONE*, 2023).\n- Youyou, W., Yang, Y., & Uzzi, B. (2023). A discipline-wide investigation of the replicability of\n  psychology papers over the past two decades. *PNAS*; replies by Crockett et al. (2023) and Mottelson &\n  Kontogiorgos (2023).\n- Forecasting Research Institute (2026-07-16). AI models have likely reached parity with superforecasters\n  on ForecastBench.\n- Schoenegger, P., Jones, C. R., Tetlock, P. E., & Mellers, B. (2025). Prompt engineering large language\n  models' forecasting capabilities. arXiv:2506.01578.\n- Alur, R., Stadie, B. C., Kang, D., et al. (2025). AIA Forecaster: technical report. arXiv:2511.07678.\n- Capstick, A., et al. (2024). AutoElicit: using large language models for expert prior elicitation in\n  predictive modelling. arXiv:2411.17284.\n- Selby, D. A., et al. (2024). Had enough of experts? Elicitation and evaluation of Bayesian priors from\n  large language models. NeurIPS 2024 workshop; journal version in *Stat* (2025).\n- Hollmann, N., et al. (2025). Accurate predictions on small data with a tabular foundation model.\n  *Nature*; Müller, S., Reuter, A., Hollmann, N., Rügamer, D., & Hutter, F. (2025). Position: the future of\n  Bayesian prediction is prior-fitted. arXiv:2505.23947.\n- Bengio, Y., et al. (2025). Superintelligent agents pose catastrophic risks: can Scientist AI offer a\n  safer path? arXiv:2502.15657.\n- Bengio, Y., Cohen, M. K., Malkin, N., MacDermott, M., Fornasiere, D., Greiner, P., & Kaddar, Y. (2025).\n  Can a Bayesian oracle prevent harm from an agent? *UAI*, PMLR 286.\n- Fornasiere, D., Richardson, O., Gendron, G., Serban, I., & Bengio, Y. (2026). The Scientist AI: safe by\n  design, by not desiring. LawZero, 2026-02-05.\n- LawZero (2026). Safety from honesty in a disinterested AI predictor. arXiv:2606.29657.\n\n## Cross-references\n\n`relevance-and-the-two-factors.md` (§§ 6–7) · `strength-calculus-literature/bayesian-and-informal-logic.md`\n· `strength-calculus-literature/computational-argumentation.md` · `what-counts-as-evidence.md` (§ 8c) ·\n`what-counts-as-evidence/philosophy-of-evidence.md` (§§ 4.3–4.4) ·\n`what-counts-as-evidence/formal-and-computational-models-of-evidence.md` (§§ 2.3–2.4) ·\n`forecasting-room-at-the-top-and-deliberus.md` · `docs/worldview-lenses.md` · `docs/convergence.md` ·\n`session26-open-rulings.md` (P2, P3, P4, Q4, Q34, Q35) · `bengio-safety-from-honesty-and-deliberus.md` ·\n`proving-a-pass-reliable.md` · `core-periphery-and-the-concept-layer.md` ·\n`listening-for-the-calculus-and-sturdy-microtasks.md` (recorded talks and episodes on these questions,\nfound 2026-10-05 and not yet heard: Pettigrew's 2025 Oxford lecture on priors, Dahlman's talk on a prior\nfixed by rule, Fitelson on argument strength as two numbers, Hájek).\n"}