{"path":"research/burden-of-proof-and-priors.md","content":"# Burden of proof, sensible priors, and the causes of belief\n\n**Date**: 2026-10-23 · **Type**: the founder's question, the corpus re-read, two outside research passes,\narithmetic, and a synthesis · **State**: every proposal here is unruled (ledger Arc AZ). The outside sources\nare search-grade throughout (abstracts and secondary pages); the two helpers' reports are kept verbatim in\n[burden-of-proof-and-priors/](burden-of-proof-and-priors/), with their read-status tags. The arithmetic is\nin `burden-of-proof-and-priors/cases.py`.\n\n## 0. The answer first\n\n**\"Burden of proof\" names three different things, and Deliberus already has a home for two of them.**\n\n1. **Where a claim starts**, before its own evidence: the prior. *Extraordinary claims need extraordinary\n   evidence* is this one. Its home is the proposed starting value by kind (ledger Q35,\n   [starting-values-and-the-problem-of-the-prior.md](starting-values-and-the-problem-of-the-prior.md)).\n2. **Who must speak next** for a claim's number to move: the obligation to argue. Deliberus has ruled this\n   one and made it the same for everyone: a reason that fails drops out (Q34), and defeating an attack does\n   not establish its opposite (AL2). Whoever wants the number to move contributes; nobody has to argue for\n   a claim to exist.\n3. **How sure one must be to act**: the threshold, or standard of proof. It comes from what each kind of\n   error costs, which is a value, and it belongs to a decision. Deliberus stores no threshold on any claim.\n\nThe causes of belief (identity, ideology, groupthink, ego) work on all three, and the research shows where.\n**Ideology picks the reference class**, so what counts as *extraordinary* is a worldview's prior in\ndisguise. **Motivated reasoning moves the threshold by side**: a preferred conclusion needs only to be\npossible, a dispreferred one must be proved (Ditto and Lopez 1992; Gilovich's *can I believe it?* against\n*must I believe it?*). **Groupthink counts heads where it should count origins**: ten people who heard one\nreport are one report (Bovens and Hartmann). **A bias accusation moves the obligation** from the claim to\nthe person's motive, which is Bulverism, already guarded by the fourth bias question (E21).\n\nSo when someone in a debate says *the burden is on you*, Deliberus's job is to ask which of the three they\nmean, and file it as a claim of that kind, where it can be argued. Most such sentences turn out to be the\nthird disguised as the first.\n\n## 1. His questions, verbatim\n\nFrom the 2013 test map of 19 April, in the Demeter archive, put to him one item at a time on 2026-10-23:\n*\"Argumentera inte för nej – angrip ja-sidans argument? Bevisbörda?\"* (\"Don't argue for no, attack the yes\nside's arguments? Burden of proof?\")\n\nHis reply, 2026-10-23: *\"The \"burden of proof\" question feels very connected to the question of what\nsensible priors one could/should have for all claims. And all this is affected by, on at least a\npsychological and cultural plane, everything re motivated reasoning, \"causes for belief\", groupthink,\nideology and identification/ego and bias-accusations etc etc. Read up on this and make the connections you\nsee persistent in the docs. Ultrathink\"*\n\n**What is ruled**: the obligation half (Q34, AL2), the fourth bias question (E21), that the machine never\npublicly says why someone believes something (AN5). **What is open**: honest starting values (Q35, Q36),\npriors in the lens and standards in the symmetry (AN6), the symmetry instruments (AN7, AN8). **What is new\nhere**: that *burden of proof* splits three ways, and where each cause of belief acts on each part.\n\n## 2. What the corpus already holds\n\n| Angle | What it says | Home |\n|---|---|---|\n| Presumption versus base rate | argumentation's prior is presumption and burden; probability's is the base rate; they answer different questions | `starting-values-and-the-problem-of-the-prior.md` § 5 |\n| Reasonable doubt across priors | the strongest verdict survives every reasonable start (Abrams 2025) | same note § 2.7, § 6.6 |\n| Laudan: the presumption of innocence is not a prior | an instruction about proof, not a probability | same note § 2.7, § 6.2 |\n| Carneades | a proof standard per statement; an *assumption* holds until questioned, an *exception* until proved | `what-counts-as-evidence/formal-and-computational-models-of-evidence.md` § 1.5 |\n| Hahn and Oaksford on the argument from ignorance | strength depends on the search, not on who carries a burden | `strength-calculus-literature/bayesian-and-informal-logic.md` |\n| Burden as a frame-level crux in real debates | *presume harm or presume valid choice when unverifiable* (assisted dying); *burden of proof under asymmetric error costs* (psychedelics); *which evidential standard governs one-shot irreversible risks* (AI risk) | `cross-source-premise-run.md`; `frontier-extraction-experiment.md` H9 |\n| The `ignorance` scheme | its questions include *the search was thorough* and *the burden of proof is correctly allocated* | `schemes.py` (`search_adequate`, `burden_correct`) |\n| Causes of belief, three entries | as attack (Bulverism), as stake (testimony), as pattern (measured, never attributed) | `causes-of-belief-and-the-three-minds.md` |\n| Priors versus standards | different priors can be rational (Dorst); different standards by side are the motivated part (Ditto et al. 2018, Kahan) | same note § 7, ledger AN6 |\n| Independence of reasons | reasons that share a likely explanation count once | ledger Q40, `how-independent-reasons-add.md` |\n\n**Not in the corpus before this note** (counted by search): motivated skepticism (Ditto and Lopez 1992),\nGilovich's two questions, Taber and Lodge, self-affirmation, Abelson's *beliefs as possessions*, status quo\nbias, Whately's presumption, Ullmann-Margalit's presumption rule, the decision threshold from error costs\n(Kaplan 1968, Blackstone), the Sagan standard and its history, precaution as a reversed burden, and every\nhistorical case in § 5.\n\n## 3. What the literature adds\n\n### 3.1 The three burdens of evidence law, and a fourth in argument\n\nEvidence law separates the **burden of persuasion** (who loses if the evidence is balanced) from the\n**burden of production** (who must put evidence forward on an issue); Williams (2003) adds the **tactical\nburden**, a running reading of who is winning (Prakken and Sartor use the three; theory report § B).\nWalton's own split is global against local burdens. For Deliberus: persuasion is a tiebreak for a\ndecision, so it is part of the threshold; production and tactical burden are about who speaks next, so they\nare the obligation, and the tactical one is derived state, never stored.\n\n### 3.2 Burden matters where a decision is made\n\nHahn and Oaksford (2007b, abstract only): in law the burden matters because law decides; in critical\ndiscussion the aim is *\"frequently simply to increase or decrease degree of belief\"*, and whether belief\ncrosses a threshold is *\"not necessarily important\"*. They call the carry-over from law to everyday argument\na *\"hasty transference\"* (secondary). Their companion paper adds that an argument from ignorance does not\nfail *because* it fails a burden; it fails because it is weak, which its content and the search decide. This\nis the strongest outside support for keeping thresholds off claims, and for Q34's third route, the argued\nthorough search.\n\n### 3.3 A threshold is a ratio of error costs\n\nKaplan (1968): act on *true* when the probability exceeds the cost of a false positive over the sum of both\nerror costs. Blackstone's *\"better that ten guilty persons escape than that one innocent suffer\"* gives\nabout 0.91. So the same evidence can rightly convict in a civil court and acquit in a criminal one, and\nprecaution is the same formula with the costs the other way round (§ 4, case 1). Ullmann-Margalit's\npresumption has the form *when P holds, act as if Q unless there is sufficient reason to think otherwise*: a\nrule for acting, not a probability (secondary). Whately's presumption for existing institutions was\ncriticised, by Kauffeld among others, for giving no criteria and falling back on common sense: a default\ndressed as a probability.\n\n### 3.4 \"Extraordinary\" is a prior, and whose prior\n\nThe line runs from Hume (1748) and Laplace (*\"the more extraordinary a fact, the more it needs to be\nsupported by strong proofs\"*) through Flournoy and Truzzi to Sagan. Its Bayesian reading is exact: a claim\nthat starts at one in a thousand needs a likelihood ratio near a thousand to reach even odds (§ 4, case 2).\nThe standing objection is that *extraordinary* depends on the worldview of whoever says it. Both are true,\nwhich is why the start must be a stated reference class anyone can attack (Q35's proposal), not the\nspeaker's sense of what is normal. Since ledger Q46 was built (2026-10-23) a claim's strength is the\nlog-odds of its start plus each independent argument's shift, so a start other than 0.5 is one input to\nchange, not a new calculus (the outside reader's catch).\n\n### 3.5 The psychology of which side must prove\n\n- **Motivated skepticism** (Ditto and Lopez 1992): people needed less information to conclude that a\n  disliked person was less intelligent than that he was more; given an unwelcome medical test result, they\n  retested more, found more reasons it might be wrong, and rated the test less accurate. The bar rises for\n  the unwanted conclusion.\n- **Gilovich's two questions**: for a welcome conclusion one asks *can I believe this?*, for an unwelcome\n  one *must I believe this?* Dawson, Gilovich and Regan (2002) found the stricter test can make people\n  *more* accurate on a logic task. Skepticism is not the bias; its direction is.\n- **Taber and Lodge (2006)**: arguments for one's own side rated stronger, opposing ones counter-argued,\n  most of all by the most sophisticated (no direct replication found; Druckman 2012 asks how far it\n  reaches outside the lab).\n- **Status quo bias** (Samuelson and Zeckhauser 1988; three of four scenarios replicated): a default is an\n  implicit burden rule.\n- **Beliefs as possessions** (Abelson 1986) and the identity reading of resistance (Kaplan, Gimbel and\n  Harris 2016, one fMRI sample of 40): resistance is strongest where a belief is part of who one is.\n- **Self-affirmation** (letting people affirm a value first, so that threatening evidence costs less) is\n  **not reliable** for political or factual belief: several teams found little effect (Lyons et al. 2021),\n  and a large educational replication ruled out benefits above d = 0.10. Recorded so nobody builds on it.\n- **Correlated testimony**: agreement adds support only as far as its sources are independent; ten copies\n  of one report add no more than the report (Bovens and Hartmann 2003). Informational cascades make\n  following the crowd rational for each person while the crowd's agreement carries one person's\n  information (Bikhchandani, Hirshleifer and Welch 1992).\n- **Bias attributions** track disagreement, not the opponent's reasoning: people see the other side as\n  biased (Pronin, Gilovich and Ross 2004; Kennedy and Pronin 2008). No study found shows that a bias\n  accusation predicts anything about the quality of the accused reasoning.\n\n## 4. The arithmetic, made plain\n\nFrom `cases.py`, nothing fitted.\n\n1. **One body of evidence, three thresholds.** Evidence with likelihood ratio 4 from a neutral start gives\n   0.80. With equal error costs the threshold is 0.50: act. With Blackstone's 10 to 1, 0.91: do not act.\n   With precaution, where the harm is ten times worse than the lost benefit, 0.09: act. The claim's strength\n   is the same 0.80 in all three. Only the decision differs.\n2. **What \"extraordinary\" costs.** To reach even odds a claim needs a likelihood ratio of 1 from a start of\n   0.5, 9 from 0.1, 99 from 0.01, 999 from 0.001, and about a million from one in a million. To reach 0.95,\n   nineteen times as much. The demand is exact; the start is the argument.\n3. **His 2013 question.** A *yes* claim with two reasons, each at relevance 0.5, both refuted to 0.1.\n   Under Q34 the claim moves nothing and reads *nothing stands (2 checked)*. Under the rejected rule\n   (subtract each failure from 0.5, in the linear form the strength review uses) it falls by 0.40, which is\n   the straw-man exploit: attach two weak reasons to a true claim, refute them, and it reads 0.10. An\n   objection of strength 0.8 at relevance 0.5 moves the claim down 0.15 under either rule. So *attack the\n   yes side's reasons* and *argue for no* do different jobs: the first removes support, the second adds\n   evidence against. To make *yes* probably false, the *no* side must argue.\n4. **Motivated skepticism in one line.** One piece of evidence with likelihood ratio 3 gives 0.75. Read\n   with *can I believe it?* (a bar of 0.30) it is accepted; with *must I believe it?* (0.95) it is\n   rejected. The posterior is the same; the bar moved with the side.\n5. **Consensus by origin.** Ten agreeing sources, each worth a likelihood ratio of 2: from ten independent\n   lines, 0.999; from two lines, 0.80; from one line copied ten times, 0.67.\n\n## 5. Tests on real cases\n\nNone of these domains appears in the corpus before this note. Each case asks: which of the three burdens\ndecided it, and would the split have helped?\n\n| Case | Which burden decided it | What the split does with it |\n|---|---|---|\n| **Thalidomide, FDA, 1960–62** (Kelsey asked for more data than physician testimonials; the US largely escaped the birth-defect epidemic) | the threshold, for a decision (approval) whose false *safe* was catastrophic | correct and stated in the right place: a high bar on *approve*, which says nothing about the drug claim's start. The burden on the maker is the decision's error costs |\n| **Leaded petrol, Kehoe's paradigm, 1925–1970s** (restriction only on clear evidence of harm) | the threshold set backwards for a decision whose false *safe* fell on the public | the split names the move: *no action without clear evidence of harm* is a claim about error costs, and it would have stood in the graph as an attackable claim rather than as the default of the debate |\n| **Tobacco, \"doubt is our product\", 1969** (wording unverified; the strategy is well documented) | the obligation, gamed: demand that the other side prove causation to a standard never accepted, and manufacture doubt | under Q34 a doubt carries weight only as an argued objection with its own strength; an unexamined objection moves nothing. A raised threshold is a separate claim about error costs. Doubt-making needs both to pass as one |\n| **Semmelweis, 1847** (mortality 18.3% before chlorinated handwashing, about 2% after; rejected for want of a mechanism) | the obligation, misplaced: a missing mechanism was treated as a refutation | under Q34 a missing mechanism is an absent reason, not an objection; the drop in mortality is a count of cases (Arc AY) and counts. A demand for a mechanism is a requirement claim that must itself be argued, and here it would have been attacked by the count |\n| **H. pylori, 1982–2005; continental drift, 1912–1960s** (true claims held down by consensus and a missing mechanism) | the start, too low and too stiff: *contradicts the textbook* treated as near zero | a start by kind is a claim with a strength; a thin or argued kind yields to evidence. The risk to name: a reference class such as *claims against medical consensus* would slow them, so that class's own record (how often such claims later held) is what sets it |\n| **Cold fusion, 1989; Bem's precognition, 2011** (extraordinary claims, failed replications: Galak et al., N = 3,289, d = 0.04) | the start, correctly low, plus independent replications | this is the split working as Sagan meant it: a low start by kind, and replications counted as independent cases. Bem shows the order: the paper was published, then the bar was applied |\n| **O. J. Simpson, 1995 and 1997** (acquitted under *beyond reasonable doubt*, liable under preponderance) | the threshold, but confounded: the civil case heard evidence the criminal jury did not | an illustration of § 4 case 1 only, not a test |\n\n**What predicts a failure**: in every case where the burden hurt (lead, tobacco, Semmelweis, H. pylori), a\nthreshold or a missing reason was treated as if it were the claim's probability. In every case where it\nhelped (thalidomide, Bem), it was applied to a decision, or counted as evidence. That is the split.\n\n## 6. Where the causes of belief act on each burden\n\n| Cause | On the start | On the obligation | On the threshold |\n|---|---|---|---|\n| **Ideology, worldview** | picks the reference class, so sets what is *extraordinary* | | decides which error is worse (precaution against proaction) |\n| **Motivated reasoning** | | more scrutiny for the unwelcome side (Ditto and Lopez; Dorst's selective scrutiny) | a higher bar for the unwelcome conclusion (*must I*) |\n| **Groupthink, conformity** | consensus taken as a start without asking how many origins it has | the dissenter is made to prove; the majority is not | |\n| **Identity, ego** | | a challenge to a belief one owns is met as an attack on the self | the bar for *I was wrong* rises with what the belief protects |\n| **Status quo** | | the change must prove; the default need not | inertia as an unstated cost of changing |\n| **Bias accusation** | its own start is low by kind: naive realism predicts it on both sides whatever the reasoning | moves the obligation from the claim to the motive (Bulverism) | |\n\nWhat the graph can read without accusing anyone, given the rulings already made:\n\n- **The start**: a stated reference class anyone can attack (Q35's path), and a lens as a set of starts,\n  with *robust* and *prior-dependent* readings (starting-values note § 6.6). Ideology becomes visible as\n  which class a lens accepts.\n- **The obligation**: the same for every claim (Q34, AL2). Whether one side was scrutinised more shows as\n  uneven critical-question coverage per debate (AN7). Agreement is counted by independent origin (Q40), so a\n  majority is not a reason by its number.\n- **The threshold**: never stored on a claim, so motivated thresholds never enter strength. They show only\n  where people act on the graph, in their votes: the belief-bias measure across raters (ledger G4), never\n  per person (E7).\n\n## 7. Recommendations (proposed, unruled), in order of dependency\n\n1. **AZ1, three burdens.** A sentence that asserts a burden of proof is filed as one or more of three\n   claims: about the claim's kind (its start), about an answer or search that is owed (the obligation), or\n   about which error costs more in a stated decision (the threshold). The `ignorance` scheme's\n   `burden_correct` question asks which of the three is meant.\n2. **AZ2, no free default.** No claim stands by status quo, majority or authority. A default (the status\n   quo, a presumption, precaution or its reverse) enters as a claim about a decision's error costs, and can\n   be argued like any claim.\n3. **AZ3, extraordinary is a start.** A demand for extraordinary evidence is filed as a claim about the\n   claim's kind (its base rate, with its strength), never as a refutation. So a true but unusual claim rises\n   as its evidence is counted, and the demand cannot be raised by status.\n4. **AZ4, motivated skepticism read where it shows.** Selective scrutiny is read by the per-debate coverage\n   count (AN7); head-counting by the independence reading (Q40); a moved threshold by belief bias across\n   raters (G4). The graph adds no per-person measure.\n5. **AZ5, a bias accusation starts low.** Once starts exist, a bias accusation's start comes from its kind,\n   which naive realism makes uninformative by default; with E21's fourth question it then moves a claim only\n   as far as its own argued evidence does.\n\n**His 2013 question** is answered by Q34 and AL2 as ruled, and gets its home here (§ 4 case 3): attacking\nthe *yes* side's arguments leaves *yes* at *nothing stands*; to make it probably false, the *no* side must\nargue.\n\n## 8. What only he decides\n\n- Whether *burden of proof* is filed as three kinds of claim (AZ1), and whether a default may ever stand\n  unargued (AZ2).\n- Whether a decision's thresholds and error costs belong in Deliberus at all, as weighing claims, or stay\n  outside it.\n\n## Sources\n\nRead-depth: abstracts and secondary pages; the two reports give each source's link and tag.\n\n- Hahn & Oaksford, *The burden of proof and its role in argumentation*, Argumentation 21 (2007)\n- Prakken & Sartor on burdens and presumptions; Williams 2003 on the tactical burden\n- Ullmann-Margalit, *On presumption*, J Phil 80 (1983); Whately (1828); Kauffeld (2003)\n- Kaplan, *Decision theory and the factfinding process*, Stan. L. Rev. 20 (1968); Blackstone's ratio\n- Hume (1748), Laplace (1814), Truzzi, Sagan; Russell's teapot\n- Wingspread Statement (1998); REACH; Sunstein, *Laws of Fear*\n- Ditto & Lopez, *Motivated skepticism*, JPSP 63 (1992); Dawson, Gilovich & Regan, PSPB 28 (2002); Taber &\n  Lodge, AJPS 50 (2006); Samuelson & Zeckhauser (1988); Abelson (1986); Kaplan, Gimbel & Harris (2016);\n  Lyons et al. (2021) and Hanselman et al. on self-affirmation; Bovens & Hartmann (2003); Bikhchandani,\n  Hirshleifer & Welch (1992); Pronin, Gilovich & Ross (2004); Kennedy & Pronin (2008)\n- Cases: Kelsey and thalidomide; Semmelweis; Wegener; Marshall and Warren; cold fusion; Bem, Ritchie et al.\n  2012, Galak et al. 2012; the 1969 tobacco proposal; Kehoe and Patterson; the Simpson trials\n\n## Cross-references\n\n- [starting-values-and-the-problem-of-the-prior.md](starting-values-and-the-problem-of-the-prior.md) (the start)\n- [relevance-and-the-two-factors.md](relevance-and-the-two-factors.md) § 7 (the obligation, ruled as Q34)\n- [causes-of-belief-and-the-three-minds.md](causes-of-belief-and-the-three-minds.md) (causes as attack, stake, pattern)\n- [cross-source-premise-run.md](cross-source-premise-run.md) (burden as a frame-level crux)\n- [what-a-case-counts-for.md](what-a-case-counts-for.md) (a mortality drop as a count)\n- [how-independent-reasons-add.md](how-independent-reasons-add.md) (consensus by origin)\n- Ledger: [session26-open-rulings.md](session26-open-rulings.md) Arc AZ\n"}