{"path":"research/what-counts-as-evidence/philosophy-of-evidence.md","content":"# The Philosophy of Evidence and Confirmation — what it offers a claim-graph argumentation platform\n\n*Research report, 2026-09-26. Written against `research-briefing.md` (Deliberus: one Claim node type, typed edges, Walton schemes with critical questions, QEM gradual semantics, direction/weight split after Keynes, horizon kinds, methodology ladder unratified).*\n\n**Verification convention.** Attributions marked **[verified]** were checked this session against a primary or authoritative secondary source (SEP entry, publisher/journal page, PhilPapers record, or the article itself). Items marked **[from memory — verify before citing outward]** are standard in the literature and I am confident of them, but I did not re-check the exact wording, page or number this session. Nothing here is fabricated; where I am unsure of a detail I say so.\n\n---\n\n## 0. The short version (for a reader with ten minutes)\n\n1. **Philosophy has no single concept of evidence; it has a bundle of roles** (Kelly, SEP \"Evidence\"): what justifies belief, what a rational thinker respects, a sign or symptom that indicates truth, and a public, neutral arbiter between disputants. The last role is the one Deliberus is built for, and it is the one most philosophers think evidence can only partly fill, because what counts as evidence is itself contested (theory-ladenness; the FRI adversarial collaboration's \"different standards for what counts as evidence\" is exactly this).\n2. **The quantitative core that is closest to consensus is: evidence bears on a hypothesis *comparatively*, through the likelihood ratio P(E|H)/P(E|¬H).** The difference, ratio and log-likelihood-ratio measures of confirmation disagree on rankings; the log-likelihood ratio (Good's \"weight of evidence\") is the one that *adds* across conditionally independent items, which is why it is the natural currency for accumulation. Counting answers is none of these measures.\n3. **Two quantities, not one:** the *balance* of evidence (which way it points, how far) and the *weight* of evidence (how much has been examined, how resilient the verdict is to more). Keynes named the distinction in 1921; Joyce (2005) and Skyrms (1977/1980, \"resilience\") formalised it. Deliberus's Aug 2026 direction/weight split is squarely in this tradition — but its current weight (a raw count of non-neutral inputs, non-saturating, independence-blind) is not what any of those authors meant.\n4. **Independence is not a nicety; it is the variable.** Agreeing reports multiply evidential force *only* when their errors are independent given the truth (Condorcet; Bovens & Hartmann 2003; Easwaran et al. 2016 on \"synergy\"). With shared origins the same reports can be worth barely more than one. And the variety-of-evidence thesis — \"diverse evidence beats repeated evidence\" — is *false in general* (Bovens & Hartmann 2003; Claveau 2013; Osimani & Landes 2023): whether variety helps depends on how the sources can fail. So independence must be *represented*, not presumed.\n5. **\"There is no atomic evidence\"** is well-defended philosophy (Sellars's attack on the Given; Quine–Duhem holism; Haack's foundherentism) *if* it means \"no evidence is immune to challenge\". It overreaches if it means \"nothing plays a foundational role\". Haack's crossword is the best model for the platform: experiential clues are not beliefs and are not infallible, but they are a *different kind of input* from the intersecting entries, and a warrant structure that treats them as just more entries loses what makes evidence evidence.\n6. **The founder's \"same evidence → same probability\" hope is the Uniqueness Thesis.** It has serious defenders (White 2005; Feldman; Christensen) and serious critics (Kelly 2013/14; Schoenfield 2014). The defensible platform version is *not* \"everyone must end at the same number\" but \"differences in conclusion must be traceable to a stated difference in prior, standard, or evidence\" — which is the *attribution* requirement the project already holds, and what permissivists themselves concede must be articulable.\n7. **Counting CQ answers is not defensible as a measure of evidential strength**, and every tradition surveyed agrees for different reasons (Bayesian: diagnosticity not number; Mayo: severity not number of passes; Keynes: weight is not balance; testimony literature: three answers from one account are one witness). Counting *is* defensible as a crude, honestly-labelled measure of *scrutiny coverage* — how much of the question space has been visited — which is a weight-like quantity, not a balance-like one.\n\n---\n\n## 1. What evidence is\n\n### 1.1 The roles of the concept (Kelly)\n\nThomas Kelly's SEP entry \"Evidence\" (first published 2006, substantive revision 28 July 2014) **[verified]** organises the concept around four theoretical roles:\n\n- **Evidence as that which justifies belief.** Evidence is what confers epistemic justification.\n- **Rational thinkers respect their evidence.** Rationality consists (at least partly) in proportioning belief to evidence — Hume's \"a wise man proportions his belief to the evidence\".\n- **Evidence as a guide to truth: evidence as sign, symptom, or mark.** Smoke is evidence of fire; Koplik spots of measles. This is the *indicator* role, which is factive-leaning and need not involve any believer.\n- **Evidence as neutral arbiter: objectivity, publicity, intersubjectivity.** Evidence is what settles disputes between people who disagree, precisely because it is (supposedly) accessible to all parties and neutral between their theories.\n\nKelly's central observation is that **these roles pull apart**. What an individual's justified beliefs rest on (role 1) may be private (her experiences); what settles a public dispute (role 4) must be public. What is a reliable sign (role 3) may be unknown to anyone. No single ontology of evidence — sense data (Russell), stimulations of sensory receptors (Quine), observation sentences (the positivists), known propositions (Williamson) — satisfies all four roles **[verified: the list is Kelly's]**.\n\n*Consensus/contested.* That the roles exist and can conflict: broadly accepted. Which role is primary: contested.\n\n*Implication for Deliberus.* The platform is building the **neutral-arbiter** role at scale, and philosophers' main doubt about that role (§7 below, theory-ladenness) is the doubt Deliberus should expect to meet in practice. Its design answer — make the standards themselves claims, attributable and attackable — is the right response to Kelly's tension: where evidence cannot be neutral, the *disagreement about what counts* can at least be public. The FRI finding (\"different standards for what counts as evidence\") is then not a failure of the arbiter but a datum the arbiter should record as its own region of the graph.\n\n### 1.2 Propositional vs non-propositional evidence\n\n**Propositionalism**: only propositions (or true propositions, or known propositions) are evidence, because evidence must be able to stand in logical and probabilistic relations (entail, be inconsistent with, raise the probability of) and must be usable as a premise.\n\n**Williamson's E=K** (*Knowledge and its Limits*, OUP 2000, ch. 9) **[verified via SEP summary]**: one's evidence is all and only the propositions one knows. Consequences: evidence is factive (false propositions are not evidence); evidence is not always luminous (one is not always in a position to know what one's evidence is); and the view underwrites a knowledge-first epistemology in which conditioning on one's evidence is conditioning on one's knowledge.\n\n**Non-propositional (experiential/phenomenal) evidence**: **Conee & Feldman**, \"Evidentialism\" (*Philosophical Studies* 48, 1985) and *Evidentialism: Essays in Epistemology* (OUP 2004) **[from memory — verify]**: a doxastic attitude D toward p is justified for S at t iff having D toward p fits the evidence S has at t. They are *mentalists*: evidence consists of mental states, including experiences, so two subjects alike in all mental respects are alike in justification. **Phenomenal conservatism** (Michael Huemer, *Skepticism and the Veil of Perception*, 2001; \"Compassionate Phenomenal Conservatism\", *PPR* 74, 2007) **[from memory — verify]**: if it seems to S that p, then, in the absence of defeaters, S thereby has at least some justification for believing p. Seemings — perceptual, memorial, intellectual, *moral* — are defeasible evidence.\n\n*Consensus/contested.* Deeply contested. E=K is influential but minority; mentalist evidentialism and phenomenal conservatism have large followings; many epistemologists are pluralists (propositional evidence for public reasoning, experiential grounds for individual justification).\n\n*Implications for Deliberus.*\n- A **claim graph is necessarily propositionalist** about what it can store: only propositions sit at the tail of a SUPPORTS edge. This is not a defect but it has a cost the corpus already half-names: first-person *felt* experience (the incorrigible part of impact testimony) enters the graph only as the proposition *\"S reports that it seems to S that p\"* — a REPORTS-style object — and its evidential force for p depends on a further premise about the reliability of such seemings. Phenomenal conservatism is the philosophical warrant for giving such reports *some* default weight rather than zero; Conee & Feldman are the warrant for insisting that what is stored is not the experience itself.\n- Williamson's non-luminosity point applies to the platform: *a contributor may not know what their evidence is*. The design already allows others to decompose one's claims; E=K is a reason that should include surfacing premises the author did not know they were relying on.\n\n### 1.3 Achinstein's four concepts and his critique of \"increase in probability\"\n\nPeter Achinstein, *The Book of Evidence* (OUP 2001) **[verified: four concepts, definitions via potential evidence]**:\n\n- **Potential evidence**: e is potential evidence that h iff (roughly) e is true, e does not entail h, P(h | e) > ½, and — crucially — the probability that there is an *explanatory connection* between h and e, given e, is > ½. Potential evidence is objective and does not depend on anyone's epistemic situation.\n- **Veridical evidence**: potential evidence where h is also true and there is in fact an explanatory connection. It is \"a good reason to believe h\" in a factive sense.\n- **ES-evidence** (epistemic-situation evidence): e is veridical evidence relative to a given epistemic situation — what someone in that situation would be justified in taking as evidence.\n- **Subjective evidence**: what a particular person takes to be evidence.\n\nAchinstein's critique of the Bayesian **increase-in-probability** criterion (e is evidence for h iff P(h|e) > P(h)) **[verified in outline via SEP \"Evidence\"; the examples from memory]**: it is **neither necessary nor sufficient**. Not sufficient: learning that a lottery ticket-holder bought one ticket rather than none raises the probability he wins from 0 to one-in-a-million, but is not *evidence that he will win*. Not necessary: evidence can be good evidence for h even when it does not raise h's probability relative to background already containing strong evidence (e.g., a second, redundant confirming test). Hence his threshold (P > ½) plus the explanatory-connection requirement.\n\n*Consensus/contested.* Most Bayesian philosophers of science reply that Achinstein runs together two concepts — *incremental confirmation* (raising) and *absolute confirmation* (being high) — which the Bayesian already distinguishes (see §2.3). The explanatory-connection requirement is more distinctive and less widely adopted, though it resonates with IBE (§3.6) and with Russo–Williamson (§3.8).\n\n*Implications for Deliberus.*\n- The lottery case is a direct warning about **what a SUPPORTS edge should mean**. An edge licensed only because E makes H \"a bit more likely\" will be pervasive and nearly worthless; the platform needs either a relevance threshold or a measure (likelihood ratio) that displays *how much*.\n- Achinstein's explanatory-connection condition maps onto a critical question the platform already asks in several schemes (\"is there a mechanism/link?\"). It is a principled reason to make \"explanatory connection\" a CQ of the generic *evidence-to-hypothesis* scheme and not only of causal ones.\n- His four concepts give Deliberus a clean vocabulary for its own provenance problem: an **extracted SUPPORTS edge is subjective evidence** (someone took it as evidence), and becomes a candidate for **potential evidence** only after its CQs are examined. That is the machine-proposes / human-examines boundary restated in the philosophy of evidence.\n\n---\n\n## 2. Confirmation theory\n\nThe SEP entry \"Confirmation\" (Vincenzo Crupi; first published 30 May 2013, substantive revision 4 Aug 2025) **[verified]** is the best current map.\n\n### 2.1 Hypothetico-deductivism (HD)\n\n**Definition [verified]**: e HD-confirms h relative to background k iff h ∧ k ⊨ e and k ⊭ e. Evidence confirms a hypothesis when the hypothesis (with auxiliaries) predicts it.\n\n**Problems [verified]**: the tacking/irrelevant-conjunction problem (if h entails e, so does h ∧ x for any irrelevant x); failure of the predictive inference condition; and **radical underdetermination via auxiliaries** — for any contingent h and e there is some k making h ∧ k ⊨ e. HD is also purely qualitative: it gives no degrees.\n\n*Implication.* Every Walton \"evidence-to-hypothesis\" edge in Deliberus is HD-shaped, and inherits the auxiliary problem: the edge is only as good as the unstated k. This is the formal reason that implicit-premise recall is (as the corpus found empirically) the pipeline's most important quality axis — the auxiliaries *are* the implicit premises.\n\n### 2.2 Hempel's paradoxes\n\nCarl Hempel, \"Studies in the Logic of Confirmation\" (*Mind* 54, 1945) **[verified via SEP]**. The **ravens paradox**: \"All ravens are black\" is logically equivalent to \"all non-black things are non-ravens\"; by Nicod's criterion a green apple confirms the latter, so by the equivalence condition it confirms the former. The Bayesian resolution (standard since Good, Mackie, Horwich, Fitelson & Hawthorne) is that the apple confirms, but by a *negligible* amount, because the population of non-black things is vast relative to ravens. Goodman's **grue/blite** paradox shows that which predicates are projectible is not given by logic.\n\n*Implication.* The Bayesian answer to the ravens is the single most important lesson for a counting system: **confirmation comes in degrees, and a system that treats every confirming instance as one unit will be led by the paradox**. A green apple and a black raven are each \"one answer\" in a count.\n\n### 2.3 Bayesian confirmation: incremental vs absolute\n\n**[verified]** Two notions:\n- **Absolute (firmness)**: e confirms h iff P(h | e ∧ k) is high (e.g. > ½).\n- **Incremental (relevance)**: e confirms h iff P(h | e ∧ k) > P(h | k).\n\nCarnap (*Logical Foundations of Probability*, 1950; preface to 2nd ed. 1962) distinguished them after confusing them in the first edition **[from memory — verify]**. The SEP stresses the \"complementary evidence\" principle: under incremental confirmation, e confirms h iff ¬e disconfirms h — no \"heads I win, tails you lose\". (This principle is the Bayesian root of \"absence of evidence is evidence of absence\", §5.)\n\n*Implication.* Deliberus's badge conflates these. A claim's σ near 0.58 after one well-answered edge is being read both as \"how firmly held\" (absolute) and \"how much has been added\" (incremental). The direction/weight split partly separates them; making explicit that σ is an *incremental* reading relative to a neutral 0.5 prior — not a probability of truth — would close the confusion old docs created by calling QEM \"Bayesian\".\n\n### 2.4 Measures of confirmation\n\n**[verified formulas, SEP]**:\n- **Difference**: D = P(h|e) − P(h).\n- **(Log-)ratio**: R = log[P(h|e)/P(h)].\n- **Log-likelihood ratio**: L = log[P(e|h)/P(e|¬h)].\n- Others: Kemeny–Oppenheim's normalised likelihood measure; Crupi & Tentori's relative-distance measure Z.\n\n**Measure sensitivity.** Branden Fitelson, \"The Plurality of Bayesian Measures of Confirmation and the Problem of Measure Sensitivity\" (*Philosophy of Science* 66, 1999, S362–S378) **[from memory — verify page range]** showed that many classic Bayesian \"solutions\" (to the ravens, to irrelevant conjunction, to variety of evidence) hold under some measures and fail under others. So which measure is used is substantive.\n\n**Why Good and Fitelson favour the log-likelihood ratio.** I. J. Good (*Probability and the Weighing of Evidence*, 1950; \"Weight of evidence: a brief survey\", 1985 **[from memory — verify]**) called log[P(e|h)/P(e|¬h)] the **weight of evidence** (Turing's \"decibans\" at Bletchley). Its properties:\n1. **Additivity for conditionally independent evidence**: if e1 and e2 are independent given h and given ¬h, then L(h, e1 ∧ e2) = L(h, e1) + L(h, e2). Evidence accumulates by addition, and the posterior log-odds = prior log-odds + Σ weights. **This is the formal justification for \"summed offsets\" in the corroborative channel — and it shows exactly the condition under which summing is licit: conditional independence.**\n2. **Prior-independence of the evidence term**: the LR depends only on the likelihoods, so two people with different priors can agree on *how much* the evidence says, while disagreeing on the posterior. (This matters for §4.)\n3. **Symmetry**: L(h,e) = −L(¬h,e).\n\nFitelson (\"Likelihoodism, Bayesianism, and Relational Confirmation\", *Synthese* 156, 2007 **[from memory — verify]**) argues that among the Bayesian measures the LR measure best satisfies a set of adequacy conditions and connects Bayesianism to likelihoodism. The SEP records defenders of L (Park 2014, Pruss 2014, Vassend 2015), of D (Huber 2008), of Z (Crupi & Tentori), and **pluralists** (Hájek & Joyce 2008: different measures capture distinct, complementary notions) **[verified]**.\n\n*Consensus/contested.* That measures differ and it matters: consensus. That the LR is *the* right measure: contested but the most defended for the *accumulation* purpose; its additivity-under-independence is a theorem.\n\n*Implications for Deliberus.*\n- The QEM offset added per answered CQ (+0.15, capped) has **no likelihood interpretation**. If the corroborative channel is to be \"≈ Good's additive log-odds\" (as the briefing says), each input should carry an estimate of its *own* LR (diagnosticity), not a fixed increment, and the sum is only licit over conditionally independent inputs.\n- A practical, cheap route: ordinal diagnosticity bands (Royall-style, §2.5) rather than point LRs, attached as claims (\"this test would rarely come out this way if H were false\") — which are then themselves attackable. That keeps \"propose, never assert\" while giving the arithmetic something to add.\n\n### 2.5 Likelihoodism, the law of likelihood, the likelihood principle\n\n**Law of Likelihood** (Hacking 1965, *Logic of Statistical Inference* **[from memory]**; Edwards 1972; Royall 1997): e favours h1 over h2 iff P(e|h1) > P(e|h2), and the degree of favouring is the ratio P(e|h1)/P(e|h2).\n\n**Royall's three questions** — Richard Royall, *Statistical Evidence: A Likelihood Paradigm* (Chapman & Hall, 1997) **[verified that the three questions exist; the mapping below is the standard reading]**:\n1. What should I **believe**, now that I have this observation? (Bayesian posterior; needs a prior.)\n2. What should I **do**? (Decision theory / Neyman–Pearson; needs losses.)\n3. What does this observation **tell me about A versus B** — how should it be interpreted *as evidence*? (Likelihood ratio; needs neither prior nor losses.)\n\nRoyall's claim is that statistics had conflated these, and only the likelihood ratio answers question 3. He also offers benchmark LR values — commonly quoted as **8 = \"fairly strong\"** and **32 = \"strong\"** evidence **[from memory — verify exact wording in Royall 1997, ch. 1]**.\n\n**The likelihood principle** (Birnbaum 1962, *JASA* 57 **[from memory]**): all the evidential import of data about parameters, given a model, is in the likelihood function. Accepted by Bayesians and likelihoodists; *rejected* by error statisticians, because it makes stopping rules and sampling plans evidentially irrelevant.\n\n*Implications for Deliberus.*\n- **Royall's trichotomy is the cleanest possible statement of what Deliberus does and does not do.** The platform answers Q3 (what the evidence says), refuses to answer Q1 for anyone (no verdicts; priors are the user's), and leaves Q2 to people. This is a better public articulation of \"maps reasoning, does not endorse conclusions\" than anything in the current copy.\n- It also locates the *normative* horizon kinds: fittingness and weighing live in Q2 (and Q1 priors), where no amount of Q3 settles things — consistent with the oracle gradient.\n- The likelihood principle debate matters practically: if the platform follows the likelihood principle, *how the evidence was gathered* (optional stopping, p-hacking, publication bias) is irrelevant; if it follows Mayo, it is central. The gaps list (publication bias, replication status, funding conflict) are all error-statistical concerns, and Deliberus should adopt Mayo's side here (§2.6).\n\n### 2.6 Error statistics and severe testing (Mayo)\n\nDeborah Mayo, *Error and the Growth of Experimental Knowledge* (Chicago, 1996); *Statistical Inference as Severe Testing* (CUP, 2018) **[verified]**. **Severity requirement**: data x provide evidence for H only if H passes a test that would, with high probability, have *failed* H were H false. Formally: H passes a severe test with x if (S-1) x agrees with H and (S-2) with very high probability the test would have produced a result that accords less well with H, were H false **[verified in substance]**. \"If little or nothing has been done to rule out flaws in inferring a claim, then it has not passed a severe test.\"\n\nSeverity is a property of the **test procedure**, not only of the data — so selection effects, data-dredging, optional stopping and publication bias *lower* severity even when the data look the same. This is the error statistician's objection to the likelihood principle.\n\n*Consensus/contested.* Mayo's framework is influential among philosophers of statistics and some methodologists; Bayesians contest it; the practical point (selection effects matter) is broadly accepted by the replication-crisis literature.\n\n*Implications for Deliberus.*\n- **Critical questions are severity probes.** A CQ answered \"yes, the flaw was looked for and not found\" is evidence *only if* the search could have found it. So an answer's value depends on the probe's power — which is information the current count ignores. The design move: each CQ answer carries (as an attackable claim) *how the check was done*, and \"answered by assertion\" is distinguished from \"answered by a check that could have failed\".\n- Publication bias, replication status, preregistration and funding conflicts are all **severity-lowering properties of the procedure** and belong as claims attached to the *edge* (the inference) rather than to the conclusion. This is a principled home for several listed gaps.\n- Severity gives a principled reading of why a meta-analysis ranks high on a methodology ladder *and* why the ladder must be per-claim: severity is a property of the test of *that* claim, not of the document.\n\n### 2.7 Popper: corroboration vs probability\n\nKarl Popper, *Logik der Forschung* (1934) / *The Logic of Scientific Discovery* (1959) **[from memory]**: theories cannot be made probable by evidence (a universal law has logical probability zero); what they can be is **corroborated** — having survived severe attempts at refutation. Corroboration is backward-looking (a report on the test record), not a probability, and *less* probable (more content-rich) theories are *more* testable and so more corroborable. Mayo's severity is the modern descendant.\n\n*Implication.* Popper's distinction is exactly Deliberus's need for a number that says \"this has survived challenge\" distinct from \"this is probably true\". The current badge fuses them. A **corroboration record** — which attacks were mounted, which failed, which were strong — is a first-class object Deliberus already half-has (attack edges with their own strength); it should be displayable as such.\n\n### 2.8 Keynes's weight of arguments and its formalisations\n\nJ. M. Keynes, *A Treatise on Probability* (1921), ch. VI, \"The Weight of Arguments\" **[from memory — chapter number verify]**: new evidence may raise or lower a probability, but it always increases the **weight** — the amount of relevant evidence on which the probability rests. Probability measures the *balance* between favourable and unfavourable evidence; weight measures their *absolute amount*. Keynes doubted weight was numerically measurable and was unsure how it should affect action.\n\n**Joyce**, \"How Probabilities Reflect Evidence\" (*Philosophical Perspectives* 19, 2005, 153–178) **[verified]**: distinguishes **balance**, **weight**, and **specificity** of evidence; argues precise credences cannot represent weight, and that imprecise (set-valued) credences or resilience can. Weight shows up in how *stable* a credence is under further evidence.\n\n**Skyrms**, **resilience** (*Causal Necessity*, Yale 1980; earlier papers c. 1977) **[from memory — verify]**: a probability is resilient to the extent it would not change much under conditioning on further plausible evidence. Weight = resilience.\n\n**Carnap's requirement of total evidence** (\"On the Application of Inductive Logic\", *PPR* 8, 1947; *LFP* 1950) **[from memory — verify]**: in applying inductive logic, one must use the total evidence available. **Good**, \"On the Principle of Total Evidence\" (*BJPS* 17, 1967) **[from memory]**: proves that for a Bayesian decision-maker, cost-free evidence never lowers expected utility — so gathering more evidence is rational when free. (Good's theorem is the formal backbone of \"ask when information gain exceeds cost\" — already in the corpus via active inference.)\n\n*Consensus/contested.* That balance and weight differ: near-consensus. How to formalise weight: contested (imprecise probability, resilience, second-order probability, Keynesian non-numerical weight). A 2025 *Erkenntnis* paper (\"A New Look at Keynesian Weight of Evidence\") shows the topic is still live **[existence verified; content not read]**.\n\n*Implications for Deliberus.*\n- The Aug 2026 direction/weight split is **correctly grounded**; the defect is the operationalisation. Weight as a raw count of non-neutral inputs, **non-saturating and independence-blind**, is not Keynes's weight (amount of *relevant* evidence), nor Joyce's (breadth/specificity), nor Skyrms's (resilience). Three concrete upgrades in increasing sophistication:\n  1. **Count only independent inputs** (collapse inputs sharing an origin; §3.5).\n  2. **Saturate**: weight should approach a ceiling as the *derivable question space* (the CQ set for the schemes present) is covered — \"coverage of the requirement space\" (already the corpus's own term in the derivability doc) is Joyce's specificity/comprehensiveness in practice.\n  3. **Resilience probe**: report how far σ would move if the single most influential input were removed or reversed. That is Skyrms's resilience, computable deterministically by the hinge machinery already built.\n- \"An unexamined claim carries no number\" is a Keynesian move (zero weight ≠ probability ½). Keep it.\n- Carnap's total-evidence requirement is the formal name for the platform's **coherent-absence** worry: a verdict computed from less than the total available evidence is rationally defective even if every step is valid. It justifies ingestion-balance instrumentation.\n\n---\n\n## 3. Evidence aggregation and structure\n\n### 3.1 Confirmation holism (Duhem–Quine) and its limits\n\nPierre Duhem, *La théorie physique* (1906): a physical hypothesis is never tested in isolation, only together with auxiliary assumptions, so a failed prediction refutes the conjunction, not a specified member. W. V. O. Quine, \"Two Dogmas of Empiricism\" (1951): \"our statements about the external world face the tribunal of sense experience not individually but only as a corporate body\" **[from memory; quote standard]**.\n\n**Limits** (widely agreed): holism does not imply that blame is *arbitrarily* distributable. Bayesians (Dorling 1979, Howson & Urbach, Strevens 2001 \"The Bayesian treatment of auxiliary hypotheses\", *BJPS* 52 **[from memory — verify]**) show that, given priors, a refutation of h ∧ a lowers the probability of the less entrenched conjunct more; scientists localise blame in practice through independent tests of auxiliaries. Quine himself later moderated to a \"moderate holism\".\n\n*Implications.* \"Link, never divide\" is sound; but the Bayesian limit shows **blame allocation is computable once auxiliaries are explicit and have independent support**. That gives the decomposition programme a concrete payoff: making auxiliaries explicit claims is what turns holism from a counsel of despair into a distribution rule. (Consistent with the weakest-link result: a refuted conjunction transmits most blame to its weakest-supported part.)\n\n### 3.2 Variety of evidence\n\n**Variety-of-evidence thesis (VET)**: ceteris paribus, evidence from diverse, independent lines of inquiry confirms more than an equal amount of evidence from one line. Intuitively central to science (Whewell's consilience; triangulation; robustness).\n\n**It is not a theorem.** **[verified]**\n- **Bovens & Hartmann**, *Bayesian Epistemology* (OUP 2003); also \"The Variety-of-Evidence Thesis and the Reliability of Instruments: A Bayesian-Network Approach\" (PhilPapers record verified): in a Bayesian network where each instrument is either reliable (tracks the truth) or a randomiser, **two positive reports from the *same* instrument can confirm more than one each from two independent instruments**. Mechanism: two concordant positives from one instrument are themselves evidence that the instrument is reliable, which raises the credibility of both reports; with two instruments that coherence-to-reliability boost is split.\n- **Claveau**, \"The Independence Condition in the Variety-of-Evidence Thesis\" (*Philosophy of Science* 80, 2013, 94–118): models unreliability as deterministic *bias* (a positively biased source always reports positive) and refines the notion of source independence; VET's validity turns on how independence and unreliability are modelled.\n- **Claveau & Grenier**, \"The variety-of-evidence thesis: a Bayesian exploration of its surprising failures\" (*Synthese*, 2019; online 2017).\n- **Landes**, \"The variety of evidence thesis and its independence of degrees of independence\" (*Synthese*, 2020/2021) and **Osimani & Landes**, \"Varieties of Error and Varieties of Evidence in Scientific Inference\" (*BJPS* 74, 2023): with probabilistic bias, VET can fail in the *opposite* direction too, and whether it holds depends on the comparison of Bayes factors for (say) sponsored versus non-sponsored studies.\n\n*Consensus.* VET is **conditionally** true: it holds when source-failure modes are independent and a failure would not produce the same concordant output; it fails when concordance is itself evidence of (un)reliability in ways that favour replication. This is now the consensus among formal epistemologists who work on it.\n\n*Implications.*\n- The corpus's rule \"**a known shared source becomes a node, never a blind discount**\" is exactly what this literature recommends: model the *source* (instrument, lab, sponsor, dataset) as a node, with its reliability as an attackable claim, and let the network compute whether concordance helps. A blanket \"diverse = better\" multiplier would be wrong.\n- Concretely: sources need **reliability** as a property distinct from any claim's truth (a listed gap), and reports need **provenance to a source node** so that (a) shared-origin correlation is visible and (b) concordant reports can update reliability.\n\n### 3.3 Robustness analysis\n\nRichard Levins, \"The Strategy of Model Building in Population Biology\" (*American Scientist* 54, 1966): \"our truth is the intersection of independent lies\". William Wimsatt, \"Robustness, Reliability, and Overdetermination\" (1981, in Brewer & Collins eds., *Scientific Inquiry and the Social Sciences*). Michael Weisberg, \"Robustness Analysis\" (*Philosophy of Science* 73, 2006) **[all from memory — verify]**: a result is robust when it holds across models making different idealising assumptions; robustness is evidence the result depends on a shared *core* rather than on the idealisations. Orzack & Sober (1993) objected that robustness among false models does not in itself confer truth; Weisberg's reply requires the models share a common causal core and that the result be derivable from it.\n\n*Implication.* Robustness is the model-side twin of variety of evidence, and it carries the same independence condition. For the platform: the claim \"the result holds across methods A, B, C\" is valuable **only if** the methods' failure modes are distinct — which should be a CQ on any \"consilience\"/\"robustness\" edge.\n\n### 3.4 Consilience of inductions (Whewell)\n\nWilliam Whewell, *The Philosophy of the Inductive Sciences* (1840; 2nd ed. 1847) **[from memory]**: consilience occurs when an induction from one class of facts unexpectedly explains a different class — the strongest mark of truth. Modern readings: Thagard's explanatory coherence; Forster's unification; Myrvold (\"A Bayesian account of the virtue of unification\", *Phil Sci* 70, 2003 **[from memory]**) shows unification can be cashed out as mutual information between evidence items given the hypothesis.\n\n*Implication.* Consilience is the structural signal a graph can *compute*: a claim supported by chains originating in disjoint regions of the graph (different concept clusters, different source clusters) is consilient in Whewell's sense. That is a buildable, deterministic \"cross-region support\" indicator — distinct from, and more honest than, raw support count. It should be displayed as a structural fact, not added to strength.\n\n### 3.5 Coherence, the impossibility results, independent witnesses\n\n**Coherence and truth.** Bovens & Hartmann (2003) and Erik Olsson, *Against Coherence: Truth, Probability, and Justification* (OUP 2005) **[from memory — verify]** proved **impossibility results**: there is no coherence measure such that, ceteris paribus, more coherent information sets are always more likely to be true (given plausible conditions on the measure and on witness reliability). Coherence is truth-conducive only under special conditions — notably that the sources are *independent* and *partially reliable*. With fully independent, somewhat reliable witnesses, agreement is powerful evidence (C. I. Lewis's \"congruence of independent witnesses\", *An Analysis of Knowledge and Valuation*, 1946 **[from memory]**); without independence it is nearly worthless.\n\n**Condorcet jury theorem** (1785): if voters are independent and each more likely than not to be correct, majority correctness → 1 as numbers grow. **Correlation problem**: with correlated votes the theorem fails; Ladha (1992) and Dietrich & Spiekermann, \"Independent Opinions? On the Causal Foundations of Belief Formation and Jury Theorems\" (*Mind* 122, 2013) **[from memory — verify]** show that the relevant independence is independence *conditional on the shared evidence/problem*, and that when voters share an evidential basis, the effective number of independent judgments can be tiny.\n\n**Agreement and synergy.** Easwaran, Fenton-Glynn, Hitchcock & Velasco, \"Updating on the Credences of Others: Disagreement, Agreement, and Synergy\" (*Philosophers' Imprint* 16(11), 2016) **[verified]**: under independence, learning that several peers each assign moderately high credence can rationally push one's credence *higher than any of theirs* (\"synergy\") — the multiplicative (log-odds-additive) pooling rule. This is the formally exact version of \"independent corroboration adds\".\n\n*Implications.*\n- The brief's arithmetic fact — **\"three one-word answers from one account = three RCTs\"** — violates every result above. Three reports from one account are, to a first approximation, *one* witness whose reliability is unknown; three RCTs from independent groups with independent failure modes are closer to three independent LRs. Independence must be represented, because **the correct aggregation rule changes with it**: multiplicative (log-odds addition) under independence; something close to \"take one\" under full dependence; intermediate under partial dependence.\n- The coherence impossibility results warn against any \"coherence score\" as a quality signal: a tightly coherent cluster of claims from one source family is not thereby likely true. Coherence is a *diagnostic of consistency*, not of truth.\n- Easwaran et al. give the platform a principled alternative to \"votes are telemetry, never strength\": votes *can* be evidence when voters are independent and calibrated — the reason not to use them is that the platform *cannot currently establish* either condition. That is a better stated reason than a blanket rule, and says what would change it.\n\n### 3.6 Inference to the best explanation (IBE) and Bayes\n\nPeter Lipton, *Inference to the Best Explanation* (Routledge 1991; 2nd ed. 2004): we infer the hypothesis that would, if true, provide the \"loveliest\" (most explanatory) explanation among those \"likeliest\"; loveliness guides judgments of likeliness. Igor Douven, SEP \"Abduction\" and *The Art of Abduction* (MIT 2022) **[from memory — verify]**: IBE can be read as an updating rule that gives an explanatory bonus beyond Bayes; van Fraassen (1989) argued any such bonus makes the agent Dutch-bookable; Douven's simulations show explanationist updating can converge faster to truth (at the cost of more error in some regimes). The **compatibilist** view (Lipton; Okasha 2000; Weisberg 2009) treats explanatory considerations as heuristics for priors and likelihoods rather than a rival rule.\n\n*Implications.* The Walton *abductive* scheme is IBE; its critical questions (\"are there alternative explanations?\", \"how good is this explanation compared with them?\") are exactly Lipton's contrastive structure. IBE's key lesson for a graph: **an abductive edge is comparative** — its strength depends on the *set of rival explanations present*. A claim graph that stores the abductive edge without the rival set cannot compute it; rivals should be first-class sibling claims linked to the same evidence, and \"no rival has been proposed\" should be marked as a coverage gap, not read as strength.\n\n### 3.7 Evidential pluralism: the Russo–Williamson thesis\n\nFederica Russo & Jon Williamson, \"Interpreting Causality in the Health Sciences\" (*International Studies in the Philosophy of Science* 21(2), 2007, 157–170) **[verified]**: to establish a causal claim in medicine one normally needs to establish *both* that cause and effect are appropriately correlated (**difference-making** evidence) *and* that there is a mechanism that can account for the correlation (**mechanistic** evidence). Developed in the EBM+ programme; Clarke, Gillies, Illari, Russo & Williamson, \"Mechanisms and the Evidence Hierarchy\" (*Topoi* 33, 2014) **[from memory]**; Williamson, *Evidential Pluralism in the Social Sciences* (Routledge 2024?) **[title/year unverified]**. Critics (Howick 2011, *The Philosophy of Evidence-Based Medicine*; Broadbent) argue mechanistic reasoning has a poor track record in medicine and that correlation evidence from RCTs can suffice.\n\n*Implications.*\n- A **missing mechanistic-evidence tier** (a listed gap) is not a hierarchy rung but an **orthogonal axis**: RCTs and meta-analyses are difference-making evidence; mechanistic evidence is a different kind whose absence is a distinct weakness. A single methodology ladder (meta-analysis 0.70 … expert opinion 0.30) cannot represent this; a causal claim needs *two slots*, and the scheme for cause-to-effect should carry a CQ \"is there mechanistic evidence?\" separate from \"is there difference-making evidence?\".\n- The ladder itself is the EBM hierarchy that philosophers (Worrall 2002/2007, \"Evidence in medicine and evidence-based medicine\"; Cartwright; Howick) have criticised as ranking *designs* rather than *particular studies' quality for a particular question*. That is the corpus's \"methodology per claim, not per source\" point, independently made in the literature. The ladder should remain unratified.\n\n### 3.8 Cartwright & Hardie: evidence for use\n\nNancy Cartwright & Jeremy Hardie, *Evidence-Based Policy: A Practical Guide to Doing It Better* (OUP 2012) **[from memory — verify]**; Cartwright, \"A philosopher's view of the long road from RCTs to effectiveness\" (*Lancet* 377, 2011). Key claims:\n- **Efficacy vs effectiveness**: an RCT shows \"it worked *there*\" (for that population, in that setting). The policy question is \"it will work *here*\".\n- To get from there to here one needs evidence that the intervention plays the same **causal role** here and that the **support factors** (other conditions needed for it to work) are present here.\n- **Relevance** of evidence is not a property of the study alone but of study + target + argument; a \"hierarchy of evidence\" answers \"is this study internally valid?\" not \"is this evidence for *our* claim?\".\n\n*Implications.* This is the most directly transferable framework for Deliberus's fact layer. Every SUPPORTS edge from a study claim to a policy/prediction claim is an \"it worked there → it will work here\" inference, and should be typed as such with CQs: *same causal structure? support factors present? population similar in the relevant respects?* The current schemes (cause-to-effect, analogy, example) each partly cover this; an explicit **external-validity** scheme (or CQ set) would catch the most common real-world error in evidence use. It also shows why **relevance is an edge property**: the same RCT claim can be strong evidence for one target and weak for another.\n\n---\n\n## 4. Higher-order evidence, disagreement, uniqueness vs permissivism\n\n### 4.1 Higher-order evidence\n\nDavid Christensen, \"Higher-Order Evidence\" (*PPR* 81(1), 2010, 185–215) **[verified]**: evidence about one's own epistemic malfunction — being fatigued, drugged, or learning that competent peers disagree — is distinctive because its import is **agent-relative**: it rationally requires one to \"bracket\" or discount one's first-order reasoning, even if that reasoning was in fact correct. Maria Lasonen-Aarnio, \"Higher-Order Evidence and the Limits of Defeat\" (*PPR* 88(2), 2014, 314–345) **[verified]**: accommodating such defeat requires a two-tier theory of justification and generates puzzles in which correct epistemic rules issue conflicting recommendations; she is sceptical that higher-order evidence always defeats. The SEP has an entry \"Higher-Order Evidence\" (existence verified).\n\n*Implications.*\n- The platform's own instruments generate higher-order evidence **about the machine passes**: the measured 76% misfiling rate of decomposition edges (⚠ corrected 2026-10-07: 84 of those 197 were the automatic linker's premise links, stored backwards, which could not have been judged parts; among the other 113, 59% are not parts: `auto-connect-upgrade.md` § 2c) is higher-order evidence that *every* decomposition edge's classification is less trustworthy than its local confidence suggests. Christensen's point says that evidence should discount *all* such outputs, not only the ones shown wrong — i.e., **pass-level calibration should feed edge-level credibility**. The \"machine work as a standing property\" watermarking proposal is how this would be implemented.\n- For human contributors: evidence of bias (funding conflict, motivated position) is higher-order evidence about a *source's* reasoning — it belongs on the source/inference, not on the conclusion (so it does not become ad hominem against the claim). Walton's *bias* scheme already has this shape.\n\n### 4.2 Peer disagreement\n\n**Conciliationism** (Christensen 2007, \"Epistemology of Disagreement: The Good News\", *Phil Review* 116; Elga 2007, \"Reflection and Disagreement\", *Noûs* 41 — the \"Equal Weight View\"; Feldman 2006 **[from memory]**): on learning that an epistemic peer, with the same evidence, disagrees, one should move substantially toward the peer. **Steadfastness** (Kelly 2005, \"The Epistemic Significance of Disagreement\", *Oxford Studies in Epistemology* 1; Kelly 2010 \"Peer Disagreement and Higher-Order Evidence\" — \"Total Evidence View\"; Lackey's justificationist view **[from memory]**): one may sometimes remain confident, especially when one has in fact reasoned correctly; the first-order evidence still counts.\n\n*Implication.* Deliberus's design is a structural answer to the disagreement literature: instead of asking \"how much should I defer to someone who disagrees?\", it asks \"*where* do our reasons diverge?\" — converting peer disagreement (higher-order) back into first-order disagreement about identifiable premises. That is the platform's clearest philosophical selling point, and it can be stated: **the conciliation-vs-steadfast dilemma arises because the peer's reasons are opaque; making them addressable dissolves much of it** (the founding \"opacity is a cost\" conviction in epistemological terms). Residual peer disagreement *after* full decomposition is exactly the permissivism question.\n\n### 4.3 Uniqueness vs permissivism\n\n**Uniqueness Thesis**: for any body of evidence E and proposition p, there is exactly one doxastic attitude toward p that is rational given E. Defenders: Roger White, \"Epistemic Permissiveness\" (*Philosophical Perspectives* 19, 2005, 445–459) **[verified]**, argues permissive cases license arbitrary switching of belief and make rational belief insensitive to truth; Feldman (2007); Christensen; Matheson.\n\n**Permissivism**: sometimes more than one attitude is rational given the same evidence. Thomas Kelly, \"Evidence Can Be Permissive\" (in Steup, Turri & Sosa eds., *Contemporary Debates in Epistemology*, 2nd ed., Wiley-Blackwell 2013/2014, 298–313) **[verified]**: rational agents can differ in how they weigh competing epistemic values (avoid error vs believe truth — James's two commandments), yielding different thresholds. Miriam Schoenfield, \"Permission to Believe: Why Permissivism Is True and What It Tells Us About Irrelevant Influences on Belief\" (*Noûs* 48, 2014, 193–218) **[verified]**: rationality is relative to one's *epistemic standards*; different reasonable standards license different credences; learning that one's standards were shaped by an irrelevant influence (upbringing) need not require abandoning them. A standard technical distinction: *intrapersonal* uniqueness (one agent, one standard) is more plausible than *interpersonal* uniqueness (all agents).\n\nFormally, subjective Bayesianism is permissive by construction (any coherent prior is allowed), objective Bayesianism (Jaynes; Williamson, *In Defence of Objective Bayesianism*, 2010 **[from memory]**) aims at uniqueness via maximum entropy — but faces the problem that the partition to which indifference is applied is itself a choice (Bertrand's paradox). Merging-of-opinions theorems (Blackwell & Dubins 1962; Gaifman & Snir 1982 **[from memory]**): agents with mutually absolutely continuous priors who condition on the same growing stream of evidence converge almost surely — convergence *in the limit*, not agreement at finite evidence.\n\n### 4.4 Assessment (b): the founder's same-evidence-same-probability hope\n\nThe hope (\"given the same evidence, probability should ideally not differ between people\"; fear of \"fuzzy gooey subjectivity\") **is the interpersonal Uniqueness Thesis**. The literature gives it a mixed verdict that can be stated precisely:\n\n1. **Where it is well supported.** The *likelihood ratio* — how much the evidence favours H over ¬H — is, given an agreed model, prior-independent. Two people with different priors *should* agree on the LR. So \"the same evidence says the same thing to everyone\" is defensible **for Royall's Q3** (what the evidence says). This is also where decomposition does its work: disagreements about likelihoods can often be decomposed into disagreements about identifiable empirical premises.\n2. **Where it is contested but defensible as an ideal.** Convergence theorems show agents with open-minded priors converge as evidence accumulates. So the hope is well-grounded as a **regulative ideal over growing evidence**, the \"wager\" in the corpus's terms. It is not grounded as a claim about finite evidence.\n3. **Where it fails, on the best current arguments.** (i) Priors: no argument commands universal assent that fixes a unique prior; objective Bayesian principles depend on a choice of partition. (ii) Epistemic standards: Kelly's and Schoenfield's point that trading off error-avoidance against truth-seeking is a value choice. (iii) Standards of proof and what counts as evidence (the FRI finding) are Q2-type, practical and value-laden. (iv) Higher-order evidence makes the *rational* response agent-relative (Christensen) — two agents with the same first-order evidence but different track records of their own reliability can rationally differ.\n4. **What survives as a platform principle.** Not \"everyone ends at the same number\", but: **every difference in conclusion from shared evidence must be attributable to a stated difference in (a) prior, (b) epistemic standard/threshold, (c) model/likelihood, or (d) evidence** — each of which becomes a claim. This is *compatible with permissivism* (permissivists say the differences are licit, not that they are unstateable) and it is precisely the \"fuzzy gooey subjectivity\" defence the founder wants: subjectivity is allowed only in *named, bounded slots*. Schoenfield's standards and Kelly's thresholds become explicit horizon kinds (they are, in effect, the \"permissive zone\" and \"axiom choice\" kinds already in the enum). The honest wager-statement: *decomposition should shrink the set of rational credences toward a point on the empirical leg, and should localise the residual spread to named standards on the evaluative legs.*\n\nThis also sharpens the badge: the platform should display the **evidence term** (direction and weight, prior-free) as its own number, and never a posterior, which would import a prior the platform has no right to choose. *(Annotation 2026-09-30, proposed and unruled: a best guess may be shown beside the evidence term when its prior is itself a named, attributed claim, labelled as such; [../starting-values-and-the-problem-of-the-prior.md](../starting-values-and-the-problem-of-the-prior.md) § 6.7.)*\n\n---\n\n## 5. Evidence of absence\n\n**Bayesian result.** By the complementarity principle (§2.3), if finding E would confirm H (P(E|H) > P(E|¬H)), then *not* finding E disconfirms H. So \"absence of evidence is evidence of absence\" is a theorem — *but* the strength is the likelihood ratio P(¬E|¬H)/P(¬E|H), which depends on the **search's sensitivity** P(E found | H true). If the search could hardly have found E even were H true, the ratio is ≈ 1 and the absence is nearly worthless. The slogan \"absence of evidence is not evidence of absence\" is therefore false as a universal claim but correct as a warning about insensitive searches.\n\n**Sober**, \"Absence of evidence and evidence of absence: evidential transitivity in connection with fossils, fishing, fine-tuning, and firing squads\" (*Philosophical Studies* 143, 2009, 63–90) **[verified]**: uses the Law of Likelihood to assess the slogan; distinguishes the qualitative question (is non-observation evidence?) from the quantitative (how much?); the fossil-intermediate case shows that failure to find an intermediate fossil is evidence against common ancestry but typically *very weak* because of the imperfection of the fossil record; also treats observation selection effects (fine-tuning, firing squad). Strevens's comment (\"Objective Evidence and Absence\", 2008/9) **[existence verified]**.\n\n**Walton**, *Arguments from Ignorance* (Penn State UP, 1996) **[from memory — verify]**: the argument from ignorance (\"not known to be true, therefore false\") is a *legitimate defeasible scheme* when the knowledge base is reasonably complete — the **epistemic closure** or \"negative evidence\" condition: *if A were true, A would be known (would have been found)*. Its critical questions ask how far the search has progressed and whether it was thorough enough that A, if true, would have been found. Classic good instances: a train timetable that doesn't list a stop; a thorough background check that finds no record.\n\n*Implications for Deliberus.*\n- A listed gap — **\"no representation of 'we searched and found nothing' as data\"** — has a precise philosophical specification: a *negative search result* is a claim with three required parts: **what was searched for, where/how (the search's scope), and an estimate of its sensitivity**. Only then can it bear on anything. Without the sensitivity part it should have zero evidential weight (LR ≈ 1) and be displayed as such.\n- This also explains the brief's arithmetic complaint that **lack_of_evidence (one CQ) can never leave the neutral band**: Walton's scheme has its strength *entirely* in the closure condition. The fix is not more CQs but letting the answer to the closure CQ *carry the sensitivity estimate*, so that a well-documented exhaustive search can move far and a casual one cannot.\n- It is also the formal link between evidence-of-absence and the platform's *coherent absence* worry. The graph failing to contain an attack on a claim is evidence the claim is unattackable **only to the extent that the graph's search of that region was sensitive** — i.e. that people with reason to attack it looked. That is why \"an unattacked claim may be uncontroversial or merely unexamined\" (a caveat already in the corpus) is exactly Sober's point, and why exposure (who has looked) must be measured before absence of attacks is read as strength.\n\n---\n\n## 6. Testimony as evidence\n\n### 6.1 Reductionism vs anti-reductionism\n\n**Reductionism** (Hume, *Enquiry* §X, 1748): testimony's credibility derives from our experience of the conformity of testimony to fact; belief in testimony must be justified by non-testimonial (inductive) reasons. **Anti-reductionism** (Thomas Reid, *Inquiry into the Human Mind*, 1764 — the \"principle of credulity\" and \"principle of veracity\"; C. A. J. Coady, *Testimony: A Philosophical Study*, OUP 1992 **[from memory]**): testimony is a basic source of justification, like perception; we are default-entitled to accept it absent defeaters. Coady's argument: the Humean inductive base is itself largely testimonial, so global reduction is impossible. **Hybrid views**: Jennifer Lackey, *Learning from Words: Testimony as a Source of Knowledge* (OUP 2008) **[from memory — verify]**: testimonial justification requires *both* that the speaker's statement be reliable (speaker-side) *and* that the hearer have positive reasons (hearer-side, but not global reduction) — and her \"statements, not beliefs\" view: one can learn from a statement even if the speaker does not believe it (the creationist teacher case).\n\n### 6.2 Expert testimony and the novice-expert problem\n\nAlvin Goldman, \"Experts: Which Ones Should You Trust?\" (*Philosophy and Phenomenological Research* 63(1), 2001, 85–110) **[verified]**. A novice choosing between disagreeing putative experts has **five sources of evidence**: (1) the **arguments** presented by the contending experts (direct or, more realistically, *indirect* argumentative justification — dialectical superiority, e.g. quick, confident rebuttals); (2) **agreement from additional putative experts** — with the crucial caveat that numbers matter only if the agreeing experts are not mere followers of a common guru (**non-independence**); (3) appraisals by **meta-experts** (credentials, ratings); (4) evidence about the experts' **interests and biases**; (5) evidence of their **past track records**. Goldman flags (5) as potentially the most useful and (2) as the most treacherous because of dependence.\n\n**John Hardwig**, \"Epistemic Dependence\" (*Journal of Philosophy* 82, 1985) and \"The Role of Trust in Knowledge\" (*J Phil* 88, 1991) **[from memory]**: modern knowledge is irreducibly collective; the rational layperson (and scientist) *must* defer, and knowledge in science rests on trust in the moral and epistemic character of others. \"A person who is epistemically dependent should defer.\"\n\n*Implications for Deliberus.*\n- Goldman's five sources are a ready-made CQ set for the **expert-opinion** scheme that is more discriminating than the standard Walton list, and source (2)'s independence caveat is the same independence requirement as §3.5. It argues for separating **source reliability** (track record, interests, credentials — properties of a source node) from the **claim** — a listed gap.\n- Goldman's argument (1) is the platform's raison d'être: Deliberus is a machine for giving novices *direct* access to experts' arguments, reducing reliance on sources (2)–(5). That is a strong pitch point, grounded in the classic paper.\n- Lackey's statements-not-beliefs view supports the always-mint / REPORTS design: what enters the graph is the *statement*, and its evidential value for p is a separate question from the speaker's sincerity. Reid's credulity principle gives the default: testimony has some positive weight absent defeaters — i.e., the tiered positional-testimony design (\"weight rises with tier; nothing gated on tier\") is anti-reductionist with a reductionist ratchet, very close to Lackey's dualism.\n- Hardwig's point is the strongest argument *against* an overly demanding decomposition-for-everything ideal: some deference is not a failure of rationality but a condition of it. The platform should represent deference explicitly (a claim resting on trust in a named source) rather than treat undecomposed testimony as a defect.\n\n### 6.3 A note on credibility and chain of custody (legal evidence scholarship)\n\nThe philosophy literature is thin on chain of custody; the evidence-law and \"science of evidence\" literature is not. David Schum, *Evidential Foundations of Probabilistic Reasoning* (Wiley 1994; Northwestern UP 2001) and Anderson, Schum & Twining, *Analysis of Evidence* (2nd ed., CUP 2005) **[from memory — verify]** decompose a testimonial report into a **chain of inferences** — from the report to the witness's belief (**veracity**), from belief to what the senses registered (**objectivity**), from sensory evidence to the event (**observational sensitivity**) — each a separate attackable link; they extend Wigmore's charts (1913/1937). Schum also formalises how redundant and corroborative evidence combine, and shows that the force of evidence depends on these credibility links.\n\n*Implication.* This is the single most directly buildable framework for the listed gaps \"no chain of custody / quote verification / page or timestamp locator\" and \"no source reliability separate from claim\". A testimony or citation edge should decompose into the Schum credibility chain on demand: *did the source say it* (locator, quote verification) → *did the source believe it* → *was the source positioned to know* → *did the event occur*. Each link is a claim; the corpus's \"position to know\" scheme covers one link, and REPORTS covers the first.\n\n---\n\n## 7. Theory-ladenness, the Given, and whether any evidence is \"atomic\"\n\n### 7.1 Theory-ladenness\n\nN. R. Hanson, *Patterns of Discovery* (CUP 1958): \"there is more to seeing than meets the eyeball\" — Tycho and Kepler at dawn; observation is shaped by background theory. Thomas Kuhn, *The Structure of Scientific Revolutions* (1962): paradigm-relative observation, incommensurability. Later moderation: Fodor (1984) on encapsulated perception; the distinction between *perceptual* and *semantic* theory-ladenness; the practical view (Hacking 1983, *Representing and Intervening*; Franklin) that experimental practice gives observations a degree of theory-independence through instrument calibration and intervention.\n\n*Consensus.* That observation *reports* are theory-laden in vocabulary and relevance: broad consensus. That this makes evidence incapable of adjudicating between theories: minority view. The usual resolution: evidence can be *neutral enough* — laden with theories not in dispute.\n\n*Implication.* The practical criterion for the platform: **a piece of evidence is usable as neutral arbiter between two positions to the extent that the theory it is laden with is not itself in dispute between them.** That is checkable in a graph: if the claim at the tail of an evidence edge has a decomposition that runs through a premise one party attacks, it is not neutral between those parties. The platform can *compute* when an alleged \"fact\" is party-laden. (It is also the right reading of the FRI finding.)\n\n### 7.2 Sellars and the Myth of the Given\n\nWilfrid Sellars, \"Empiricism and the Philosophy of Mind\" (1956) **[from memory]**: the \"Given\" — something that is both (a) non-conceptual, acquired without any learning, and (b) capable of *justifying* belief — is a myth. Anything that justifies must stand in the \"logical space of reasons\", i.e., be conceptually articulated and itself open to being asked for reasons; but anything so articulated presupposes a web of concepts and knowledge. Hence \"empirical knowledge, like its sophisticated extension, science, is rational, not because it has a *foundation* but because it is a self-correcting enterprise which can put *any* claim in jeopardy, though not *all* at once.\" **[famous passage; quote from memory — verify wording, EPM §38]**\n\n### 7.3 Foundationalism vs coherentism vs foundherentism\n\n**Classical foundationalism**: justified beliefs rest on basic beliefs that are self-justifying (infallible, indubitable, or self-evident). **Modest foundationalism** (e.g., Audi, Alston): basic beliefs are justified non-inferentially but *defeasibly*. **Coherentism** (BonJour 1985; Lehrer): justification is a matter of coherence within a system; no basic beliefs. Problems: foundationalism's regress stopper seems either unjustified or not basic; coherentism's isolation objection (a coherent fairy tale) and the Bovens–Hartmann/Olsson impossibility results (§3.5).\n\n**Haack's foundherentism**, *Evidence and Inquiry: Towards Reconstruction in Epistemology* (Blackwell 1993; expanded 2nd ed. Prometheus 2009) **[verified]**: warrant depends on (i) **supportiveness** — how well experiential evidence and reasons support the belief; (ii) **independent security** — how well-warranted the supporting reasons are *independently of the belief in question*; (iii) **comprehensiveness** — how much of the relevant evidence is included. The **crossword analogy**: clues ≈ experiential evidence; intersecting completed entries ≈ reasons. An entry is well-warranted to the extent it fits its clue and the intersecting entries, *and* those entries are themselves reasonable independently of it, *and* enough of the puzzle has been filled. Mutual support without vicious circularity. Crucially, experience (the clue) contributes causally and evidentially without being a *belief* and without being infallible.\n\n### 7.4 Assessment (a): \"There is no atomic evidence, only claims whose decomposition nobody has demanded yet\"\n\n**What the slogan gets right, with strong philosophical backing:**\n1. **No evidence is immune to challenge.** This is Sellars's conclusion (any claim can be put in jeopardy), Quine's (no statement immune to revision), and modest foundationalism's own concession (basic beliefs are *defeasible*). Consensus in contemporary epistemology is against infallible foundations.\n2. **Whatever functions as evidence in a public graph must be propositional**, and every proposition can be asked for reasons — Sellars's \"space of reasons\". So in the graph, *every* evidence node is a claim, and every claim is decomposable in principle.\n3. **The \"nobody has demanded yet\" clause is Sellars's \"not all at once\"** — and Peirce's and Wittgenstein's point that doubt needs a reason (see 7.5). It correctly makes atomicity *relative to the current dialectical state*, which is also how dialectical/dialogical theories of justification (Rescorla, \"Epistemic and Dialectical Regress\", 2009 **[from memory]**; Brandom's default-and-challenge structure, *Making It Explicit*, 1994) treat it: a claim is entitled by default until a *legitimate* challenge is issued.\n\n**Where it overreaches or needs a qualification:**\n1. **It erases a real difference in kind.** Haack's crossword shows the problem: if clues were merely more entries, the puzzle could be completed coherently while ignoring the clues — the coherentist fairy tale. Evidence *as evidence* has a causal/experiential or measurement anchor that other claims lack. In the graph, *observation reports, measurements and records* are still claims, but they carry a different kind of support — **provenance to an observation event** — that makes their warrant partly *independent* of the rest of the graph (Haack's \"independent security\"). A slogan that makes all claims alike invites the coherence-without-anchor failure the Bovens–Hartmann/Olsson results warn about.\n2. **\"Nobody has demanded\" conflates *demanded* with *legitimately demandable*.** Brandom and Rescorla make the challenge a normative notion: a bare \"why?\" does not automatically shift the burden. A platform where any demand triggers decomposition invites **regress-by-demand** (a strategy-class adversary can demand decomposition indefinitely to stall). The better slogan: *only claims whose decomposition no one has **reasonably** challenged yet* — with what counts as a reasonable challenge itself a matter of stated grounds (which is how the platform's \"contributions argue for themselves\" rule already works).\n3. **Hinge commitments (§7.5) are a principled counterexample** to \"everything decomposes productively\": some propositions are not improved by further evidential support because they are presupposed by the practice of giving evidence at all.\n4. **Decomposition is not the only thing that changes an evidence claim's standing.** Independent replication, calibration of an instrument, and a track record are *lateral* evidence (they support the source, not the claim's parts). The slogan's vertical picture needs the lateral dimension.\n\n**Proposed restatement, consistent with the literature:** *\"No evidence is foundational in the sense of immune to challenge; every piece of evidence is a claim, decomposable on reasonable challenge. But claims differ in the kind of support they rest on, and a claim anchored to an observation event carries warrant independent of the claims it supports — which is what makes it evidence rather than one more entry in the crossword.\"* In graph terms: keep \"no atomic evidence\", but add a **provenance kind** for claims grounded in observation/measurement/record (distinct from inferential support), so that the graph can compute Haack's independent security and detect coherent-but-unanchored clusters.\n\n**On the founder's question \"what constitutes un-decomposable evidence?\"** The literature's best answer is: *nothing is undecomposable in principle; what stops decomposition in practice is (i) a hinge (not doubtable within the practice), (ii) a report of a seeming whose content is the seeming itself (the incorrigible first-person core — Huemer/Conee–Feldman), or (iii) a measurement whose further decomposition would only re-enter its own calibration record (instrument reliability, which is lateral).* These are three different horizon kinds for the empirical leg, and naming them would sharpen \"empirical = hands over to evidence\".\n\n### 7.5 Hinge epistemology\n\nLudwig Wittgenstein, *On Certainty* (written 1949–51; pub. 1969): \"the questions that we raise and our doubts depend on the fact that some propositions are exempt from doubt, are as it were like hinges on which those turn\" (§341) **[from memory — standard quote]**. Duncan Pritchard, *Epistemic Angst: Radical Skepticism and the Groundlessness of Our Believing* (Princeton 2016) **[from memory — verify]**: hinge commitments (\"I have hands\", \"the earth has existed for a long time\") are not *beliefs* in the knowledge-apt sense and are not rationally grounded; they are the background against which rational evaluation takes place. Rational evaluation is essentially local. Rivals: Coliva's framework readings; Wright's \"entitlement\" (non-evidential warrant to trust hinges).\n\n*Implications.*\n- Hinge epistemology directly supports the corpus's **structural / axiom-choice horizon kinds**: some descents end not in missing evidence but in commitments whose status is framework, not conclusion.\n- It also warns: hinges should not be represented as unsupported value premises (which the completeness oracle would flag as \"copout axioms\"). A hinge *reached* is not a copout; it is a place where \"the spade is turned\" (*Philosophical Investigations* §217). The completeness instrument needs a hinge kind that is not counted as a defect.\n- Caution: **hinges differ between people and communities only in some cases** (deep disagreement, Fogelin 1985). Where they differ, evidence cannot settle; the platform can only make the hinge difference legible — which is the \"ceiling counts against the wager\" rule the project already holds.\n\n---\n\n## 8. Evidence in normative domains\n\n### 8.1 Reflective equilibrium\n\nJohn Rawls, \"Outline of a Decision Procedure for Ethics\" (*Phil Review* 60, 1951) and *A Theory of Justice* (1971, §4, §9) **[from memory]**: moral justification is a matter of mutual adjustment between principles and **considered judgments** about cases, until they cohere. Norman Daniels, \"Wide Reflective Equilibrium and Theory Acceptance in Ethics\" (*Journal of Philosophy* 76, 1979) **[from memory]**: *wide* RE includes background theories (of the person, of society, empirical facts), not just principles and judgments. Critics: considered judgments may be bias (Singer 1974; Hare); RE is coherentism with the coherence-truth gap (§3.5).\n\n### 8.2 Are intuitions evidence?\n\nMainstream methodology treats intuitions as defeasible evidence (Bealer 1998; Sosa; Huemer's phenomenal conservatism extends to intellectual seemings). **Herman Cappelen**, *Philosophy Without Intuitions* (OUP 2012) **[from memory]**: philosophers do not in fact rely on intuitions as evidence; \"intuitive\" is a hedge; arguments do the work. Max Deutsch, *The Myth of the Intuitive* (MIT 2015) makes a similar case. **Experimental philosophy** (Weinberg, Nichols & Stich 2001; Machery et al. 2004 on reference; later replication work — Cova et al. 2021 found a high replication rate for x-phi studies **[from memory — verify]**) challenges the *reliability* of intuitions by showing variation with culture, order and framing — though several headline cultural-variation results failed to replicate.\n\n### 8.3 The fact/value line for evidence\n\nPositions range from sharp separation (Humean: no ought from is; moral claims have no evidence in the empirical sense) through moral realism with intuitions as evidence (Huemer, *Ethical Intuitionism*, 2005), to pragmatists and \"thick concept\" theorists (Putnam, *The Collapse of the Fact/Value Dichotomy*, 2002) who deny a sharp line. Inductive-risk literature (Rudner 1953; Douglas, *Science, Policy, and the Value-Free Ideal*, 2009) **[from memory]**: even empirical evidence thresholds (how much evidence is enough to accept) involve value judgments about the costs of error — so **standards of proof are value-laden even on the empirical leg**.\n\n*Implications for Deliberus.*\n- Reflective equilibrium is **what decomposition of a value claim does**: principles (mother claims) and case judgments (children) adjusted until they cohere. The graph can represent both, and the \"attributed interior library\" of the permissive zone is a store of individuals' reflective equilibria.\n- Intuitions **can** enter the graph as evidence under a phenomenal-conservatism reading — as *reports of seemings* with defeasible default weight — but the x-phi literature is a reason to make **framing and order sensitivity** a CQ on any claim supported chiefly by intuition. Cappelen's challenge is a reason not to over-label: a claim tagged \"intuition\" is often really an argument-in-waiting, i.e., an undecomposed claim.\n- **Impact testimony**: under Huemer/Conee–Feldman it is evidence (of what the position costs), with incorrigibility limited to the experiential core (as the brief already says). Under inductive-risk reasoning it bears directly on *standards of proof* (how bad an error would be for whom). That resolves the brief's contradiction \"whether impact testimony is weighed as evidence\": it is **evidence for premises about costs and stakes**, which then enter the *weighing* and the choice of standard — not evidence for empirical conclusions. That is consistent with \"testimony is evidence for a premise, never a verdict\".\n- **Standard of proof** (a listed gap) is best modelled, per Douglas and legal epistemology (Laudan, *Truth, Error, and Criminal Law*, 2006 **[from memory]**), as an explicit, attributable claim about the acceptable error trade-off for a *given use* — separate from the evidence. Two parties may agree on what the evidence says (Q3) and disagree on whether it is enough (Q2), and the platform should be able to show that.\n\n---\n\n## 9. Explicit assessments\n\n### (a) \"There is no atomic evidence, only claims whose decomposition nobody has demanded yet\"\nSee §7.4 for the full argument. **Verdict: sound as anti-foundationalism, strongly backed (Sellars, Quine, modest foundationalism's own concessions, dialectical default-and-challenge theories); incomplete in two ways.** (1) It erases the difference in *kind* of support that makes evidence evidence (Haack's clues vs entries; \"independent security\"); the graph needs a provenance kind for observation/measurement/record-grounded claims so it can detect coherence without anchor. (2) \"Demanded\" should be \"reasonably challenged\", or the principle becomes a stalling vector. Plus: hinges are a principled place where decomposition stops being productive, and should be a non-defect horizon kind.\n\n### (b) Same evidence → same probability\nSee §4.4. **Verdict: correct for what evidence says (the likelihood ratio, prior-free given a model) and as a limit ideal over growing evidence (merging theorems); not established, and on the best arguments false, as a claim about posteriors at finite evidence** (priors, epistemic standards, value-laden thresholds, agent-relative higher-order evidence). **Platform-level restatement that keeps the founder's anti-gooey intent**: *every divergence from shared evidence must be attributable to a named difference in prior, standard, model or evidence* — permissivism is allowed only in named slots. Display the evidence term, never a posterior.\n\n### (c) Counting CQ answers vs weighing them\n**Verdict: counting is not defensible as a measure of evidential strength**, on every framework surveyed:\n- **Bayesian/likelihoodist**: strength is diagnosticity (LR), which varies by orders of magnitude between answers; the ravens paradox is the classic demonstration that counting instances misleads.\n- **Error-statistical**: an answer's value depends on the severity of the check behind it; an unfalsifiable \"yes\" is worth nothing.\n- **Keynesian**: counting inputs conflates weight with balance; and without independence it does not even measure weight.\n- **Testimony / independence**: three answers from one account are one witness (Goldman's non-independence; Condorcet correlation; Bovens–Hartmann). The brief's \"three one-word answers = three RCTs\" is the paradigm failure.\n- **Walton's own framework**: CQs are not votes; unanswered CQs shift the *burden* (defeasibility), and one strong undercutting answer can defeat the whole inference (Pollock's undercutting defeaters, \"Defeasible Reasoning\", *Cognitive Science* 11, 1987 **[from memory]**). Additive counting cannot represent a single decisive defeater.\n\n**What *is* defensible**: counting as a labelled measure of **scrutiny coverage** — \"k of n critical questions have been addressed\" — which is a weight-like, not balance-like, quantity, and should be displayed as such. And a **minimal weighing** that stays within \"propose, never assert\": each answer carries (as attackable claims) (i) its *polarity* (rebutting vs undercutting vs confirming), (ii) an *ordinal diagnosticity band* (would this answer be surprising if the inference were bad?), (iii) its *source* for independence grouping. Arithmetic then: sum log-odds bands over *independent* answers; take the max/one for dependent groups; let a strong undercutter cap the edge (a min, as for necessary parts). The current +0.15/answer rule can survive as the fallback when none of these are supplied — explicitly labelled as \"counted, not weighed\".\n\n---\n\n## 10. Consensus vs contested — a summary table\n\n| Topic | Broad consensus | Live controversy |\n|---|---|---|\n| Nature of evidence | Multiple roles that conflict (Kelly) | Propositional (E=K) vs mentalist/experiential |\n| Confirmation | Degrees matter; incremental ≠ absolute | Which measure; LR favoured for accumulation |\n| Likelihood principle | LR captures \"what data say\" given a model | Whether sampling/stopping matter (Mayo: yes) |\n| Weight vs balance | They differ (Keynes, Joyce) | How to formalise weight |\n| Holism | Tests are of conjunctions | How much blame is localisable (Bayes: a lot, given priors) |\n| Variety of evidence | Helps *conditionally* | Exact conditions (B&H, Claveau, Landes/Osimani) |\n| Coherence | Not truth-conducive in general | Special conditions under which it is |\n| Independence | Essential for aggregation | Right notion (conditional on shared evidence) |\n| IBE | Explanatory reasoning is ubiquitous | Rival rule to Bayes vs heuristic within it |\n| Causal evidence | Both kinds useful | Whether mechanistic evidence is *required* (R–W) |\n| External validity | \"Worked there\" ≠ \"will work here\" | How to evidence transfer |\n| Higher-order evidence | It is real and agent-relative | Whether it always defeats |\n| Disagreement | — | Conciliationism vs steadfastness |\n| Uniqueness | Intrapersonal more plausible than interpersonal | Interpersonal uniqueness |\n| Absence of evidence | Evidential iff search sensitive | — (settled in Bayesian terms) |\n| Testimony | Some default entitlement | Reductionism vs anti-reductionism vs hybrid |\n| Given | No infallible Given | Whether anything plays a non-inferential role |\n| Moral intuitions | Used in practice via RE | Whether they are evidence (Cappelen) and reliable (x-phi) |\n\n---\n\n## 11. Most transferable to a claim-graph argumentation platform like the one above\n\nConcrete, ordered roughly by leverage and by how little new machinery each needs.\n\n1. **Say what the number is, in Royall's terms.** The badge answers \"what does the evidence say\" (Q3), never \"what should you believe\" (Q1) or \"what should you do\" (Q2). Adopt this trichotomy in UI copy, docs, and the agent API; drop any residual \"Bayesian\"/\"probability\" wording for σ. It is the most economical philosophical defence of the no-verdict design.\n\n2. **Replace counted CQ answers with minimally weighed ones, keeping counting as coverage.** Per answer: polarity (confirm / rebut / undercut), an ordinal diagnosticity band (e.g. *weak ≈ LR 2, moderate ≈ 8, strong ≈ 32*, Royall's benchmarks — verify wording), and a source for independence grouping — each an attackable claim. Aggregate: add bands in log-odds across independent groups; within a dependent group count once; a strong undercutter caps the edge by min. Display \"k of n CQs addressed\" as **coverage (weight)**, separately.\n\n3. **Represent sources and independence as nodes, never as discounts.** A source node (person, lab, instrument, dataset, sponsor) with reliability as an attackable claim; every report/answer/citation linked to it. Aggregation rules then follow Bovens–Hartmann/Easwaran et al.: multiplicative (log-odds additive) only across independent sources. Concordant reports from one source update the *source's* reliability rather than multiplying the claim's strength. This retires \"three one-word answers = three RCTs\".\n\n4. **Make weight Keynesian, not a tally.** Weight = coverage of the scheme's question space by *independent* inputs, saturating; plus a **resilience** readout (how far σ moves if the single most influential input is removed or reversed) — computable with the existing hinge machinery. Keep \"unexamined → no number\".\n\n5. **A negative-search claim type.** \"We searched for E and found nothing\" = {target, scope/method, estimated sensitivity}. Without sensitivity it carries LR ≈ 1 and is shown as such. This fixes the stuck lack_of_evidence scheme (let the closure CQ answer carry the sensitivity) and gives coherent absence a measurable form: \"no attacks found\" counts only in proportion to how exposed the claim was.\n\n6. **Schum's credibility chain for testimony and citations.** Decompose any testimony/citation edge on demand into: said it (locator: page/timestamp, quote match) → believed it (veracity) → positioned to know (objectivity/position to know) → event occurred (sensitivity). Covers chain of custody, quote verification and locators with one pattern, and uses existing REPORTS and position-to-know machinery.\n\n7. **Two causal-evidence slots (Russo–Williamson) and an external-validity scheme (Cartwright & Hardie).** Cause-to-effect edges carry separate CQs for difference-making and mechanistic evidence; a missing mechanism is shown as its own gap, not as a lower rung. Every \"study → policy/prediction\" edge gets the \"it worked there → it will work here\" CQs: same causal structure, support factors present, relevant population similarity. Methodology stays per claim and per edge (relevance is an edge property); keep the ladder unratified.\n\n8. **Severity-lowering properties attach to the inference, not the conclusion.** Publication bias, replication status, preregistration, funding conflict, optional stopping → claims attached to the evidence edge (Mayo). This gives several listed gaps one principled home and keeps bias from becoming ad hominem against the conclusion.\n\n9. **Rebutting vs undercutting typed** (Pollock). An ATTACK on a claim's truth (rebut) and an attack on the inferential link (undercut) behave differently: an undercutter caps the edge; a rebutter competes with the claim. The CQ answers are mostly undercutters; typing them makes item 2's min-cap rule available.\n\n10. **A provenance kind for observation-anchored claims (Haack's independent security).** Keep \"no atomic evidence\", but distinguish claims whose support includes an observation/measurement/record event from purely inferential ones. Then compute a diagnostic: clusters that are highly mutually supporting but contain no anchored claim are *coherent-but-unanchored* (the Bovens–Hartmann/Olsson warning) and should be flagged, not rewarded.\n\n11. **\"Reasonably challenged\", not merely \"demanded\".** Decomposition requests should carry a stated ground (default-and-challenge, Brandom/Rescorla) — the platform's \"contributions argue for themselves\" rule applied to challenges — to prevent regress-by-demand as a strategy-class stall.\n\n12. **Hinge and seeming as non-defect horizon kinds on the empirical leg.** Add (or name within the enum) hinge (framework, not doubtable within the practice), first-person seeming (incorrigible core; causes and satisfiers still decomposable), and instrument calibration (lateral, not vertical). Stop the completeness oracle from counting any of them as copout axioms.\n\n13. **An \"open\" epistemic status.** The ladder of epistemic statuses has no slot for \"this is an open question\"; confirmation theory's distinction between zero weight and probability ½ is the theoretical case for it (Keynes), and the Lindqvist mapping is the practical case. Add it.\n\n14. **Attributable divergence as the operational form of the uniqueness hope.** When two contributors reach different conclusions from shared evidence, prompt for the locus: prior, standard of proof, model/likelihood, or evidence — each minted as an attributed claim. Standards of proof become explicit, attributable, attackable claims tied to a use (inductive risk; Douglas). This turns \"fuzzy gooey subjectivity\" into named, bounded slots and is compatible with both uniqueness and permissivism.\n\n15. **Pass-level calibration feeds edge-level credibility (Christensen's higher-order evidence, applied to the machine).** A measured misfiling rate for a pass is evidence about *every* output of that pass. Watermark every machine judgment with pass and version, and let pass calibration discount the confidence of all its unexamined outputs — not only those shown wrong.\n\n16. **Consilience as a displayed structural fact.** Compute and show whether a claim's support arrives from disjoint regions (source clusters, concept clusters, methods with distinct failure modes). Do not add it to strength; use it as a CQ on robustness claims (\"do these methods fail independently?\").\n\n17. **Frame the platform's answer to peer disagreement.** Public copy and funder framing: the conciliation-vs-steadfast dilemma exists because a peer's reasons are opaque; Deliberus converts higher-order disagreement into first-order disagreement at identified premises, and what remains after full decomposition is the permissive residue, which it records rather than resolves. Goldman (2001) supplies the novice's-access argument: direct access to experts' arguments reduces dependence on credentials and head-counts.\n\n---\n\n## References\n\n(✔ = verified this session; ◦ = from memory, verify before external use.)\n\n- Achinstein, P. (2001). *The Book of Evidence*. Oxford University Press. https://global.oup.com/academic/product/the-book-of-evidence-9780195171716 ✔\n- Anderson, T., Schum, D., & Twining, W. (2005). *Analysis of Evidence* (2nd ed.). Cambridge University Press. ◦\n- Birnbaum, A. (1962). On the foundations of statistical inference. *JASA* 57, 269–306. ◦\n- Blackwell, D., & Dubins, L. (1962). Merging of opinions with increasing information. *Annals of Mathematical Statistics* 33, 882–886. ◦\n- Bovens, L., & Hartmann, S. (2003). *Bayesian Epistemology*. Oxford University Press. ✔ (VET failure verified via Osimani & Landes summary)\n- Bovens, L., & Hartmann, S. The Variety-of-Evidence Thesis and the Reliability of Instruments: A Bayesian-Network Approach. PhilPapers: https://philpapers.org/rec/HARTVT ✔\n- Brandom, R. (1994). *Making It Explicit*. Harvard University Press. ◦\n- Cappelen, H. (2012). *Philosophy Without Intuitions*. Oxford University Press. ◦\n- Carnap, R. (1947). On the application of inductive logic. *PPR* 8, 133–148. ◦ — (1950/1962). *Logical Foundations of Probability*. Chicago. ◦\n- Cartwright, N. (2011). A philosopher's view of the long road from RCTs to effectiveness. *The Lancet* 377, 1400–1401. ◦\n- Cartwright, N., & Hardie, J. (2012). *Evidence-Based Policy: A Practical Guide to Doing It Better*. Oxford University Press. ◦\n- Christensen, D. (2007). Epistemology of disagreement: the good news. *Philosophical Review* 116, 187–217. ◦\n- Christensen, D. (2010). Higher-order evidence. *Philosophy and Phenomenological Research* 81(1), 185–215. https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1933-1592.2010.00366.x ✔\n- Claveau, F. (2013). The independence condition in the variety-of-evidence thesis. *Philosophy of Science* 80, 94–118. ✔\n- Claveau, F., & Grenier, O. (2019). The variety-of-evidence thesis: a Bayesian exploration of its surprising failures. *Synthese*. https://link.springer.com/article/10.1007/s11229-017-1607-5 ✔\n- Coady, C. A. J. (1992). *Testimony: A Philosophical Study*. Oxford University Press. ◦\n- Conee, E., & Feldman, R. (1985). Evidentialism. *Philosophical Studies* 48, 15–34; (2004) *Evidentialism: Essays in Epistemology*. OUP. ◦\n- Crupi, V. (2013; rev. 2025). Confirmation. *Stanford Encyclopedia of Philosophy*. https://plato.stanford.edu/entries/confirmation/ ✔\n- Daniels, N. (1979). Wide reflective equilibrium and theory acceptance in ethics. *Journal of Philosophy* 76, 256–282. ◦\n- Dietrich, F., & Spiekermann, K. (2013). Independent opinions? On the causal foundations of belief formation and jury theorems. *Mind* 122, 655–685. ◦\n- Douglas, H. (2009). *Science, Policy, and the Value-Free Ideal*. University of Pittsburgh Press. ◦\n- Douven, I. Abduction. *Stanford Encyclopedia of Philosophy*; (2022) *The Art of Abduction*. MIT Press. ◦\n- Duhem, P. (1906). *La théorie physique: son objet et sa structure*. ◦\n- Easwaran, K., Fenton-Glynn, L., Hitchcock, C., & Velasco, J. D. (2016). Updating on the credences of others: disagreement, agreement, and synergy. *Philosophers' Imprint* 16(11), 1–39. https://philarchive.org/rec/FENUOT-3 ✔\n- Elga, A. (2007). Reflection and disagreement. *Noûs* 41, 478–502. ◦\n- Fitelson, B. (1999). The plurality of Bayesian measures of confirmation and the problem of measure sensitivity. *Philosophy of Science* 66, S362–S378. ◦\n- Fitelson, B. (2007). Likelihoodism, Bayesianism, and relational confirmation. *Synthese* 156, 473–489. ◦\n- Goldman, A. I. (2001). Experts: which ones should you trust? *Philosophy and Phenomenological Research* 63(1), 85–110. https://onlinelibrary.wiley.com/doi/10.1111/j.1933-1592.2001.tb00093.x ✔\n- Good, I. J. (1950). *Probability and the Weighing of Evidence*. Griffin. ◦ — (1967). On the principle of total evidence. *BJPS* 17, 319–321. ◦ — (1968) cited in SEP \"Confirmation\" ✔\n- Haack, S. (1993/2009). *Evidence and Inquiry: Towards Reconstruction in Epistemology*. Blackwell / Prometheus. https://friesian.com/haack.htm (review); Wikipedia \"Foundherentism\" ✔\n- Hacking, I. (1965). *Logic of Statistical Inference*. CUP. ◦ — (1983). *Representing and Intervening*. CUP. ◦\n- Hájek, A., & Joyce, J. (2008). Confirmation. In Psillos & Curd (eds.), *Routledge Companion to Philosophy of Science*. (cited in SEP) ✔\n- Hanson, N. R. (1958). *Patterns of Discovery*. Cambridge University Press. ◦\n- Hardwig, J. (1985). Epistemic dependence. *Journal of Philosophy* 82, 335–349; (1991) The role of trust in knowledge. *J Phil* 88, 693–708. ◦\n- Hempel, C. G. (1945). Studies in the logic of confirmation. *Mind* 54, 1–26, 97–121. ✔ (via SEP)\n- Howick, J. (2011). *The Philosophy of Evidence-Based Medicine*. Wiley-Blackwell. ◦\n- Huemer, M. (2001). *Skepticism and the Veil of Perception*; (2005) *Ethical Intuitionism*; (2007) Compassionate phenomenal conservatism. *PPR* 74, 30–55. ◦\n- Joyce, J. M. (2005). How probabilities reflect evidence. *Philosophical Perspectives* 19, 153–178. https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1520-8583.2005.00058.x ✔\n- Kelly, T. (2006; rev. 2014). Evidence. *Stanford Encyclopedia of Philosophy*. https://plato.stanford.edu/entries/evidence/ ✔\n- Kelly, T. (2005). The epistemic significance of disagreement. *Oxford Studies in Epistemology* 1. ◦\n- Kelly, T. (2013/2014). Evidence can be permissive. In Steup, Turri & Sosa (eds.), *Contemporary Debates in Epistemology* (2nd ed.), 298–313. https://philpapers.org/rec/KELECB ✔\n- Keynes, J. M. (1921). *A Treatise on Probability*. Macmillan (ch. VI, \"The Weight of Arguments\"). ◦ (chapter number)\n- Kuhn, T. S. (1962). *The Structure of Scientific Revolutions*. Chicago. ◦\n- Lackey, J. (2008). *Learning from Words: Testimony as a Source of Knowledge*. Oxford University Press. ◦\n- Landes, J. (2020). The variety of evidence thesis and its independence of degrees of independence. *Synthese*. https://link.springer.com/article/10.1007/s11229-020-02738-5 ✔\n- Lasonen-Aarnio, M. (2014). Higher-order evidence and the limits of defeat. *PPR* 88(2), 314–345. https://onlinelibrary.wiley.com/doi/10.1111/phpr.12090 ✔\n- Laudan, L. (2006). *Truth, Error, and Criminal Law*. CUP. ◦\n- Levins, R. (1966). The strategy of model building in population biology. *American Scientist* 54, 421–431. ◦\n- Lipton, P. (2004). *Inference to the Best Explanation* (2nd ed.). Routledge. ◦\n- Mayo, D. G. (1996). *Error and the Growth of Experimental Knowledge*. Chicago. ◦ — (2018). *Statistical Inference as Severe Testing*. Cambridge University Press. https://assets.cambridge.org/97811070/54134/frontmatter/9781107054134_frontmatter.pdf ✔\n- Myrvold, W. (2003). A Bayesian account of the virtue of unification. *Philosophy of Science* 70, 399–423. ◦\n- Olsson, E. J. (2005). *Against Coherence: Truth, Probability, and Justification*. OUP. ◦\n- Osimani, B., & Landes, J. (2023). Varieties of error and varieties of evidence in scientific inference. *BJPS* 74. https://www.thebsps.org/short-reads/evidence-osimani-landes/ ✔\n- Pollock, J. L. (1987). Defeasible reasoning. *Cognitive Science* 11, 481–518. ◦\n- Popper, K. R. (1934/1959). *The Logic of Scientific Discovery*. ◦\n- Pritchard, D. (2016). *Epistemic Angst*. Princeton University Press. ◦\n- Quine, W. V. O. (1951). Two dogmas of empiricism. *Philosophical Review* 60, 20–43. ◦\n- Rawls, J. (1951). Outline of a decision procedure for ethics. *Phil Review* 60; (1971) *A Theory of Justice*. ◦\n- Reid, T. (1764). *An Inquiry into the Human Mind on the Principles of Common Sense*. ◦\n- Royall, R. (1997). *Statistical Evidence: A Likelihood Paradigm*. Chapman & Hall. https://www.routledge.com/Statistical-Evidence-A-Likelihood-Paradigm/Royall/p/book/9780412044113 ✔ (existence, three questions); benchmarks 8/32 ◦\n- Russo, F., & Williamson, J. (2007). Interpreting causality in the health sciences. *International Studies in the Philosophy of Science* 21(2), 157–170. https://www.tandfonline.com/doi/abs/10.1080/02698590701498084 ✔\n- Schoenfield, M. (2014). Permission to believe. *Noûs* 48, 193–218. https://philpapers.org/rec/SCHPTB ✔\n- Schum, D. A. (1994). *Evidential Foundations of Probabilistic Reasoning*. Wiley. ◦\n- Sellars, W. (1956). Empiricism and the philosophy of mind. *Minnesota Studies in the Philosophy of Science* 1. ◦\n- Skyrms, B. (1980). *Causal Necessity*. Yale University Press (resilience). ◦\n- Sober, E. (2009). Absence of evidence and evidence of absence. *Philosophical Studies* 143, 63–90. https://link.springer.com/article/10.1007/s11098-008-9315-0 ✔\n- Strevens, M. (2001). The Bayesian treatment of auxiliary hypotheses. *BJPS* 52, 515–537. ◦\n- Walton, D. (1996). *Arguments from Ignorance*. Penn State University Press. ◦\n- Weisberg, M. (2006). Robustness analysis. *Philosophy of Science* 73, 730–742. ◦\n- Whewell, W. (1840/1847). *The Philosophy of the Inductive Sciences*. ◦\n- White, R. (2005). Epistemic permissiveness. *Philosophical Perspectives* 19, 445–459. https://philpapers.org/rec/WHIEP ✔\n- Williamson, J. (2010). *In Defence of Objective Bayesianism*. OUP. ◦\n- Williamson, T. (2000). *Knowledge and its Limits*. OUP. ✔ (via SEP)\n- Wimsatt, W. (1981). Robustness, reliability, and overdetermination. In Brewer & Collins (eds.), *Scientific Inquiry and the Social Sciences*. ◦\n- Wittgenstein, L. (1969). *On Certainty*. Blackwell. ◦\n- Worrall, J. (2007). Evidence in medicine and evidence-based medicine. *Philosophy Compass* 2, 981–1022. ◦\n\n### Unverified points flagged in the text\n- Royall's exact benchmark wording (8 = \"fairly strong\", 32 = \"strong\").\n- Exact Sellars quotation (EPM §38) and Wittgenstein *On Certainty* §341 wording.\n- Page ranges marked ◦ above.\n- Williamson (Jon), *Evidential Pluralism in the Social Sciences* — title and year not checked.\n- Cova et al. 2021 x-phi replication rate — not checked.\n- The 2025 *Erkenntnis* paper \"A New Look at Keynesian Weight of Evidence\" — existence verified, content not read.\n"}