{"path":"research/what-counts-as-evidence/law-intelligence-history-journalism-forensics.md","content":"# Evidence in the non-scientific professions: law, intelligence, history, journalism, forensic science\n\n*Research report for Deliberus, 2026-09-26. Scope: how the professions that decide facts under\nadversarial or high-stakes conditions classify evidence, grade it, combine it, and state their\nconclusions, and what a claim-graph argumentation platform can take from them.*\n\n**How to read the verification marks.** Everything quoted in \"double quotes\" with a source was\nread in the primary document during this pass, unless marked otherwise. Items marked\n**[recalled, not re-verified]** come from standard literature I know well but did not re-open in\nthis session; they are safe to use internally and should be checked before any outward citation.\nItems marked **[contested]** are live scholarly disputes, not settled practice.\n\n**Two corrections to the brief before starting.**\n\n1. The five Carneades proof standards named in the brief (scintilla, preponderance, clear and\n   convincing, beyond reasonable doubt, dialectical validity) are **not** all in Gordon, Prakken &\n   Walton (2007). Reading the 2007 *Artificial Intelligence* paper itself: it defines exactly three\n   — **SE** (scintilla of evidence), **BA** (best argument) and **DV** (dialectical validity) —\n   orders them DV > BA > SE, and says plainly \"we do not claim that any of the standards defined\n   here adequately model legal proof standards\", listing preponderance and beyond-reasonable-doubt\n   as \"further work\". The numeric-threshold versions of preponderance, clear-and-convincing and\n   BRD come from the later Gordon & Walton (2009) \"Proof burdens and standards\" **[recalled, not\n   re-verified]**. Details in §1.9.\n2. Schum's later writing uses **\"divergent\"** where the 1994 book and much secondary literature\n   say **\"conflicting\"** for dissonant evidence about *different* events. Both names are in use for\n   the same pattern (§2.4).\n\n---\n\n## 0. The one-paragraph answer\n\nAcross all five professions the same four moves recur, and none of them is \"count the supporting\nitems\". **(a) Separate the source from the report**: who said it, and how good is that sayer in\ngeneral, is a different question from whether this particular statement is true (NATO's two\naxes; Schum's competence-versus-credibility; history's external versus internal criticism; law's\nauthentication versus weight). **(b) Break credibility into named attributes, each of which can be\nattacked separately** (Schum: veracity, objectivity, observational sensitivity for testimony;\nauthenticity, accuracy, reliability for tangible evidence; the hearsay tradition's perception,\nmemory, narration, sincerity). **(c) Weigh evidence by how much it discriminates between\nhypotheses, not by how much of it there is** (the likelihood ratio in forensic science; Heuer's\ndiagnosticity; Schum's inferential force; Rule 403's exclusion of \"needlessly … cumulative\nevidence\"). **(d) Treat agreement between reports as worth something only in proportion to their\nindependence** (Schum's corroborative redundancy; Langlois & Seignobos's \"necessity of proving\nthem to be independent\"; the WMD Commission's finding that daily briefs \"left an impression of\nmany corroborating reports where in fact there were very few sources\"). And a fifth, specific to\nlaw: **the threshold for accepting a claim is set by the question and the cost of error, not by a\nuniversal scale** (proof standards, burdens, Carneades). Deliberus today implements none of (a),\n(c), (d) or the fifth, and implements (b) only through Walton critical questions whose answers are\ncounted rather than weighed. §7 says what to build.\n\n---\n\n## 1. Law\n\n### 1.1 Categories of evidence\n\nThe standard common-law typology **[recalled; textbook consensus, e.g. McCormick on Evidence;\nLempert, Gross & Liebman]**:\n\n| Category | What it is | Characteristic doubt |\n|---|---|---|\n| **Testimonial** | a person's assertion, oral or written, offered for its truth | the witness's perception, memory, narration, sincerity |\n| **Documentary** | a writing or recording offered for its content | authenticity; whether the content is an assertion (and then hearsay) |\n| **Real / physical** | the object itself (the knife, the bloodstain) | authenticity; chain of custody; alteration |\n| **Demonstrative** | a representation made to illustrate (map, model, chart, animation) | accuracy of the representation |\n\nCrossing this is **direct vs circumstantial**: direct evidence, if believed, establishes the fact in\nissue without further inference (an eyewitness who saw the stabbing); circumstantial evidence\nestablishes a fact from which the fact in issue is inferred (the defendant's fingerprint on the\nknife). The distinction is relative to the proposition: Tillers & Schum note Wigmore's point that\n\"the classification of some proposition depends upon its position in a chain of reasoning\" — an\nitem can be direct on one proposition and circumstantial on a higher one (Tillers & Schum 1988,\n*Cardozo L. Rev.*). Schum later adds that, apart from contrived cases, **all** inference from\nevidence is \"catenated\": \"apart from just two instances, all inductive or probabilistic reasoning\nbased on evidence seems to be catenated or cascaded in nature\" (Tillers & Schum 1988). Direct\neyewitness testimony still carries one hidden link — the witness's credibility.\n\nThe real/demonstrative split reappears almost verbatim in Schum's intelligence work (§2.2), which\ncites it to Lempert, Gross & Liebman.\n\n### 1.2 Relevance and probative value vs prejudice (FRE 401–403)\n\n- **FRE 401** (relevance) **[recalled text]**: evidence is relevant if \"(a) it has any tendency to\n  make a fact more or less probable than it would be without the evidence; and (b) the fact is of\n  consequence in determining the action.\" The threshold is deliberately minimal (\"any tendency\").\n  Note the probabilistic language: relevance is defined as *change in probability*, i.e. a\n  likelihood ratio different from 1 — the same quantity as forensic science's LR and Heuer's\n  diagnosticity.\n- **FRE 402**: relevant evidence is admissible unless excluded; irrelevant evidence is not.\n- **FRE 403** (verified text): \"The court may exclude relevant evidence if its probative value is\n  substantially outweighed by a danger of one or more of the following: unfair prejudice,\n  confusing the issues, misleading the jury, undue delay, wasting time, or needlessly presenting\n  cumulative evidence.\" The Advisory Committee Note defines unfair prejudice as \"an undue tendency\n  to suggest decision on an improper basis, commonly, though not necessarily, an emotional one.\"\n\nTwo things in 403 matter here. First, **probative value and weight are separate from\nrelevance**: relevance is a gate (does it move probability at all?), probative value is a\nmagnitude. Second, **\"needlessly presenting cumulative evidence\"** is a codified recognition that\nmore items saying the same thing do not add proportionate weight and can mislead — law's version\nof Schum's redundancy (§2.4).\n\n### 1.3 Hearsay and its exceptions — the epistemology\n\n**Definition** (FRE 801, verified): a *statement* is \"a person's oral assertion, written assertion,\nor nonverbal conduct, if the person intended it as an assertion\"; hearsay is a statement that\n\"(1) the declarant does not make while testifying at the current trial or hearing; and (2) a party\noffers in evidence to prove the truth of the matter asserted.\" FRE 801(d) removes two classes from\nthe definition altogether: certain prior statements of a witness who *is* now subject to\ncross-examination, and statements of an opposing party.\n\n**Why it is excluded.** The canonical rationale is that the *declarant's* testimonial capacities\ncannot be tested. Tribe's \"Triangulating Hearsay\" (1974) organised the four **testimonial\ninfirmities**: **perception, memory, narration (ambiguity), and sincerity** — the hearsay\n\"dangers\" that oath, demeanour and above all **cross-examination** are supposed to expose\n(Morgan 1948; Tribe 1974) **[recalled, not re-verified; standard]**. Wigmore called\ncross-examination \"beyond any doubt the greatest legal engine ever invented for the discovery of\ntruth\" **[recalled]**.\n\n**The exceptions** follow a single logic: admit the statement when **circumstances substitute for\ncross-examination** — i.e. when some structural feature makes one of the four dangers unlikely\n**[recalled structure of FRE 803/804/807]**:\n\n- *FRE 803 (availability immaterial)*: present sense impression and excited utterance (little time\n  to fabricate → sincerity, memory); statements for medical diagnosis (incentive to be truthful);\n  recorded recollection; **business and public records** (regularity, duty, routine → sincerity,\n  memory); learned treatises; ancient documents.\n- *FRE 804 (declarant unavailable)*: former testimony (was cross-examined then); dying\n  declarations; **statements against interest** (people rarely lie against themselves → sincerity).\n- *FRE 807*: residual exception for statements with equivalent \"guarantees of trustworthiness\".\n\n**The epistemological lesson**: the hearsay rule is not \"second-hand information is worthless\" —\nhistorians explicitly reject that reading (Gottschalk: \"hearsay evidence would not be discarded by\nthe historian, as it would be by a law court merely because it is hearsay\"). It is **\"a claim whose\ncredibility attributes cannot be interrogated should not be weighted as if they had been.\"** Each\nexception names a reason why a particular attribute need not be interrogated. For a claim graph\nthis maps directly: a reported statement (Deliberus's REPORTS edge) carries the declarant's four\nattributes *unexamined*; answering critical questions about them is the platform's substitute for\ncross-examination; and an exception-type fact (\"statement against interest\", \"made in routine\nrecord-keeping\") is itself a claim that justifies a higher starting weight.\n\n### 1.4 Authentication, chain of custody, best evidence\n\n- **Authentication** (FRE 901(a), verified): \"the proponent must produce evidence sufficient to\n  support a finding that the item is what the proponent claims it is.\" 901(b) lists ten\n  illustrative methods, including testimony of a witness with knowledge, comparison with\n  authenticated specimens, \"distinctive characteristics\", public records, ancient documents\n  (20+ years), and \"evidence describing a process or system and showing that it produces an\n  accurate result\".\n- **Chain of custody** is the usual way to authenticate fungible real evidence (a blood sample,\n  drugs): a documented sequence of custodians showing the item was not substituted or altered.\n  In US practice gaps usually go to **weight, not admissibility** once a threshold is met\n  **[recalled, standard]**. Schum treats chain of custody for intelligence as a chain of reasoning\n  whose every link (each custodian, each transformation, each translation) is a possible\n  credibility failure (Schum et al. 2009, \"Analyzing Evidence and its Chain of Custody\").\n- **Best evidence rule** (FRE 1002, **[recalled text]**): \"An original writing, recording, or\n  photograph is required in order to prove its content unless these rules or a federal statute\n  provides otherwise\" — with duplicates admissible under 1003. The epistemic point: when the\n  *content* of a document is at issue, a paraphrase or recollection of it adds a transmission\n  link that the original avoids. This is the legal form of \"quote, don't paraphrase\".\n\n### 1.5 Expert evidence gatekeeping\n\n- **Frye v. United States** (D.C. Cir. 1923): scientific evidence admissible if the technique is\n  \"generally accepted\" in the relevant scientific community **[recalled]**.\n- **Daubert v. Merrell Dow** (1993): the trial judge is gatekeeper; non-exclusive factors —\n  (1) whether the theory/technique **can be and has been tested**; (2) **peer review and\n  publication**; (3) **known or potential error rate**; (4) existence and maintenance of\n  **standards** controlling operation; (5) **general acceptance** **[recalled; consensus\n  statement]**.\n- **Kumho Tire v. Carmichael** (1999): gatekeeping applies to *all* expert testimony, technical and\n  experience-based, not only \"scientific\" **[recalled]**.\n- **FRE 702 as amended 1 Dec 2023** (verified): expert opinion is admissible \"if the proponent\n  demonstrates to the court that it is more likely than not that: (a) the expert's … specialized\n  knowledge will help the trier of fact …; (b) the testimony is based on sufficient facts or data;\n  (c) the testimony is the product of reliable principles and methods; and (d) the expert's opinion\n  reflects a reliable application of the principles and methods to the facts of the case.\" The\n  committee note says courts had incorrectly treated sufficiency of basis and application as\n  questions of *weight* rather than admissibility, and that \"each expert opinion must stay within\n  the bounds of what can be concluded from a reliable application of the expert's basis and\n  methodology\" — flagging forensic claims of \"absolute or one hundred percent certainty\".\n  **Note (c) vs (d)**: validity of the *method* and validity of *this application* are separate\n  gates. PCAST (§5.2) says its \"foundational validity\" and \"validity as applied\" were written to\n  correspond exactly to 702(c) and 702(d).\n- **England and Wales, Criminal Practice Directions 2023, 7.1.2** (verified via CPS guidance; the\n  same list was CrimPD 2015 19A.5): factors bearing on reliability of expert opinion include\n  \"(a) the extent and quality of the data on which the expert opinion is based; (b) the validity of\n  the methodology employed by the expert; (c) … whether the opinion properly explains how safe or\n  unsafe the inference is …; (d) … the degree of precision or margin of uncertainty …; (e) the\n  extent to which any material … has been reviewed by others with relevant expertise …; (f) the\n  extent to which the expert's opinion is based on material falling outside the expert's own field\n  of expertise; (g) the completeness of the information which was available to the expert …;\n  (h) if there is a range of expert opinion on the matter in question, where in the range the\n  expert's own opinion lies and whether the expert's preference has been properly explained; and\n  (i) whether the expert's methods followed established practice in the field …\". 7.1.3 adds the\n  flaws to watch for: an untested hypothesis, \"an unjustifiable assumption\", \"flawed data\", a\n  technique \"not properly carried out or applied\", or \"an inference or conclusion which has not\n  been properly reached\".\n\n**For Deliberus this list is a ready-made critical-question set** for the *expert opinion* scheme,\nricher than Walton's six, and two items are absent from Walton entirely: **(h) where in the\ndistribution of expert opinion this expert sits**, and **(g) whether the expert had the\ncomplete information**. Both are exactly the \"counted not weighed\" gap: an expert-opinion edge\nshould carry *range position*, not just *expert: yes*.\n\n### 1.6 Burdens and standards of proof\n\n**Two burdens** (Thayer 1898; Wigmore; McCormick — quoted in Gordon, Prakken & Walton 2007): the\n**burden of production** (\"the liability to an adverse ruling … if evidence on the issue has not\nbeen produced\") and the **burden of persuasion** (\"if the party having that burden has failed to\nsatisfy it, the issue is to be decided against that party\"). Wigmore: \"the risk of\nnon-persuasion never shifts, but the duty of producing evidence … does have this shifting\ncharacteristic.\" Both burdens can be **allocated per issue**: the plaintiff bears production for\nthe operative facts, the defendant for exceptions; in criminal law, once the defendant produces\nsome evidence of an exception (e.g. self-defence), the prosecution bears persuasion that it does\nnot hold.\n\n**Standards of proof (US)** **[recalled; consensus]**:\n\n| Standard | Where | Usual gloss | Survey-based numeric readings |\n|---|---|---|---|\n| Preponderance of the evidence | civil | \"more likely than not\" | >0.5 |\n| Clear and convincing | fraud, civil commitment (*Addington v. Texas*, 1979), termination of parental rights | \"highly probable\" | roughly 0.7–0.8 in judge surveys |\n| Beyond reasonable doubt | criminal (*In re Winship*, 1970) | no reasonable doubt | judge surveys centre around 0.85–0.95 (e.g. McCauliff 1982, ~90%) |\n\nThe numeric readings are **[recalled, approximate; contested]** — US courts overwhelmingly refuse\nto quantify, and the Laudan (2006) critique is that BRD has no stable meaning at all.\nEngland uses \"sure\" for the criminal standard; Sweden \"ställt utom rimligt tvivel\"\n**[recalled]**.\n\n**Civil-law \"intime conviction\".** French criminal procedure historically instructed jurors that\nthe law \"does not ask them to account for the means by which they were convinced\" and asks a\nsingle question: \"Avez-vous une intime conviction?\" (CPP art. 353) **[recalled]**. After *Taxquet\nv. Belgium* (ECtHR 2010) France required assize courts to give reasons (the *feuille de\nmotivation*, law of 10 Aug 2011) **[recalled]**. Clermont & Sherwin (2002) argue the civil-law\nworld applies essentially one high standard across civil and criminal cases where common law uses\npreponderance for civil **[recalled; contested]**. The contrast matters for Deliberus: *intime\nconviction* is the \"bare judgment\" model (a verdict with no inspectable reasons), and the European\ntrend has been toward requiring reasons — the direction Deliberus already chose.\n\n**Transferable principle**: the standard is **a property of the question and the cost of each kind\nof error**, not of the evidence. The same evidence can satisfy preponderance and fail BRD. A\nsingle universal badge scale cannot represent this.\n\n### 1.7 The probability debate\n\n- **Conjunction paradox** (Cohen 1977): burdens attach to *each element* of a claim. If a plaintiff\n  proves elements A and B each at 0.6, preponderance is met on each, but P(A∧B) = 0.36 if\n  independent — so a probabilistic reading of \"more likely than not\" would require the *conjunction*\n  to exceed 0.5, which is not what courts do **[recalled; the paradox is uncontested, its\n  resolution is contested]**. Allen & Pardo call it \"the conjunction problem, which arises from the\n  fact that proof burdens apply to the individual elements of crimes, civil claims, and defenses\n  rather than a party's case as a whole\" (Allen & Pardo 2019, verified). **Deliberus relevance**:\n  its *necessary* decomposition uses **min**, which is element-wise acceptance — the court's\n  practice — rather than product. Legal scholarship supplies both the defence (courts do it;\n  elements are correlated through a single story) and the attack (it overstates the whole).\n- **Naked statistical evidence** (Cohen 1977, *The Probable and the Provable*) **[recalled]**:\n  *Gatecrasher* — 1,000 attend a rodeo, 499 paid; the organiser sues an arbitrary attendee; P(did\n  not pay) = 0.501 > 0.5, yet no court would find liability. *Blue Bus* — from *Smith v. Rapid\n  Transit* (Mass. 1945) and Tribe (1971): a bus hit the plaintiff; 80% of buses on the route belong\n  to the Blue company; the statistic alone does not suffice. Explanations offered: lack of\n  **individualised** evidence (Thomson 1986), failure of **sensitivity** / knowledge (Enoch,\n  Spectre & Fisher 2012), low **Keynesian weight** (Cohen; Nance), incentive effects, and the\n  empirical **Wells effect** — people refuse to find liability on naked statistics even when their\n  subjective probability equals that from an eyewitness (Wells 1992) **[all recalled; contested\n  which explanation is right]**. **Deliberus relevance**: the \"Keynesian weight\" reading is the one\n  Deliberus already adopted (Aug 2026, direction vs weight). Law's refusal to act on a 0.8 with\n  weight near zero is the institutional precedent for \"an unexamined claim carries no number\".\n- **Relative plausibility theory** (Allen & Pardo; verified 2019 *IJEP* 23(1–2):5–59): proof is\n  **explanatory and comparative**, not probabilistic per element. \"The central fact-finding task is\n  not to attach probabilities to the individual elements; it is to determine whether potential\n  explanations of the evidence and events satisfy the applicable standard of proof\", and \"the\n  explanatory account is inherently comparative\". Under preponderance the fact-finder asks which\n  available explanation is best; under BRD, conviction requires that the prosecution's\n  explanation be plausible \"and there is no plausible defense explanation\". Criteria of explanatory\n  quality: \"consistency, coherence, fit with background knowledge, simplicity, absence of gaps, and\n  the number of unlikely assumptions\". **[contested]**: Nance, Clermont, Schwartz & Sober, and\n  others argue RPE is probabilism in disguise or a \"friendly amendment\" to likelihoodism\n  (e.g. the 2019 *IJEP* responses).\n- **Story model** (Pennington & Hastie 1986, 1991, 1992 **[recalled]**): jurors construct\n  narratives and choose the verdict whose story best fits; story acceptance depends on **coverage**\n  (accounts for the evidence), **coherence** (consistency, plausibility, completeness) and\n  **uniqueness** (no rival story equally good). Descriptively robust; its normative status is the\n  question.\n- **Bayesian networks for legal evidence** (Fenton, Neil & Lagnado 2013, *Cognitive Science*\n  37(1):61–102, verified): legal arguments built from a small set of reusable **idioms** — the\n  *evidence idiom*; the **evidence-accuracy idiom** (every evidence node gets a parent \"accuracy\"\n  node, so a DNA match is conditioned on whether the test was accurate; knowing the test was\n  accurate \"explains away\" the alternative cause); *opportunity* (a necessary condition, e.g.\n  presence at scene); *motive* (raises probability, not necessary); the **dependency idiom**\n  (two pieces of evidence not independent — e.g. two CCTV cameras of the same make and fault);\n  the **alibi idiom** (where the hypothesis itself influences the accuracy of the evidence, as when\n  the alibi witness is the defendant's friend); and *explaining away*. **Deliberus relevance**:\n  the evidence-accuracy idiom *is* the source/report split in graph form; the dependency idiom *is*\n  the shared-origin edge; the alibi idiom is \"the claim under test influences the reliability of\n  its own evidence\", which is Deliberus's circularity hazard.\n- **Bex's hybrid theory** (Bex 2011, *Arguments, Stories and Criminal Evidence*; Bex, van Koppen,\n  Prakken & Verheij 2010, *AI & Law* 18:123–152) **[recalled]**: stories (causal-abductive,\n  judged by coherence and fit to story schemes) and arguments (evidential, from sources via\n  schemes like witness testimony) are combined: arguments support or attack story elements, and\n  stories are compared on how much evidence they explain and contradict. Critical questions\n  apply both to arguments *and to stories* (is the story internally consistent? does it fit a\n  known scheme? does it have gaps?).\n\n### 1.8 What the probability debate settles and what it does not\n\n**Settled**: the likelihood ratio is the correct measure of the *probative force* of a single item\nrelative to two stated propositions (there is no serious dissent on this inside forensic science,\nand Schum uses it throughout); element-wise proof is how courts operate; people are poor\nintuitive Bayesians (weak evidence effect, prosecutor's fallacy, §5). **Contested**: whether the\n*final* decision should be probabilistic at all (Allen & Pardo vs the Bayesians); what explains the\nnaked-statistics intuition; whether Keynesian weight is a separate dimension or reducible to\nexpected information gain. Deliberus has chosen a side — weight as a separate dimension, no\nverdicts — and the legal literature is the richest source of arguments *for* that side\n(Cohen, Nance) and *against* it (Kaye, Schwartz & Sober).\n\n### 1.9 Carneades: proof standards as per-statement functions\n\nGordon, Prakken & Walton (2007), \"The Carneades model of argument and burden of proof\",\n*Artificial Intelligence* 171(10–15):875–896 (read in full text this session):\n\n- **Premise types** that allocate burden: \"those that must always be supported with further\n  grounds; those that can be assumed until they have been questioned; and those that do not hold\n  in the absence of evidence to the contrary. In this article we will call the first kind 'ordinary\n  premises', the second kind 'assumptions' and third kind 'exceptions'.\" Critical questions map\n  onto these: some CQs merely *ask* (shift the burden back to the proponent — an assumption),\n  others must be *backed by evidence* before they bite (an exception).\n- **Statements have a dialectical status**: stated, questioned, accepted, rejected.\n- **Proof standards are assigned per statement**: \"A statement s is acceptable in an argument\n  graph G if and only if it satisfies its proof standard.\" Definitions (verbatim): \"**SE (Scintilla\n  of Evidence).** A statement meets this standard iff it is supported by at least one defensible pro\n  argument.\" \"**BA (Best Argument).** … supported by some defensible pro argument with priority\n  over all defensible con arguments.\" \"**DV (Dialectical Validity).** … supported by at least one\n  defensible pro argument and none of its con arguments are defensible.\" Every standard has a\n  **complement** (swap pro and con), and \"in principal a statement can satisfy both a proof\n  standard and its complement, highlighting the difference between acceptability and truth.\"\n- Verheij's four roles for critical questions, quoted in the paper: CQs can (1) question a premise,\n  (2) point to exceptional situations (these **undercut**), (3) set conditions for proper use\n  (these refute implicit assumptions), (4) point to counterarguments (these **rebut**). This is the\n  rebutting/undercutting distinction the Deliberus brief lists as untyped.\n- **Gordon & Walton (2009)** extended this to preponderance (PE: max weight of applicable pro\n  arguments exceeds max con), clear and convincing (PE plus the pro weight exceeds a threshold α and\n  the pro–con difference exceeds β), and beyond reasonable doubt (CCE plus max con weight below γ)\n  **[recalled, not re-verified: check the exact definitions before citing]**. Van Gijzel & Prakken\n  (2012) relate Carneades to ASPIC+ and Dung semantics.\n\n**Why this matters most for Deliberus**: Carneades demonstrates, in a formal system from the same\nWalton lineage Deliberus already uses, that (i) acceptability is **relative to a per-claim\nstandard**, (ii) the same argument graph can meet a claim's standard and its complement at once,\nand (iii) **burden allocation is expressed by the type of premise a CQ attacks**. All three are\nabsent from the current QEM badge.\n\n---\n\n## 2. Wigmore charts and the \"new evidence scholarship\"\n\n### 2.1 Wigmore's chart method and its modern revival\n\nWigmore published the chart method in *The Principles of Judicial Proof* (1913; 2nd ed. 1931;\n3rd ed. as *The Science of Judicial Proof*, 1937). It has two halves (Schum 1999, verified):\n\"The analytic part consists of recording, on what Wigmore called a **key list**, propositions that\ndefine what we would today term the nodes on an inference network … In addition, Wigmore also\nsuggested that we record inductive **generalizations** that provide a warrant or a license for\nreasoning from one node to another … The synthetic part consists of drawing a **chart** showing the\ninferential relationships among the nodes.\"\n\nWigmore's symbol set (Tillers & Schum 1988, verified) distinguished: offeror (prosecution /\ndefence); **affirmative vs negative** testimonial and circumstantial evidence; **ancillary**\nevidence that is **explanatory** (weakens a link) or **corroborative** (strengthens it);\nevidence observable by the fact-finder; judicially noticed facts; and graded arrows expressing\nthe \"verbally-expressed strength\" of each link. Twining called the method a \"lead balloon\" for\nits long neglect (Twining 1985, cited in Tillers & Schum).\n\n**Probanda.** Schum (1999, verified): \"the ultimate probandum … is the major hypothesis, civil\nclaim, or criminal charge at issue\"; the elements the substantive law requires are **penultimate\nprobanda** (for first-degree murder: victim killed; defendant killed victim; intent; premeditation).\nThen the striking observation: \"inferences in legal contexts are well-structured at the top … This\nis an inferential luxury that is not provided in any other context I know about. There are no\n'substantive laws' … in fields such as science, medicine, history, or intelligence analysis that\ntell us what penultimate matters we must prove.\" **For Deliberus this is the decomposition\nproblem stated from law's side**: the law hands the analyst a mother claim's necessary parts;\neverywhere else, decomposing the mother is itself contestable work.\n\n**Directly relevant vs ancillary evidence** (Schum 1999, verified): evidence is directly relevant\n\"if a plausible chain of reasoning can be constructed that links the evidence to the penultimate\nprobandum\", and **ancillary** \"if it bears on the strength or weakness of links … in chains of\nreasoning set up by directly relevant evidence. In short, ancillary evidence is evidence about\nother evidence … We might also term such evidence metaevidence.\" Schum (2001, *Cardozo*):\n\"generalizations plus ancillary evidence form the 'glue' holding an argument together.\"\n\n### 2.2 Anderson, Schum & Twining, *Analysis of Evidence* (2nd ed., CUP 2005)\n\nThe \"modified Wigmorean analysis\" proceeds in seven steps **[recalled; check chapter 5 before\nciting]**: (1) clarify the **standpoint** (who is analysing, at what stage, for what purpose);\n(2) formulate the potential **ultimate probandum**; (3) formulate the **penultimate probanda**;\n(4) formulate the provisional **theory of the case**; (5) record the data — the **key list**;\n(6) prepare the **chart**; (7) refine and complete. The standpoint step is load-bearing: Schum\n(1999, verified) reports Anderson and Twining's argument that \"the standpoint a person adopts\nstrongly influences the ideas and evidence one will generate …, the arguments one will then\nconstruct … and the inferential force one will assign\", and that failure to declare it \"causes\nmany difficulties\". Chapters also treat outlines and chronologies (the outline method: each\npenultimate probandum as a heading), narrative and its dangers, and probabilities.\n\n**Generalizations.** Every inferential link rests on a generalization (\"people who flee usually\nhave something to hide\"). Twining's view (e.g. \"Necessary but dangerous? Generalizations and\nnarrative in argumentation about 'facts' in criminal process\", 1999) is that they are\n**indispensable and the chief locus of prejudice** — stereotype enters here **[recalled]**. Schum\n(2001, verified): \"what is at issue in any legal contest is whether an asserted generalization\napplies in the particular situation at hand\", which is what ancillary evidence tests.\n**Deliberus relevance**: a Walton scheme's warrant *is* the generalization; making the\ngeneralization an explicit, attackable claim (rather than implicit in the scheme label) is what\nWigmoreans insist on.\n\n### 2.3 Schum's substance-blind classification of evidence\n\nSchum, *The Evidential Foundations of Probabilistic Reasoning* (Wiley 1994; Northwestern UP\nreprint 2001), chapters 3–8; summarised in Schum 1999, 2001, 2009 and Schum et al. 2009 (all\nverified in text). The core idea: \"although evidence varies substantively in a near infinite\nfashion … there is a finite and manageable number of logically distinguishable species\" (Tillers &\nSchum 1988). \"Substance-blind\" means the classes \"are recurrent and apply regardless of the\nsubstance or content of the evidence\" (Schum 2011 chapter abstract).\n\n**The organising question** (Schum et al. 2009, OIC paper, verified): \"How do you, the analyst,\nstand in relation to this item of evidence? Can you examine it for yourself to see what events it\nmight reveal? If you can, we say that the evidence is **tangible** … But suppose instead you must\nrely upon other persons … Their reports to you about these events are examples of **testimonial**\nevidence.\"\n\n**Forms of evidence** (Schum 1999: \"the forms identify whether the evidence is tangible, testimonial\nin various ways, is missing, or refers to accepted facts from authoritative records\"):\n\n1. **Tangible evidence**\n   - **Real tangible**: \"a thing itself and has only one major believability attribute:\n     **authenticity**. Is this object what it is represented as being or is claimed to be?\"\n   - **Demonstrative tangible**: a representation (map, image, sensor record, table). Three\n     attributes: **authenticity**; **accuracy** (\"the extent to which the device that produced the\n     representation … had a degree of sensitivity (resolving power or accuracy) that allows us to\n     tell what events were observed\"); **reliability** (\"A system, sensor, or test of any kind is\n     reliable to the extent that the results it provides are repeatable or consistent\").\n2. **Testimonial evidence**\n   - **Unequivocal**, subdivided by *how the source knows*: (a) **based on direct observation**;\n     (b) **obtained at second hand** (hearsay: \"I … heard about its occurrence … from another\n     person\"); (c) **based on opinion** (\"I did observe events C and D and inferred from them that\n     event E definitely occurred\") — which Schum reclassifies as *equivocal* because it requires\n     assessing the source's reasoning.\n   - **Equivocal**: **completely equivocal** (\"I don't know\", \"I can't remember\");\n     **probabilistically equivocal** (\"I'm 60 percent sure\"; \"I'm fairly sure\"); and testimony\n     based on opinion.\n3. **Missing evidence**: evidence expected but not produced — whose absence can itself be\n   informative (the adverse-inference tradition in law; Heuer's absence of evidence, §3.4). Schum\n   treats the *failure to obtain* as an item that needs its own chain of reasoning (why is it\n   missing: never existed, not looked for, concealed?) **[recalled detail; the category is\n   verified]**.\n4. **Accepted facts / authoritative records**: things taken as given without credibility inquiry\n   (tide tables, judicially noticed facts, well-established scientific constants). Locke already\n   said their probability \"rises near to certainty\" (Schum 2009, verified).\n\nCrossing all forms: **positive vs negative evidence** (that E occurred vs that E did not occur).\n\n**The three credentials of evidence** **[recalled; this triad runs through all of Schum's work]**:\n**relevance** (does a defensible chain of reasoning link it to a probandum?), **credibility /\nbelievability** (did the reported event occur?), and **inferential (probative) force or weight**\n(how much does it, if credible, shift belief between hypotheses?). Schum's formal measure of force\nis the likelihood ratio; the 2025 *AI & Law* paper building on him notes \"Schum used the term\n'inferential force' … for the likelihood ratio\", with its logarithm being Good's \"weight of\nevidence\".\n\n**Competence is separate from credibility** — for human sources (Schum et al. 2009, \"Analyzing\nEvidence and its Chain of Custody\", verified):\n\n- **Competence** = **access** (\"did this source have access to, or did actually observe, the events\n  being reported?\") + **understanding** (\"the source understood the events … well enough to provide\n  … an intelligible account\").\n- **Credibility** = **veracity** (\"is this source telling us about an event he/she believes to have\n  occurred?\"), **objectivity** (\"did this source base a belief on sensory evidence … or … because\n  this source expected or wished them to occur?\"), and **observational sensitivity** under the\n  conditions of observation (\"if the source did base a belief on sensory evidence, how good was\n  this evidence?\").\n- And the warning, verbatim: \"the competence and credibility of HUMINT sources are entirely\n  distinct. Competence does not entail credibility, nor does credibility entail competence.\n  Confusing these two characteristics invites inferential disaster … 'We can believe what this\n  source has told us because he had good access'. This is a glaring non sequitur; the source may\n  have had all the access in the world but be lying through his teeth.\"\n\nMap to the hearsay dangers: veracity ≈ sincerity; objectivity ≈ (part of) perception and memory\nas distorted by expectation; observational sensitivity ≈ perception; narration/ambiguity is the\nsource of *equivocal* testimony. Schum and Morris (2007, *Law, Probability and Risk* 6:247–274)\nturned these into structured question sets for intelligence sources (the MACE system)\n**[recalled]**.\n\n**Why \"five reasons\"** (Schum 2009, verified): \"All conclusions reached from evidence are\nnecessarily probabilistic for five reasons. Our evidence is 'never complete', is usually\n'inconclusive', is frequently 'ambiguous', is commonly 'dissonant' to some degree and comes to us\nfrom 'sources' [having imperfect credibility].\" Each is a distinct uncertainty a claim graph\nshould be able to represent separately: incompleteness (missing evidence, gaps), inconclusiveness\n(LR near 1), ambiguity (equivocal testimony, contested meaning), dissonance (§2.4), and source\ncredibility.\n\n### 2.4 Schum's recurrent combinations of evidence\n\nSchum 2009 (*LPR* 8:197–231) and Schum et al. 2009 (both verified) give three families.\n\n**(A) Harmonious evidence**\n- **Corroborative**: \"when two or more sources (of whatever kind) report or reveal the '**same\n  event**'. For example, two human witnesses and a photograph, all tell us that event E occurred.\"\n  A second species: \"ancillary evidence favourable to the credibility of a source that provides\n  directly relevant evidence about some event\" (i.e. corroboration *of the source*, not of the\n  event). Schum (1994) thus distinguishes **two forms of corroboration**, which Godden (2014,\n  *Argumentation* 28, \"Modeling Corroborative Evidence\") maps to convergent and premise-bolstering\n  types.\n- **Convergent**: \"evidence from two or more sources about '**different**' events, all of which\n  favour the same conclusion.\" \"A very important subtlety associated with convergent evidence is\n  their possible '**synergism**': two or more convergent items of evidence can have increased\n  inferential force when they are considered together than they would have if they are considered\n  separately or independently.\"\n\n**(B) Dissonant evidence** (\"directionally inconsistent\"; \"two basic ways that are often confused\")\n- **Contradictory**: reports of \"mutually exclusive events, those which cannot happen together\"\n  (E and not-E; or X in Houston and X in Boston at the same moment). \"Evidential contradictions are\n  always resolved on credibility grounds\" — because one source must be \"untruthful, not objective,\n  or mistaken\". Crucially: \"counting heads assumes that all of the four sources … have equal\n  credibility … what matters in resolving evidential contradictions is the **aggregate credibility\n  of the sources on either side**.\"\n- **Divergent** (called **conflicting** in the 1994 book): evidence about \"two events that can\n  happen together but simply point us in different inferential directions\". Resolving divergence is\n  \"a more complex matter\": credibility, but also understanding, and sometimes \"considering the\n  'joint occurrence' of the events reported\".\n\n**(C) Redundant evidence** (\"the first item acts to make subsequent items redundant to some degree\")\n- **Corroborative redundancy**: repeated reports of the *same* event. \"If we believed our first\n  source … was perfectly credible, then additional accounts of this same event would be completely\n  redundant and would add no inferential force … The second account of this event springs to life\n  to the extent that we find the first source not credible. The third account … springs to life to\n  the extent to which we believe the second source not to be credible, and so on.\" Exception\n  noted in Schum et al. 2009: when sources \"collaborated in deciding what to tell us, or … one\n  source influenced or coerced another\" — then the second report is not independent at all.\n- **Cumulative redundancy**: reports of *different* events where \"if event E occurred, this would\n  reduce the force of knowing that F occurred\" — \"I have chosen the word 'cumulative' that is used\n  in the field of law to indicate evidence that adds nothing new.\"\n\n**Plus**: **synergism** (the opposite of redundancy for convergent items) and **directional\nchange** — Schum's \"redundance measure\" R, defined on likelihood ratios, takes values that\nclassify an interaction as redundancy (0 < R ≤ 1), synergy, or a reversal of direction (the 2025\n*AI & Law* paper, verified, formalises these with weight-of-evidence measures). The foundational\nformal result is **Schum & Kelly (1973)**, *OBHP* 10:404–423 (verified abstract/excerpt): for\nmultiple unreliable sources reporting on the same event, a joint conflicting report favours the\nmajority side \"iff the joint reliability of the majority sources exceeds the joint reliability of\nthe sources in the minority. Thus, **source reliability is more important than the extent of\nsource agreement** … the practice of using an odd number of sources to resolve potential reporting\nconflicts ignores what is crucial.\"\n\n**Chains of reasoning** (catenated / cascaded inference): each link from evidence E* (the report)\nthrough E (the event) through interim probanda to the ultimate probandum is a separate source of\nuncertainty, and force is attenuated at every link. Schum's LR analysis of the Sacco–Vanzetti\nwitnesses (Schum 2001, verified) shows force \"springs to life\" or vanishes depending on\ncredibility of *other* witnesses — i.e. **the force of an item is not a property of the item alone\nbut of the item in its evidential context** (conditional non-independence).\n\n**Why this is the single most important piece for Deliberus.** Deliberus currently has one\nevidential relation (SUPPORTS) whose contributions add. Schum shows that the *same* surface pattern\n— two items pointing at one claim — can be (i) corroborative-of-the-same-event and therefore\n*redundant* (second item worth little if the first source is credible), (ii) convergent about\ndifferent events and possibly *synergistic* (worth more together), (iii) cumulatively redundant,\nor (iv) collusive and worth nothing extra. Additive aggregation gets (i), (iii) and (iv) wrong in\nthe same direction — overcounting — and (ii) wrong in the other. The already-drafted Deliberus\ndouble-counting taxonomy (same claim twice; one evidence node by many paths; parallel edges;\npartial overlap; shared source) is the **structural** half of this; Schum supplies the\n**semantic** half: *same event vs different events*, and *redundant vs synergistic*.\n\n---\n\n## 3. Intelligence analysis\n\n### 3.1 The NATO Admiralty code (source reliability × information credibility)\n\nOrigin: British Admiralty Naval Intelligence Division, WWII; codified in NATO STANAG 2511 and\nAJP-2.1 (current NATO 2016 edition) (Kelly, Budescu, Dhami & Mandel 2025, verified). The scales\n**[recalled wording; stable across editions with minor variation]**:\n\n| Source reliability | | Information credibility | |\n|---|---|---|---|\n| A | Completely reliable | 1 | Confirmed by other sources |\n| B | Usually reliable | 2 | Probably true |\n| C | Fairly reliable | 3 | Possibly true |\n| D | Not usually reliable | 4 | Doubtful |\n| E | Unreliable | 5 | Improbable |\n| F | Reliability cannot be judged | 6 | Truth cannot be judged |\n\n\"A\" is \"a tried and trusted source which can be depended upon with confidence\". Credibility \"1\"\nis information that \"originates from another source than the already existing information on the\nsame subject\"; \"5\" \"positively contradicts previously reported information\" or \"conflicts with the\nestablished pattern of an intelligence target in a marked degree\"; \"6\" is completely novel\ninformation that \"provides no basis for comparison\" (STANAG 2511, 2003, quoted in Kelly et al.\n2025, verified). The system \"requires the evaluation of source reliability and information\ncredibility **independently** from one another\".\n\n**Why two axes.** A reliable source can report something false (was deceived, was out of position\nthis time); an unreliable source can report something true. Collapsing the two into one grade hides\nexactly the case that matters most for deception detection: *good source, implausible report*, or\n*bad source, independently confirmed report*.\n\n**Known problems** (Irwin & Mandel 2019, *Intelligence and National Security* 34(4):503–525;\nBlock 2021 summary; verified):\n- **The axes are not used independently.** Baker, McKendry & Mace (1968) found in ~1,400 US Army\n  field reports that **87% of ratings fell on the diagonal** (A1, B2, C3 …).\n- **Credibility is multidimensional** (consistency with other reports + plausibility) while\n  reliability is unidimensional.\n- **\"Confirmed by other sources\" does not say how many or how independent.** Irwin & Mandel ask\n  \"how many independent sources must provide confirmation … And should relationships of affinity,\n  hostility or independence between the sources be considered?\"; and an emphasis on consistency\n  with existing reporting \"may encourage confirmation bias\".\n- **No numeric anchors**, national variants, treating reliability as context-free, and failing to\n  distinguish **primary from relaying sources** or humans from sensors.\n- Experiments: Samet (1975) found credibility weighted more heavily than reliability on likelihood\n  judgments; Mandel et al. (2023) and Kelly et al. (2025) found roughly equal weighting, with\n  intra-individual reliability best when the two attributes are moderately consistent and worst\n  when maximally inconsistent (A5, E1) — i.e. **people find it hardest to judge precisely the\n  cases the two-axis design exists for.**\n\nIrwin & Mandel's proposed replacement: numeric probability estimates of information accuracy,\ncollector–analyst collaboration, and periodic re-evaluation as new information arrives.\n\n### 3.2 ICD 203 analytic standards (ODNI)\n\nICD 203, *Analytic Standards*, signed 2 January 2015 (superseding 21 June 2007), with technical\namendments 2022 (text read in full this session). Five standards: analysis must be **Objective**;\n**Independent of political consideration**; **Timely**; **Based on all available sources**; and\n**Implements and exhibits the nine Analytic Tradecraft Standards**:\n\n1. \"Properly describes quality and credibility of underlying sources, data, and methodologies\" —\n   factors include \"accuracy and completeness, possible denial and deception, age and continued\n   currency of information, and technical elements of collection as well as source access,\n   validation, motivation, possible bias, or expertise\"; source summary statements should \"explain\n   which sources are most important to key analytic judgments\" (per ICD 206).\n2. \"Properly expresses and explains uncertainties associated with major analytic judgments\" —\n   separating **likelihood** of an event from the analyst's **confidence** in the basis; noting\n   \"causes of uncertainty (e.g., type, currency, and amount of information, knowledge gaps, and the\n   nature of the issue)\" and \"to what degree and how a judgment depends on assumptions\".\n3. \"Properly distinguishes between underlying intelligence information and analysts' assumptions\n   and judgments\" — \"Assumptions are defined as suppositions used to frame or support an argument\n   … Judgments are defined as conclusions based on underlying intelligence information, analysis,\n   and assumptions. Products should state assumptions explicitly when they serve as the **linchpin**\n   of an argument or when they bridge key information gaps … explain the implications for judgments\n   if assumptions prove to be incorrect.\"\n4. \"Incorporates analysis of alternatives\" — \"identify and assess plausible alternative\n   hypotheses\" and \"indicators that, if detected, would affect the likelihood of identified\n   alternatives.\"\n5. Demonstrates customer relevance and addresses implications.\n6. \"Uses clear and logical argumentation\" — \"acknowledge significant supporting and contrary\n   information\".\n7. \"Explains change to or consistency of analytic judgments\" — state how judgments are consistent\n   with or changed from prior analysis; \"Significant differences in analytic judgment, such as\n   between two IC analytic elements, should be fully considered and brought to the attention of\n   customers.\"\n8. \"Makes accurate judgments and assessments\" — \"should not avoid difficult judgments in order to\n   minimize the risk of being wrong\".\n9. \"Incorporates effective visual information where appropriate.\"\n\n**The likelihood table** (verbatim, ICD 203 D.6.e.(2)(a)): two parallel rows of terms —\n*almost no chance / very unlikely / unlikely / roughly even chance / likely / very likely / almost\ncertain(ly)* and *remote / highly improbable / improbable (improbably) / roughly even odds /\nprobable (probably) / highly probable / nearly certain* — mapped to **01–05%, 05–20%, 20–45%,\n45–55%, 55–80%, 80–95%, 95–99%**. \"Analysts are strongly encouraged not to mix terms from different\nrows.\" And, crucially: products using a confidence level \"**must not combine a confidence level and\na degree of likelihood** … in the same sentence.\"\n\n**Transferable**: ICD 203's likelihood/confidence split is the same move as Deliberus's\ndirection/weight split, and its \"linchpin assumption\" requirement is a demand to surface the\nunsupported premise a judgment most depends on — which Deliberus's hinge score computes.\n\n### 3.3 Kent's words of estimative probability\n\nSherman Kent, \"Words of Estimative Probability\" (*Studies in Intelligence*, 1964; reprinted in\nSteury 1994). Prompted by NIE 29-51 (1951) saying a Soviet attack on Yugoslavia was a \"serious\npossibility\", which readers interpreted from roughly 20% to 80% **[recalled]**. Kent's scale\n**[recalled]**: certain 100%; almost certain 93% (±~6); probable 75% (±~12); chances about even\n50% (±~10); probably not 30% (±~10); almost certainly not 7% (±~5); impossible 0. Heuer (1999,\nverified) reports an experiment with 23 NATO officers: \"there was broad consensus about the\nmeaning of 'better than even,'\" but \"a wide disparity in interpretation of other probability\nexpressions\", and warns an ambiguous report \"may have no impact on the reader if it is couched in\nsuch ambiguous language that the reader can easily interpret it as consistent with his or her own\npreconceptions.\" Heuer's own anecdote: two analysts argued for months over whether a source was\n\"probably\" bona fide or \"probably\" under hostile control; both turned out to mean ~51%.\nLater work (Friedman, Baker, Mellers, Tetlock & Zeckhauser 2018, *ISQ*) found that rounding\nforecasters' numeric estimates to coarser verbal bins measurably degrades accuracy\n**[recalled]**.\n\n### 3.4 Heuer's Analysis of Competing Hypotheses (ACH)\n\nRichards J. Heuer Jr., *Psychology of Intelligence Analysis* (CIA CSI, 1999), ch. 8 (verified).\nThe eight steps, verbatim:\n\n1. \"Identify the possible hypotheses to be considered. Use a group of analysts with different\n   perspectives to brainstorm the possibilities.\"\n2. \"Make a list of significant evidence and arguments for and against each hypothesis.\"\n3. \"Prepare a matrix with hypotheses across the top and evidence down the side. Analyze the\n   'diagnosticity' of the evidence and arguments — that is, identify which items are most helpful in\n   judging the relative likelihood of the hypotheses.\"\n4. \"Refine the matrix. Reconsider the hypotheses and delete evidence and arguments that have no\n   diagnostic value.\"\n5. \"Draw tentative conclusions about the relative likelihood of each hypothesis. Proceed by trying\n   to disprove the hypotheses rather than prove them.\"\n6. \"Analyze how sensitive your conclusion is to a few critical items of evidence. Consider the\n   consequences for your analysis if that evidence were wrong, misleading, or subject to a\n   different interpretation.\"\n7. \"Report conclusions. Discuss the relative likelihood of all the hypotheses, not just the most\n   likely one.\"\n8. \"Identify milestones for future observation that may indicate events are taking a different\n   course than expected.\"\n\n**Diagnosticity** (verbatim): \"A high-temperature reading may have great value in telling a doctor\nthat a patient is sick, but relatively little value in determining which illness a person is\nsuffering from … Evidence is diagnostic when it influences your judgment on the relative likelihood\nof the various hypotheses … If an item of evidence seems consistent with all the hypotheses, it may\nhave no diagnostic value. A common experience is to discover that most of the evidence supporting\nwhat you believe is the most likely hypothesis really is not very helpful, because that same\nevidence is also consistent with other hypotheses.\" And: \"No matter how much information is\nconsistent with a given hypothesis, one cannot prove that hypothesis is true … a single item of\nevidence that is inconsistent with a hypothesis may be sufficient grounds for rejecting that\nhypothesis.\"\n\n**Absence of evidence** (verbatim): \"Note the absence of evidence as well as its presence … the\nsteps the adversary has not taken to ready his forces for attack may be more significant than the\nobservable steps that have been taken … the dog did not bark in the night.\" And ask why it is\nabsent: \"Is it because it has not happened, it is not normally observable, it is being concealed\nfrom you, or because you or the intelligence collectors have not looked for it?\" Heuer also\nreports that \"out of sight, out of mind\" describes how people actually treat gaps (Fischhoff,\nSlovic & Lichtenstein's car-won't-start fault-tree study).\n\n**Oversensitivity to consistency** (verbatim): \"Information may be consistent only because it is\nhighly correlated or redundant, in which case many related reports may be no more informative than\na single report. Or it may be consistent only because information is drawn from a very small sample\nor a biased sample … there is an illusion of validity caused by the consistency of the information.\"\n\n**ACH is contested.** Dhami, Belton & Mandel (2019, *Applied Cognitive Psychology* 33(6):1080–1090;\nverified): 50 analysts randomly assigned; ACH-trained analysts \"did not follow all of the steps\";\n\"mixed evidence for ACH's ability to reduce confirmation bias\"; ACH \"may increase judgement\ninconsistency and error\"; ACH users produced more **tied ranks** (80% vs 19%), partly because ACH\n\"encourages analysts to reduce probabilistic (continuous) data … to a 5-point ordinal scale\"; no\naccuracy advantage (36% vs 33% chose the correct hypothesis). Analysts in both groups preferred\nadding up *both* consistent and inconsistent evidence rather than ACH's disconfirmation-only rule.\nA separate caution: the **zero-sum fallacy** (Pilditch, Fenton & Lagnado 2019, *Psychological\nScience*): people wrongly assume evidence equally predicted by two hypotheses supports neither,\nwhich holds only when the hypotheses are mutually exclusive and exhaustive — **diagnosticity is\nrelative to the hypothesis set**, and a non-exhaustive set makes \"non-diagnostic\" misleading.\n\n**Settled vs contested**: the *concept* of diagnosticity (= likelihood ratio between rival\nhypotheses) is uncontested; ACH *as a procedure* has weak empirical support.\n\n### 3.5 Circular reporting and false corroboration\n\nThe canonical case: Iraq WMD. The Robb–Silberman Commission (31 March 2005; verified):\n- Curveball, a single source never met by any US officer, supplied \"approximately 100 detailed\n  reports\" on mobile BW facilities between 2000 and 2001; the second and fourth sources \"each\n  provided a single report\".\n- \"The NIE … attributed this judgment to multiple sources … The presentation of the material as\n  attributable to 'multiple sensitive sources,' however, gave the impression that the support for\n  the BW assessments was more broadly based than was in fact the case.\"\n- The daily briefs, \"with their attention-grabbing headlines and drumbeat of repetition, **left an\n  impression of many corroborating reports where in fact there were very few sources**.\"\n- A source the DIA had flagged as a **fabricator** in May 2002 was nevertheless used again in July\n  and in the October 2002 NIE, \"because … Defense HUMINT had failed to recall the reporting\" — a\n  retraction that did not propagate.\n- Defense HUMINT regarded itself as \"just a 'conduit'\" with no responsibility for validating the\n  source — **a relay that does not own credibility**.\n\nFour distinct failures, each with a graph analogue: (1) **one source, many reports** counted as\nvolume; (2) **\"multiple sources\" asserted without independence**; (3) **retraction not\npropagated** to dependent products; (4) **relays that pass reports without owning their\ncredibility**. The term \"circular reporting\" usually denotes the special case where a report\nreturns to its origin disguised as confirmation (e.g. a liaison service's report citing press\ncoverage that itself came from the first service) **[recalled general usage]**. A September 2026\npreprint states the machine version exactly: \"another agent is not another observation\"; with\nfixed evidence and report multiplicity rising 1 → 32, naive aggregation collapsed posterior\ncoverage from 0.940 to 0.263 (Bara 2026, arXiv:2609.01873, abstract verified).\n\n### 3.6 Source reliability vs report credibility: the principle\n\nAcross Admiralty, ICD 203/206 and Schum the principle is the same: **evaluate the source and the\nspecific report separately, record both, and do not let one silently stand in for the other.**\nSchum sharpens it into four attributes of the *report* (veracity, objectivity, sensitivity, plus\ncompetence/access) that can vary report by report for the *same* source, and Irwin & Mandel add\nthat source reliability is itself context-dependent (reliable on topic X, not Y).\n\n---\n\n## 4. History and source criticism\n\n### 4.1 Primary, secondary, tertiary\n\n**[recalled; consensus]** A **primary** source is produced by a participant or witness, or\ncontemporaneously with the events (letters, minutes, photographs, the court record). A\n**secondary** source analyses or interprets primary sources (a historian's monograph). A\n**tertiary** source compiles secondary ones (encyclopedias, textbooks). The status is relative to\nthe question: a 1950s history of the French Revolution is secondary on 1789 and primary on 1950s\nhistoriography.\n\n### 4.2 External and internal criticism\n\nThe classic manuals are Ernst Bernheim, *Lehrbuch der historischen Methode* (1889) and Charles-V.\nLanglois & Charles Seignobos, *Introduction aux études historiques* (1898; English 1898). Langlois\n& Seignobos (verified, Gutenberg text) define a \"twofold division of Historical Criticism:\n**External**, investigating the transmission and origin of documents and the statements in them;\n**Internal**, dealing with the content of the statements and their probability.\" Internal\ncriticism examines the good faith and accuracy of each statement; their list of **reasons for\ndoubting good faith**: \"(1) the author's interest; (2) the force of circumstances, official\nreports; (3) sympathy and antipathy; (4) vanity; (5) deference to public opinion; (6) literary\ndistortion\" (verified).\n\nGarraghan & Delanglez, *A Guide to Historical Method* (1946), six inquiries (verified via secondary\nsummary): **when** (date), **where** (localization), **by whom** (authorship), **from what\npre-existing material** (analysis), **in what original form** (integrity) — together external\ncriticism — and **what is the evidential value of its contents** (credibility) — internal\ncriticism. **This is exactly the source/report split**: questions 1–5 are about the document as an\nartefact (≈ Schum's authenticity + chain of custody), question 6 about what it says\n(≈ credibility of the specific report).\n\nThe German *Quellenkritik* questions taught in schools and journalism (who, when, where, why, for\nwhom, how close to the event, dependence on other sources) are the same list **[recalled; common\nScandinavian formulation: äkthet, tid, beroende, tendens — authenticity, time/closeness,\ndependence, bias]**. The Swedish four-criteria version is worth noting because it puts\n**dependence** (is this source independent of the others?) on an equal footing with bias.\n\n### 4.3 Corroboration and independence\n\nLanglois & Seignobos's chapter headings (verified) are almost a Schum outline sixty years early:\n\"The statements of authors, taken singly, do not rise above probability — The only sure results of\ncriticism are negative — To establish facts it is necessary to compare different statements —\nContradictions between statements, real and apparent — **Agreement of statements — Necessity of\nproving them to be independent — Perfect agreement not so conclusive as occasional coincidence**\n— Cases where different observations of the same fact are not independent — … **Different facts,\neach imperfectly proved, corroborate each other when they harmonise**.\" Two ideas here are not in\nSchum's vocabulary: **perfect agreement is suspicious** (verbatim agreement suggests copying), and\nthe last heading is convergence. The standard historians' maxims derived from Bernheim and L&S\ninclude \"majority does not rule\" and \"if two independently created sources agree on a matter, the\nreliability of each is measurably enhanced\" (as summarised in the historical-method literature;\nverified as secondary summary only).\n\n**Stemmatics** (Lachmann's method) is the most developed independence-detection technique in the\nhumanities: shared *errors* reveal copying, so witnesses are grouped into families and a family\ncounts once **[recalled]**. L&S discuss exactly this (\"Suppose A is the common source: perhaps B\nand C copied it independently\").\n\n### 4.4 The argument from silence\n\nLanglois & Seignobos (verified) state the implicit premise — \"If an alleged event really had\noccurred, there would be some document in existence in which it would be referred to\" — and\nrestrict the argument to cases where two conditions hold: \"(1) It is necessary not only that there\nshould be now no documents in existence which mention the fact in question, but that there should\nnever have been any. If the documents are lost we can conclude nothing\"; \"(2) The fact must have\nbeen of such a kind that it could not fail to be observed and recorded.\" They name the typical\nabuse: silence about \"the secret acts of governments and the complaints of the lower classes …\ngreatly impresses unreflecting historians; it is the origin of the widespread sophism of the 'good\nold times.'\" **This is a critical-question set for Walton's argument from ignorance, with an\nexplicit political-bias warning attached**: a record that systematically omits a class of events\ncannot support negative conclusions about that class.\n\n### 4.5 Testimony and the evidential paradigm\n\n- Historians do not exclude hearsay; Gottschalk (verified via secondary quote) asks of any\n  secondary witness: on whose primary testimony does he rely, did he report it accurately as a\n  whole, and if not in which details — then tests that primary testimony as if it were before him.\n- Marc Bloch (*The Historian's Craft*, 1949) prized \"witnesses in spite of themselves\" — evidence\n  not intended to inform posterity (a merchant's ledger over a chronicle) — because unintended\n  evidence carries no motive to mislead **[recalled]**. This is Schum's *veracity* attribute made\n  structural: intended testimony has a veracity question, unintended traces largely do not.\n- **Carlo Ginzburg**, \"Spie. Radici di un paradigma indiziario\" (1979; English \"Clues: Roots of an\n  Evidential Paradigm\" in *Clues, Myths, and the Historical Method*, 1989) **[recalled]**: a\n  **conjectural** or **evidential paradigm** shared by Morelli's attribution of paintings from ear\n  lobes and fingernails, Freud's reading of slips, Sherlock Holmes's deductions, and medical\n  semiotics — knowledge from small, involuntary, marginal traces, individualising rather than\n  generalising, abductive, and irreducibly qualitative. Ginzburg's point for us: an entire\n  respectable tradition of evidence rests on *signs* read by trained judgement that cannot be\n  fully formalised — which is where the Walton *argument from sign* lives and why its critical\n  questions (is the sign reliable? are there other causes?) are the whole of its rigour.\n\n---\n\n## 5. Journalism and fact-checking\n\n### 5.1 Verification discipline\n\nKovach & Rosenstiel, *The Elements of Journalism* (2001; 4th ed. 2021): \"the essence of journalism\nis a discipline of verification\", distinguishing a **journalism of verification** from journalisms\nof assertion and affirmation **[recalled]**. Their \"intellectual principles of a science of\nreporting\" **[recalled]**: never add anything that was not there; never deceive the audience; be\nas transparent as possible about your methods and motives; rely on your own original reporting;\nexercise humility. **Transparency** is their answer to verification's limits: tell the reader how\nyou know and what you do not know.\n\n**Two-source rule**: a newsroom convention (made famous by the *Washington Post* during Watergate)\nthat a sensitive fact from anonymous sources needs two independent confirmations **[recalled; not a\nformal code]**. Its weakness is Schum's: two sources are two only if independent. The IFCN code\n(§5.3) embeds a version: key elements are checked \"against more than one named source\"\n(verified fragment).\n\n### 5.2 OSINT verification (Bellingcat and others)\n\n**[recalled; practice consensus]** Open-source verification separates questions about a piece of\nuser-generated media into **provenance** (is this the original? who first posted it?), **source**\n(who captured it?), **date/time** (**chronolocation** — shadows and sun position via tools like\nSunCalc, weather records, metadata), **location** (**geolocation** — matching landmarks,\nskylines, road markings against satellite and street-level imagery), and **motivation** — the\nFirst Draft \"five pillars\" checklist **[recalled]**. Bellingcat's distinguishing norm is\n**publishing the work**: each geolocation is shown with the matched features so any reader can\nredo it — verification as *replicable*, not as authority. Reverse image search to find earlier\nappearances is the standard test for **recycled media** (an old image presented as new), which is\na provenance failure, not a content failure. **This maps to chain-of-custody-as-claims**: \"this\nvideo was filmed at coordinates X on date Y\" is a claim with its own evidence, separable from what\nthe video shows.\n\n### 5.3 IFCN Code of Principles and ClaimReview\n\nThe **IFCN Code of Principles** (Poynter, 2016; current version June 2025 with 31 criteria;\nverified): five commitments — **non-partisanship and fairness** (\"the same standard for every fact\ncheck … let the evidence dictate the conclusions\"); **standards and transparency of sources**\n(\"provide all sources in enough detail that readers can replicate their work\"; use \"the best\navailable primary, not secondary, sources of evidence wherever suitable primary sources are\navailable\", and explain when a secondary source is used; disclose relevant interests of sources);\n**transparency of funding and organization** (disclose funders ≥5% of revenue); **standards and\ntransparency of methodology**; and an **open and honest corrections policy** (\"seeking so far as\npossible to ensure that readers see the corrected version\").\n\n**ClaimReview** (schema.org, developed with Google and Duke Reporters' Lab, ~2016–17)\n**[recalled; check field names before implementing]**: a `ClaimReview` has `claimReviewed` (the\nclaim text), `itemReviewed` (a `Claim` with `author`/`appearance`/`datePublished`), `author` (the\nfact-checker), `datePublished`, `url`, and `reviewRating` (a `Rating` with `ratingValue`,\n`bestRating`, `worstRating`, and `alternateName` — the textual verdict like \"Mostly False\"). It is\na *verdict* format: one claim, one rating. Deliberus's use would be *import* (a fact-check as\nevidence with provenance) rather than *emit* (Deliberus does not issue verdicts).\n\n### 5.4 Rating scales and their criticism\n\n- **Uscinski & Butler (2013)**, \"The Epistemology of Fact Checking\", *Critical Review*\n  25(2):162–180 **[recalled]**: fact-checkers select claims non-randomly, rate predictions and\n  causal claims as if they were facts, and aggregate verdicts into misleading \"who lies more\"\n  statistics.\n- **Lim (2018)**, *Research & Politics* 5(3) (verified): of 1,178 PolitiFact and 325 WaPo checks\n  of 2016 candidates, only 77 statements overlapped; agreement was good on \"outright falsehoods or\n  obvious truths\" but \"much lower (14 agreements out of 23 cases) for statements in the more\n  ambiguous scoring range\". Most important finding for Deliberus: the discrepancies \"often stemmed\n  from differences in fact-checkers' subjective judgment of the *significance* of inaccuracies …\n  rather than their disagreement on whether the statement itself was true\". On Jeb Bush's \"Florida\n  led the nation in job creation\", both gave **identical rationales** (raw totals, omitted year,\n  low-wage jobs); WaPo gave 4 Pinocchios, PolitiFact \"Half True\".\n- **Lee et al. (2023)**, *HKS Misinformation Review* 4(5) (verified): Snopes vs PolitiFact, 749\n  matching claims; 69.6% identical ratings, but after adjusting for scale granularity, focus,\n  subtly different claims and timing, **only one** genuine conflict. **Cross-checking** study\n  (PLOS ONE 2023, verified): WaPo vs PolitiFact on Trump statements — 22.6% selection\n  disagreement, moderate agreement on deceptiveness *scaling*, \"nearly complete agreement\" on\n  bottom-line veracity.\n\n**Synthesis**: professional checkers **agree on the facts and disagree on the scale**. The\ndisagreement lives in the step that maps a *set of findings* onto a *single ordinal verdict* —\ni.e. in weighting the significance of each finding. A graph that keeps the findings as separate\nclaims and does not collapse them into one rating preserves the part everyone agrees on and\nexposes the part they do not.\n\n### 5.5 Lateral reading and SIFT\n\n**Wineburg & McGrew (2019)**, \"Lateral Reading and the Nature of Expertise\", *Teachers College\nRecord* 121(11) (verified): 10 PhD historians, 10 professional fact-checkers, 25 Stanford\nundergraduates. Historians and students \"often fell victim to easily manipulated features of\nwebsites, such as official-looking logos and domain names. They read **vertically**, staying\nwithin a website to evaluate its reliability. In contrast, fact checkers read **laterally**,\nleaving a site after a quick scan and opening up new browser tabs … Compared to the other groups,\nfact checkers arrived at more warranted conclusions in a fraction of the time.\" Their three\npractices: **taking bearings**, **lateral reading**, **click restraint**. In the minimumwage.com\ntask fact-checkers left the page in a mean of 37 s and identified it as a front for a PR firm.\n\n**SIFT** (Mike Caulfield, 2019) **[recalled]**: **S**top; **I**nvestigate the source; **F**ind\nbetter coverage; **T**race claims, quotes and media to the original context.\n\n**The epistemology**: you assess a source's reliability **from outside the source**, by what\nindependent others say about it — never from its self-presentation. For a claim graph: a source's\nreliability should be a claim *about* the source, supported by evidence from *other* nodes, not a\nproperty the source's own content can raise.\n\n---\n\n## 6. Forensic science\n\n### 6.1 NAS 2009\n\nNational Research Council, *Strengthening Forensic Science in the United States: A Path Forward*\n(2009). Its most-cited sentence (verified via multiple secondary sources): \"With the exception of\nnuclear DNA analysis, … no forensic method has been rigorously shown to have the capacity to\nconsistently, and with a high degree of certainty, demonstrate a connection between evidence and a\nspecific individual or source.\" It cited \"the notable dearth of peer-reviewed, published studies\nestablishing the scientific bases and validity of many forensic methods\" and recommended an\nindependent federal entity (never created in the proposed form) **[recalled]**.\n\n### 6.2 PCAST 2016\n\nPresident's Council of Advisors on Science and Technology, *Forensic Science in Criminal Courts:\nEnsuring Scientific Validity of Feature-Comparison Methods* (Sept 2016; text read this session).\n\n- **Foundational validity**: a method \"shown, based on empirical studies, to be repeatable,\n  reproducible, and accurate, at levels that have been measured and are appropriate to the intended\n  application … It is the scientific concept we mean to correspond to the legal requirement, in\n  Rule 702(c)\". **Validity as applied**: \"the method has been reliably applied in practice …\n  correspond[ing] to … Rule 702(d).\"\n- For **objective** methods, validity can be established step by step; for **subjective**\n  methods \"the method must be evaluated as if it were a 'black box' in the examiner's head\" through\n  **black-box studies** \"in which many examiners render decisions\" on samples of known ground\n  truth; \"the foundational validity of a subjective method can only be established through\n  multiple, appropriately designed black-box studies\"; and \"neither experience nor professional\n  practices can substitute for foundational validity\" (section title).\n- Findings: **latent fingerprints** — \"foundationally valid … albeit with a false positive rate that\n  is substantial …: could be as high as **1 error in 306** cases based on the FBI study and **1 error\n  in 18** cases based on a study by another crime laboratory\"; with a confirmation-bias warning that\n  examiners \"often alter the features that they initially mark in a latent print based on\n  comparison with an apparently matching exemplar\". **Firearms** — one appropriately designed\n  black-box study (Ames Laboratory): false-positive rate \"estimated at **1 in 66**, with a\n  confidence bound indicating that the rate could be as high as **1 in 46**\"; one study is not\n  enough, so \"falls short of the scientific criteria for foundational validity\". **Bitemarks** —\n  \"far from meeting the scientific standards\"; observed false positives \"typically above ten\n  percent\"; examiners \"cannot even consistently agree on whether an injury is a human bitemark\".\n  **Footwear**: \"no appropriate black-box studies\". **Single-source and simple-mixture DNA**:\n  foundationally valid.\n- The DOJ and many forensic bodies disputed PCAST's insistence on black-box studies as the *only*\n  route for subjective methods **[recalled; contested]**.\n\n### 6.3 Likelihood ratios and the ENFSI verbal scale\n\nThe **likelihood ratio** LR = P(E | Hp) / P(E | Hd) measures how much more probable the findings\nare under one proposition than under a stated alternative; it does **not** state the probability of\neither proposition. **ENFSI Guideline for Evaluative Reporting in Forensic Science** (2015;\nWillis et al.) requires evaluative reports to satisfy **balance, logic, robustness and\ntransparency**, to address at least one pair of competing propositions, and to report the value of\nfindings as an LR **[recalled for the four principles]**. Its verbal scale (verified via multiple\nsources):\n\n| LR range | Verbal equivalent (\"the findings provide … support for Hp rather than Hd\") |\n|---|---|\n| 1 – 10 | weak |\n| 10 – 100 | moderate |\n| 100 – 1,000 | moderately strong |\n| 1,000 – 10,000 | strong |\n| 10,000 – 1,000,000 | very strong |\n| > 1,000,000 | extremely strong |\n\n(An LR of 1 is \"neutral\"; below 1 the same scale runs for Hd.)\n\n**The weak evidence effect** (Martire, Kemp, Watkins, Sayle & Newell 2013, *Law & Human Behavior*\n37(3):197–207; Martire et al. 2014, *Forensic Sci. Int.* 240:61–68; verified abstracts): with\n~900 participants, belief change was \"considerably smaller than Bayesian calculations would\npredict\", and \"participants presented with evidence **weakly** supporting guilt tended to\n'**invert**' the evidence, thereby counterintuitively reducing their belief in the guilt of the\naccused.\" The effect \"was most apparent in the verbal presentation conditions\" and numerical\nexpressions were \"most resistant to the weak evidence effect\". **Direct implication for badges**:\na label meaning \"weak support for X\" can be read as evidence *against* X.\n\n### 6.4 Prosecutor's fallacy and defence fallacy\n\nThompson & Schumann (1987), \"Interpretation of statistical evidence in criminal trials: the\nprosecutor's fallacy and the defense attorney's fallacy\", *Law & Human Behavior* 11:167–187\n**[recalled]**.\n\n- **Prosecutor's fallacy** (the transposed conditional): treating P(evidence | innocent) as\n  P(innocent | evidence) — \"the chance of a random match is 1 in a million, so the chance he is\n  innocent is 1 in a million\". The *Sally Clark* case (UK, 1999; overturned 2003) combined this with\n  an independence error (squaring a SIDS rate to get \"1 in 73 million\"); the Royal Statistical\n  Society issued a public statement in 2001 **[recalled]**. *R v Adams* (1996–97) is the leading\n  English case on presenting Bayesian reasoning to juries **[recalled]**.\n- **Defence fallacy**: \"in a city of 10 million, 10 people would match, so the match makes it only\n  1 in 10 that he is the source — the evidence is nearly irrelevant.\" Wrong because the match\n  moved the probability from 1 in 10 million to 1 in 10: the LR is huge even if the posterior is\n  modest.\n\nBoth errors come from **reporting a posterior where only a likelihood was established**, or\nreading an LR as if it were a posterior. ENFSI's insistence on \"support for Hp *rather than* Hd\"\nwording exists to block them.\n\n### 6.5 Contextual bias and sequential unmasking\n\nItiel Dror's programme: examiners exposed to task-irrelevant context (a confession, the detective's\nbelief) reach different conclusions on the same evidence — including the famous study where\nfingerprint examiners contradicted their own earlier identifications when given a misleading\ncontext (Dror, Charlton & Péron 2006) **[recalled]**. **Linear Sequential Unmasking** (Dror,\nThompson, Meissner, Kornfield, Krane, Saks & Risinger 2015, *J. Forensic Sci.*; verified\ncitation): examiners \"first analyze the trace evidence in isolation from the reference material\"\nand document it, with \"a balanced restriction on the changes that are permitted after exposure to\nthe reference material\". **LSU-Expanded** (Dror & Kukucka 2021, *Forensic Sci. Int.: Synergy*)\ngeneralises the idea to all forensic decisions: sequence information by relevance and objectivity,\nrecord what was seen when, and justify any revision.\n\n**Transfer**: the order in which a contributor sees material changes their judgment; a claim graph\nthat shows a CQ answerer the target claim's current badge before they answer invites exactly the\ncircularity PCAST found in fingerprint marking.\n\n### 6.6 Hierarchy of propositions\n\nCook, Evett, Jackson, Jones & Lambert (1998), \"A hierarchy of propositions: deciding which level to\naddress in casework\", *Science & Justice* 38(4):231–239 (citation verified): propositions can be\nframed at **source level** (Level I: \"this stain came from Mr X\"), **activity level** (Level II:\n\"Mr X kicked the victim\"), or **offence level** (Level III: \"Mr X assaulted the victim\"); Evett et\nal. (2000, \"More on the hierarchy of propositions\") added **sub-source** (DNA came from X, without\nsaying which body fluid) and distinguished *explanations* from *propositions*. \"In general, the\nhigher the level of the propositions, the greater the assistance that will be given to the court\"\n— but also the more the evaluation depends on case circumstances and assumptions (transfer,\npersistence, background levels) that lie outside the scientist's measurement. Offence level is the\ncourt's job; scientists work at source and activity.\n\n**Transfer**: the same finding has a different LR at each level. \"The DNA matches\" can be strong\nsupport at sub-source level and nearly neutral at activity level (secondary transfer). A claim\ngraph that lets a source-level finding support an offence-level claim directly skips the most\ncontestable inferential steps — exactly the steps Wigmore wanted on the chart.\n\n---\n\n## 7. Most transferable to a claim-graph argumentation platform like Deliberus\n\nEach item names the gap from the briefing it addresses, the professional source, and a concrete\nshape. None requires Deliberus to issue verdicts; all keep the \"machine proposes, contributions\nthat argue for themselves move strength\" rule.\n\n### 7.1 Separate the source from the report — two nodes, not one\n\n*Gap: \"no source reliability separate from claim\".* Every evidential claim that comes from a\nperson, document or instrument should be linked to a **Source** claim-cluster (\"Outlet X\", \"Witness\nY\", \"Sensor Z\") through the existing REPORTS/always-mint machinery. Source reliability becomes a\nset of claims *about the source* (track record, access on this topic, interests), argued like any\nother — and, per lateral reading, **evidence for a source's reliability must come from nodes other\nthan that source's own output**. Report credibility stays on the report. Display both, as NATO\ndoes, and follow Irwin & Mandel: the two should be allowed to disagree (A5, E1), and the graph\nshould flag those cases as the most informative rather than smoothing them.\n\n### 7.2 Schum's credibility attributes as the critical questions of testimony and records\n\n*Gap: witness-testimony and position-to-know CQs are generic; answers are counted.* Replace or\nextend the testimony scheme's CQs with Schum's decomposition, each as a separately answerable\nclaim:\n\n- *Competence*: access (\"was the source in a position to observe this?\"), understanding.\n- *Credibility*: **veracity** (\"does the source believe what it reports?\" — interests, statement\n  against interest, unintended trace), **objectivity** (\"was the belief formed from observation\n  rather than expectation or wish?\"), **observational sensitivity** (\"how good were the conditions\n  and the source's senses/instruments?\").\n- *Mode*: direct observation / second-hand / opinion-based; equivocal or not.\n- For tangible sources: **authenticity**, **accuracy**, **reliability** (repeatability).\n- For experts: the nine CrimPD 7.1.2 factors, especially **where in the range of expert opinion\n  this expert sits** and **whether the expert had complete information**.\n\nKeep Schum's warning as a validation rule: **an answer to a competence CQ must not be allowed to\nraise a credibility CQ** (\"he had good access, so believe him\" is the named non sequitur).\n\n### 7.3 Type the combination: corroboration vs convergence vs redundancy vs synergy\n\n*Gaps: \"no source independence / shared-origin correlation\"; \"35% of decomposition edges are\nevidence misfiled as parts\" [⚠ corrected 2026-09-27: the census label reliably means only *not a part*; read by hand, about 13 of the 31 machine-written ones were support running child to parent (⚠ 2026-10-07: *child* here names the premise by its role; as stored, these links ran premise first, the reverse of a part-of link: `auto-connect-upgrade.md` § 2c), so the 29 written by the automatic linker were retired rather than re-filed (`session39-what-the-graph-stopped-claiming.md` § 3)]; weight does not saturate.* When two or more evidence edges point at one\nclaim, record (proposed by machine, contestable by anyone) **which Schum pattern they form**:\n\n- **same event, different sources → corroborative**: aggregate with **saturation**, where each\n  additional report adds only to the extent prior reports are not credible (Schum's \"springs to\n  life\" rule; Schum & Kelly's result that \"source reliability is more important than the extent of\n  source agreement\"). Deliberus's weight (count of inputs) should become a saturating function of\n  *independent credible* inputs, not raw count.\n- **different events, same conclusion → convergent**: this is where *corroborative* decomposition\n  (summed offsets) belongs, and where **synergy** can be proposed (the two together exclude a rival\n  explanation neither excludes alone).\n- **contradictory** (mutually exclusive events): resolved on **credibility**, never on headcount —\n  the badge must not show 3-vs-1 as a majority.\n- **divergent**: needs its own treatment (both may be true; seek a joint explanation).\n\nThis also gives the \"three studies are witnesses, not parts\" repair its vocabulary: three studies\nof the *same effect* are corroborative (saturating), not necessary parts (min) — they should never\nhave been DECOMPOSES_INTO.\n\n### 7.4 Circular-reporting detection: provenance as graph structure\n\n*Gap: \"no source independence\"; the \"known shared source becomes a node\" principle already\ndrafted.* Implement the WMD Commission's four failures as checks: (1) **count sources, not reports**\n— any summary that says \"N items support this\" must say from how many independent origins;\n(2) **independence is a claim**, not a default — \"these two reports are independent\" can be\nsupported (different access, different chains, different errors) or attacked (shared wire story,\ncopied phrasing — L&S's \"perfect agreement not so conclusive as occasional coincidence\", and\nstemmatics' shared-error test are machine-proposable signals); (3) **retraction propagates** — when a\nsource is shown to fabricate, every claim whose support routes through it is flagged, not silently\nkept; (4) **relays do not launder** — a report that passes through an intermediary keeps its\noriginal source's credibility questions open. The Bara (2026) result — naive aggregation of 32\nagents on fixed evidence dropping posterior coverage from 0.94 to 0.26 — is the machine-side\nreason this matters now, since LLM passes are themselves relays.\n\n### 7.5 Weigh CQ answers by diagnosticity, not by count\n\n*Gaps: \"answers to CQs are COUNTED not weighed\"; \"no likelihood ratio or diagnosticity\"; three\none-word answers = three RCTs.* The unifying quantity across law (FRE 401's \"more or less probable\"),\nforensics (LR), intelligence (diagnosticity) and Schum (inferential force) is **how much more\nexpected the evidence is under the claim than under its best rival**. A practical, non-numeric\nfirst step that fits \"propose, never assert\":\n\n- Every evidential edge gets a **rival**: \"what else would explain this evidence?\" (a CQ that\n  *every* evidential scheme should carry). If the rival explains it as well, the edge's force is\n  near zero regardless of how many CQs are answered — Heuer's fever thermometer.\n- An answer to a CQ moves the edge in proportion to how much it **discriminates** (it rules out or\n  makes implausible a rival explanation), which can be expressed on an ordinal LR-band scale\n  (ENFSI's six bands) chosen by the contributor *and argued for*, rather than a fixed +0.15.\n- Preserve Pilditch et al.'s caveat: diagnosticity is relative to the rival set; if rivals are not\n  exhaustive, say so (the platform's \"this is open\" status).\n- Keep Dhami et al.'s caution: forcing continuous judgments onto a coarse ordinal scale produced\n  ties and inconsistency in ACH — use enough bands, and record reasons.\n\n### 7.6 Proof standards chosen by the question, not one global badge\n\n*Gap: \"no standard-of-proof concept\".* Adopt Carneades' idea: **each claim (or each question a\nreader brings) carries a proof standard**, and \"meets its standard\" is computed per claim. Examples:\nscintilla (is there any defensible supporting argument? — useful for \"is this a live\npossibility?\"), best argument / preponderance (civil-style), dialectical validity (no defensible\ncounterargument survives — BRD-like), plus ICD 203's likelihood/confidence split. The same graph\ncan then say \"meets preponderance, fails BRD\" — and, as Carneades notes, can meet a standard *and\nits complement* at low standards, which is honest rather than contradictory. Two design guards:\n(a) the choice of standard is itself a contribution with reasons (cost of each error), never a\ndefault imposed silently; (b) Keynesian weight stays separate — law's refusal to act on naked\nstatistics is the precedent for \"a high number on low weight meets no standard above scintilla\".\nCarneades' **premise types** (ordinary / assumption / exception) also give Deliberus a principled\nway to type which CQs merely *ask* (burden returns to the proponent) and which must be *backed*\nbefore they bite — resolving the rebutting vs undercutting gap in the same move (Verheij's four CQ\nroles).\n\n### 7.7 Chain of custody and provenance as claims\n\n*Gaps: \"no chain of custody / quote verification / page or timestamp locator\"; \"no edge or answer\ntimestamps\".* Following the best evidence rule, FRE 901, Garraghan's first five questions and OSINT\npractice, the **artefact** and the **content** are different claims:\n\n- \"This quotation appears on page N of document D (version V, retrieved T)\" — a locator claim,\n  verifiable by anyone, attackable by \"the quote is truncated / out of context\" (SIFT's \"trace to the\n  original context\").\n- \"Document D is what it purports to be\" (authenticity) and \"D was not altered between creation and\n  retrieval\" (integrity/custody).\n- \"This image was taken at place P on date T\" (geolocation/chronolocation) — separate from what the\n  image shows.\n- Paraphrase adds a transmission link; the best-evidence principle argues that where the content of\n  a source is at issue, the original wording should be stored and the paraphrase shown as derived.\n- Every edge and answer carries a timestamp and author (ICD 203 standard 7: state how a judgment\n  changed and why).\n\n### 7.8 The hierarchy of propositions for testimony and evidence\n\n*Gap: evidence from one level silently supports claims at another.* Borrow Cook et al.'s hierarchy\nas a *typing* of how far an evidential edge jumps: **sub-source → source → activity → offence** has\nanalogues outside crime — \"the study measured X\" (sub-source), \"X occurred in the sample\" (source),\n\"the intervention caused X\" (activity), \"the policy is good\" (offence/normative). An edge that jumps\nmore than one level should be flagged and decomposed, because the skipped levels carry the\ncontestable assumptions (transfer, generalisation, value premises). This dovetails with the\nhorizon-kind typing: only the lower levels can hand over to the empirical oracle.\n\n### 7.9 Badge wording that cannot commit the prosecutor's fallacy or the weak evidence effect\n\n*Gap: badge semantics.* Three rules from forensic reporting:\n\n1. **Say \"support for X rather than Y\"**, naming the rival, never \"X is probably true\" — a\n   Deliberus badge reports force of the mapped reasoning, which is a likelihood-type quantity, not\n   a posterior; wording that reads as a posterior invites the prosecutor's fallacy.\n2. **Near the neutral band, do not use \"weak support\" wording** — Martire et al. show it is read as\n   evidence *against*. Use \"not yet discriminating\" or show the direction with the weight so the\n   reader sees *why* it is small.\n3. **Do not mix likelihood and confidence in one phrase** (ICD 203): direction and weight are\n   displayed as two things, as Deliberus now does — keep it that way in all copy, including agent\n   surfaces.\n\n### 7.10 Negative evidence and the argument from silence as first-class structure\n\n*Gap: \"no representation of 'we searched and found nothing' as data\".* Represent a **search** as a\nclaim: \"Searcher S looked in corpus C, with method M, on date T, for Z, and found nothing.\" Its force\ndepends on the Langlois & Seignobos conditions, which become the CQs: would Z have been recorded if\nit occurred? Would the record have survived? Did the searcher look where it would be? Is there a\nsystematic reason the record omits this *class* of event (the \"good old times\" sophism — whose\nvoices the archive excludes)? Heuer's four reasons for absence (not happened, not observable,\nconcealed, not looked for) are the answer menu. This also gives Deliberus's \"lack of evidence\"\nscheme more than one CQ, so it can leave the neutral band when properly examined.\n\n### 7.11 Separate findings from verdict scaling (the fact-checkers' lesson)\n\nLim's Jeb Bush case — identical findings, verdicts two notches apart — shows that the **mapping\nfrom findings to a single rating is where disagreement lives**. Deliberus already declines to\nissue the rating; the transferable rule is to make each finding (raw totals used; year omitted;\njobs low-wage) its own claim with its own *significance* claim (\"this omission matters because …\"),\nso the disagreement about significance becomes contributable instead of hidden inside a label. It\nalso supplies an **external validation target**: on claims fact-checked by several organisations,\nDeliberus's decomposition should reproduce their shared findings even where their ratings differ.\n\n### 7.12 Sequencing and blinding for contributors\n\nFrom Dror's LSU and PCAST's fingerprint finding: let a contributor record their answer to a CQ\n**before** seeing the current badge or others' answers where feasible, and record what they had\nseen when they answered. This is cheap, protects the independence that §7.3–7.4 depend on, and\ngives the ratification-tether and \"positional testimony\" work an instrument (answers given blind vs\nanchored).\n\n### 7.13 What not to import\n\n- **A fixed methodology ladder as weight.** Every profession here grades the *report and its\n  chain*, not the *genre*: PCAST grades by measured error rate of the specific method; CrimPD\n  grades the specific expert's data and application; Admiralty grades this source on this report.\n  The inert Deliberus ladder (meta-analysis 0.70 … expert opinion 0.30) should stay a *prior\n  proposal* per claim, overridable by argued CQ answers, never a per-source constant.\n- **Single ordinal verdict scales** (Pinocchios, Admiralty \"1–6\" collapsed with \"A–F\") — the\n  evidence says professionals agree on findings and disagree on scaling.\n- **ACH's disconfirmation-only integration rule** — weakly supported; analysts' preferred\n  both-sides integration was not worse.\n- **Headcount resolution of contradictions** — refuted formally (Schum & Kelly 1973) and\n  historically (L&S: \"majority does not rule\").\n\n### 7.14 Consensus vs contested, for the record\n\n| Settled across the professions | Contested |\n|---|---|\n| Source reliability ≠ report credibility | Whether proof is probabilistic or explanatory (Allen & Pardo vs Bayesians) |\n| The LR / diagnosticity as the measure of an item's force | What explains the naked-statistics intuition |\n| Agreement counts only in proportion to independence | Whether ACH improves analysis (evidence says no) |\n| Absence of evidence is evidence only under stated conditions | Whether verbal scales help or harm lay understanding |\n| Standards of proof vary with the cost of error | Numeric equivalents of BRD / clear and convincing |\n| Validity of method ≠ validity of this application (702(c)/(d); PCAST) | Whether black-box studies are the only route to validity for subjective methods |\n| Hearsay's rationale is untestable testimonial capacities | Whether civil-law systems use one standard or several |\n\n---\n\n## References\n\n**Law and legal theory**\n- Federal Rules of Evidence 401–403, 801–807, 901, 1002 — Legal Information Institute, Cornell.\n  https://www.law.cornell.edu/rules/fre (403, 702, 801, 901 read 2026-09-26).\n- FRE 702, as amended 1 Dec 2023, with Committee Note. https://www.law.cornell.edu/rules/fre/rule_702\n- *Frye v. United States*, 293 F. 1013 (D.C. Cir. 1923). *Daubert v. Merrell Dow Pharmaceuticals*,\n  509 U.S. 579 (1993). *Kumho Tire Co. v. Carmichael*, 526 U.S. 137 (1999). *In re Winship*, 397 U.S.\n  358 (1970). *Addington v. Texas*, 441 U.S. 418 (1979). *Smith v. Rapid Transit, Inc.*, 317 Mass.\n  469 (1945). *Taxquet v. Belgium*, ECtHR (GC) 2010.\n- Criminal Practice Directions 2023 (England & Wales), 7.1.1–7.1.3; as quoted in CPS, \"Expert\n  Evidence\" (updated 20 Nov 2023). https://www.cps.gov.uk/prosecution-guidance/expert-evidence ;\n  earlier CrimPD 2015 19A.5 as quoted in *Brecani v R* [2021] EWCA Crim 731.\n- Allen, R. J. & Pardo, M. S. (2019). Relative plausibility and its critics. *International Journal\n  of Evidence & Proof* 23(1–2):5–59. Working paper:\n  https://scholarship.law.ua.edu/cgi/viewcontent.cgi?article=1005&context=fac_working_papers\n- Bex, F. J. (2011). *Arguments, Stories and Criminal Evidence: A Formal Hybrid Theory*. Springer.\n- Bex, F., van Koppen, P., Prakken, H. & Verheij, B. (2010). A hybrid formal theory of arguments,\n  stories and criminal evidence. *Artificial Intelligence and Law* 18:123–152.\n- Clermont, K. M. & Sherwin, E. (2002). A comparative view of standards of proof. *American Journal\n  of Comparative Law* 50:243–275. [recalled]\n- Cohen, L. J. (1977). *The Probable and the Provable*. Oxford: Clarendon.\n- Enoch, D., Spectre, L. & Fisher, T. (2012). Statistical evidence, sensitivity, and the legal value\n  of knowledge. *Philosophy & Public Affairs* 40(3):197–224. [recalled]\n- Fenton, N., Neil, M. & Lagnado, D. A. (2013). A general structure for legal arguments about\n  evidence using Bayesian networks. *Cognitive Science* 37(1):61–102.\n  https://onlinelibrary.wiley.com/doi/10.1111/cogs.12004\n- Gordon, T. F., Prakken, H. & Walton, D. (2007). The Carneades model of argument and burden of\n  proof. *Artificial Intelligence* 171(10–15):875–896.\n  https://webspace.science.uu.nl/~prakk101/pubs/GordonPrakkenWalton2007a.pdf\n- Gordon, T. F. & Walton, D. (2009). Proof burdens and standards. In I. Rahwan & G. Simari (eds.),\n  *Argumentation in Artificial Intelligence*, Springer, 239–258. [recalled; definitions not\n  re-verified]\n- Laudan, L. (2006). *Truth, Error, and Criminal Law*. Cambridge UP. [recalled]\n- McCauliff, C. M. A. (1982). Burdens of proof: degrees of belief, quanta of evidence, or\n  constitutional guarantees? *Vanderbilt Law Review* 35:1293. [recalled]\n- Morgan, E. M. (1948). Hearsay dangers and the application of the hearsay concept. *Harvard Law\n  Review* 62:177. [recalled]\n- Pennington, N. & Hastie, R. (1986). Evidence evaluation in complex decision making. *JPSP*\n  51:242–258; (1992) Explaining the evidence: tests of the story model for juror decision making.\n  *JPSP* 62:189–206. [recalled]\n- Thomson, J. J. (1986). Liability and individualized evidence. *Law and Contemporary Problems*\n  49(3):199–219. [recalled]\n- Tribe, L. H. (1971). Trial by mathematics. *Harvard Law Review* 84:1329; (1974) Triangulating\n  hearsay. *Harvard Law Review* 87:957. [recalled]\n- van Gijzel, B. & Prakken, H. (2012). Relating Carneades with abstract argumentation via the ASPIC+\n  framework. *Argument & Computation* 3(1):21–47.\n- Wells, G. L. (1992). Naked statistical evidence of liability: is subjective probability enough?\n  *JPSP* 62(5):739–752. [recalled]\n\n**Wigmore, Schum, Anderson & Twining**\n- Wigmore, J. H. (1913). *The Principles of Judicial Proof*; 3rd ed. (1937) *The Science of Judicial\n  Proof*. Boston: Little, Brown.\n- Anderson, T., Schum, D. & Twining, W. (2005). *Analysis of Evidence*, 2nd ed. Cambridge UP.\n  https://repository.law.miami.edu/fac_books/32/\n- Schum, D. A. (1994/2001). *The Evidential Foundations of Probabilistic Reasoning*. Wiley;\n  Northwestern UP reprint.\n  https://nupress.northwestern.edu/9780810118218/the-evidential-foundations-of-probabilistic-reasoning/\n- Schum, D. A. (1999). Inference networks and the evaluation of evidence: alternative analyses.\n  *Proc. UAI 1999*. https://arxiv.org/pdf/1301.6737\n- Schum, D. A. (2001). Alternative views of argument construction from a mass of evidence.\n  *Cardozo Law Review* 22:1461. https://larc.cardozo.yu.edu/cgi/viewcontent.cgi?article=2745&context=clr\n- Schum, D. A. (2009). A science of evidence: contributions from law and probability. *Law,\n  Probability and Risk* 8(3):197–231.\n- Schum, D. A. (2011). Classifying forms and combinations of evidence: necessary in a science of\n  evidence. In Dawid, Twining & Vasilaki (eds.), *Evidence, Inference and Enquiry*, British\n  Academy/OUP. https://philpapers.org/rec/SCHCFA-2\n- Schum, D. A. & Kelly, C. W. (1973). A problem in cascaded inference: determining the inferential\n  impact of confirming and conflicting reports from several unreliable sources. *Organizational\n  Behavior and Human Performance* 10:404–423.\n- Schum, D. A. & Morris, J. R. (2007). Assessing the competence and credibility of human sources of\n  intelligence evidence. *Law, Probability and Risk* 6(1–4):247–274.\n- Schum, D., Tecuci, G., Boicu, M. & Marcu, D. (2009). Substance-blind classification of evidence for\n  intelligence analysis. *Ontology for the Intelligence Community (OIC 2009)*.\n  https://c4i.gmu.edu/OIC09/papers/OIC2009_1_SchumEtAll.pdf\n- Schum, D. et al. (2009). Analyzing evidence and its chain of custody. George Mason University\n  Learning Agents Center.\n  http://lac.gmu.edu/publications/2009/Schum%20et%20al%20-%20Chain%20of%20Custody.pdf\n- Tillers, P. & Schum, D. (1988). Charting new territory in judicial proof: beyond Wigmore. *Cardozo\n  Law Review* 9:907. https://larc.cardozo.yu.edu/cgi/viewcontent.cgi?article=1309&context=clr\n- Twining, W. (1985). *Theories of Evidence: Bentham and Wigmore*. Stanford UP. [recalled]\n- Godden, D. (2014). Modeling corroborative evidence: inference to the best explanation as\n  counter-rebuttal. *Argumentation* 28. https://link.springer.com/article/10.1007/s10503-013-9308-9\n- (2025). Measuring evidential phenomena for complex reasoning patterns about a mass of evidence.\n  *Artificial Intelligence and Law*. https://link.springer.com/article/10.1007/s10506-025-09455-9\n\n**Intelligence**\n- ODNI. *Intelligence Community Directive 203: Analytic Standards*, 2 Jan 2015, as amended 2022.\n  https://www.dni.gov/files/documents/ICD/ICD-203.pdf (read via archive.dni.gov)\n- NATO. STANAG 2511 (2003); AJP-2.1 Allied Joint Doctrine for Intelligence Procedures (2016 ed.).\n  [scales recalled; quoted fragments via Kelly et al. 2025]\n- Kent, S. (1964). Words of estimative probability. *Studies in Intelligence* 8(4); in D. P. Steury\n  (ed.), *Sherman Kent and the Board of National Estimates* (CIA CSI, 1994).\n- Heuer, R. J., Jr. (1999). *Psychology of Intelligence Analysis*. CIA Center for the Study of\n  Intelligence. https://www.cia.gov/resources/csi/static/Pyschology-of-Intelligence-Analysis.pdf\n- Irwin, D. & Mandel, D. R. (2019). Improving information evaluation for intelligence production.\n  *Intelligence and National Security* 34(4):503–525. doi:10.1080/02684527.2019.1569343\n- Baker, J. D., McKendry, J. M. & Mace, D. J. (1968). *Certitude judgments in an operational\n  environment*. US Army Behavioral Science Research Lab. [via Block 2021]\n- Block, L. (2021). Critical review of the Admiralty Code.\n  https://www.blockint.nl/intel-analysis/critical-review-of-the-admiralty-code/\n- Kelly, M., Budescu, D., Dhami, M. K. & Mandel, D. R. (2025). The effect of source reliability and\n  information credibility on judgments of information quality in intelligence analysis. *Judgment\n  and Decision Making*. https://www.cambridge.org/core/journals/judgment-and-decision-making/article/effect-of-source-reliability-and-information-credibility-on-judgments-of-information-quality-in-intelligence-analysis/E67548E8010A47345C3439D45D9EC6B3\n- Samet, M. G. (1975). Quantitative interpretation of two qualitative scales used to rate military\n  intelligence. *Human Factors* 17(2):192–202. [recalled]\n- Dhami, M. K., Belton, I. K. & Mandel, D. R. (2019). The \"analysis of competing hypotheses\" in\n  intelligence analysis. *Applied Cognitive Psychology* 33(6):1080–1090. doi:10.1002/acp.3550\n- Pilditch, T. D., Fenton, N. & Lagnado, D. (2019). The zero-sum fallacy in evidence evaluation.\n  *Psychological Science* 30(2). https://journals.sagepub.com/doi/10.1177/0956797618818484\n- Friedman, J. A., Baker, J. D., Mellers, B. A., Tetlock, P. E. & Zeckhauser, R. (2018). The value\n  of precision in probability assessment. *International Studies Quarterly* 62(2):410–422.\n  [recalled]\n- Commission on the Intelligence Capabilities of the United States Regarding Weapons of Mass\n  Destruction (Robb–Silberman), *Report to the President*, 31 March 2005.\n  https://www.govinfo.gov/content/pkg/GPO-WMD/pdf/GPO-WMD-1-23.pdf ;\n  excerpt https://nsarchive2.gwu.edu/NSAEBB/NSAEBB234/WMD_report_excerpt.pdf\n- Bara, M. (2026). Epistemic Sybil resistance: multiplying AI agents without multiplying evidence.\n  arXiv:2609.01873. https://arxiv.org/abs/2609.01873\n\n**History**\n- Bernheim, E. (1889). *Lehrbuch der historischen Methode*. Leipzig.\n- Langlois, Ch.-V. & Seignobos, Ch. (1898). *Introduction to the Study of History* (trans. G. G.\n  Berry). https://www.gutenberg.org/files/29637/29637-h/29637-h.htm (§ argument from silence, pp.\n  254–256; agreement/independence, pp. 199–204)\n- Garraghan, G. J. & Delanglez, J. (1946). *A Guide to Historical Method*. Fordham UP.\n- Gottschalk, L. (1950). *Understanding History*. Knopf. [via secondary quotation]\n- Bloch, M. (1949/1953). *The Historian's Craft*. [recalled]\n- Ginzburg, C. (1979). Spie. Radici di un paradigma indiziario. In A. Gargani (ed.), *Crisi della\n  ragione*; English: Clues: roots of an evidential paradigm, in *Clues, Myths, and the Historical\n  Method* (Johns Hopkins UP, 1989). [recalled]\n\n**Journalism and fact-checking**\n- Kovach, B. & Rosenstiel, T. (2001; 4th ed. 2021). *The Elements of Journalism*. Crown. [recalled]\n- IFCN Code of Principles (2016; June 2025 version). https://ifcncodeofprinciples.poynter.org/the-commitments ;\n  https://www.poynter.org/wp-content/uploads/2025/06/ifcn-cop-june-2025-final.pdf\n- schema.org, ClaimReview type. https://schema.org/ClaimReview [fields recalled]\n- Uscinski, J. E. & Butler, R. W. (2013). The epistemology of fact checking. *Critical Review*\n  25(2):162–180. [recalled]\n- Lim, C. (2018). Checking how fact-checkers check. *Research & Politics* 5(3).\n  https://journals.sagepub.com/doi/full/10.1177/2053168018786848\n- Lee, S. et al. (2023). \"Fact-checking\" fact checkers: a data-driven approach. *HKS Misinformation\n  Review* 4(5). https://misinforeview.hks.harvard.edu/article/fact-checking-fact-checkers-a-data-driven-approach/\n- (2023). Cross-checking journalistic fact-checkers: the role of sampling and scaling in\n  interpreting false and misleading statements. *PLOS ONE*.\n  https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0289004\n- Wineburg, S. & McGrew, S. (2019). Lateral reading and the nature of expertise. *Teachers College\n  Record* 121(11). https://journals.sagepub.com/doi/10.1177/016146811912101102\n- Caulfield, M. (2019). SIFT (The Four Moves). https://hapgood.us/2019/06/19/sift-the-four-moves/\n  [recalled]\n- Silverman, C. (ed.) (2014). *Verification Handbook*. European Journalism Centre. [recalled]\n\n**Forensic science**\n- National Research Council (2009). *Strengthening Forensic Science in the United States: A Path\n  Forward*. National Academies Press.\n- PCAST (2016). *Forensic Science in Criminal Courts: Ensuring Scientific Validity of\n  Feature-Comparison Methods*.\n  https://obamawhitehouse.archives.gov/sites/default/files/microsites/ostp/PCAST/pcast_forensic_science_report_final.pdf\n- ENFSI (2015). *ENFSI Guideline for Evaluative Reporting in Forensic Science*. See also Champod,\n  Biedermann, Vuille, Willis & De Kinder (2016), *A Primer for Legal Practitioners*.\n  https://wp.unil.ch/forensicdecision/files/2015/01/ChampodBiedermannVuilleWillisDeKinder_2016PDF.pdf\n- Cook, R., Evett, I. W., Jackson, G., Jones, P. J. & Lambert, J. A. (1998). A hierarchy of\n  propositions: deciding which level to address in casework. *Science & Justice* 38(4):231–239.\n  doi:10.1016/S1355-0306(98)72117-3\n- Evett, I. W., Jackson, G. & Lambert, J. A. (2000). More on the hierarchy of propositions: exploring\n  the distinction between explanations and propositions. *Science & Justice* 40(1):3–10.\n- Thompson, W. C. & Schumann, E. L. (1987). Interpretation of statistical evidence in criminal\n  trials: the prosecutor's fallacy and the defense attorney's fallacy. *Law and Human Behavior*\n  11:167–187. [recalled]\n- Martire, K. A., Kemp, R. I., Watkins, I., Sayle, M. A. & Newell, B. R. (2013). The expression and\n  interpretation of uncertain forensic science evidence: verbal equivalence, evidence strength, and\n  the weak evidence effect. *Law and Human Behavior* 37(3):197–207. doi:10.1037/lhb0000027\n- Martire, K. A., Kemp, R. I., Sayle, M. & Newell, B. R. (2014). On the interpretation of likelihood\n  ratios in forensic science evidence: presentation formats and the weak evidence effect. *Forensic\n  Science International* 240:61–68.\n- Dror, I. E., Charlton, D. & Péron, A. (2006). Contextual information renders experts vulnerable to\n  making erroneous identifications. *Forensic Science International* 156:74–78. [recalled]\n- Dror, I. E., Thompson, W. C., Meissner, C. A., Kornfield, I., Krane, D., Saks, M. & Risinger, M.\n  (2015). Context management toolbox: a Linear Sequential Unmasking (LSU) approach for minimizing\n  cognitive bias in forensic decision making. *Journal of Forensic Sciences* 60(4):1111–1112.\n  doi:10.1111/1556-4029.12805\n- Dror, I. E. & Kukucka, J. (2021). Linear Sequential Unmasking–Expanded (LSU-E). *Forensic Science\n  International: Synergy* 3:100161.\n  https://www.sciencedirect.com/science/article/pii/S2589871X21000310\n"}