{"path":"research/what-counts-as-evidence.md","content":"# What Counts as Evidence — the academic landscape, and how evidence should enter Deliberus\n\n**Date**: 2026-09-26 · **Status**: research synthesis plus proposals. **Nothing here is ruled.** Every\nproposal is a row in the ruling ledger, Arc Q (`session26-open-rulings.md`). · **Occasion**, the\nfounder, verbatim: *\"deeply research what in academic context is considered evidence and what\ndifferent tiers of evidence can consist of. Then review how evidence can be translated optimally to\nthe ontology and processes and rationality idealized in Deliberus\"*, grounded first *\"on everything\nthat's been said on 'evidence' in the Deliberus corpus\"*, and then *\"Document everything carefully\nbefore you spit out the audio walkthrough.\"*\n\n**The dossier behind this document** is in [`what-counts-as-evidence/`](what-counts-as-evidence/README.md):\nthree corpus maps (what the docs, the code, and the founder's own words say about evidence) and five\nacademic research strands (philosophy of evidence; grading in science and policy; testimony, lived\nexperience and normative evidence; formal and computational models; law, intelligence, history,\njournalism and forensics). This document is the synthesis; the strands carry the citations, the\nnumbers, and their verification marks. Where a number below was not checked against a primary\nsource during the research pass it says so.\n\n**The listening version**: [`../listening/2026-09-26-what-counts-as-evidence.md`](../listening/2026-09-26-what-counts-as-evidence.md).\n\n---\n\n## 0. The answer in one page\n\n1. **Evidence is a relation, not a kind of object.** Every mature tradition treats \"E is evidence for\n   H\" as a three-place relation — this item, this hypothesis, against this background — and measures\n   its force *comparatively*: how much more expected is E if H is true than if it is false (the\n   likelihood ratio; Good's \"weight of evidence\"; forensic science's reported LR; Heuer's\n   diagnosticity). Deliberus already holds the relational half (\"evidence is a position in the\n   graph\"). It does not yet hold the comparative half: the code **counts** critical-question answers\n   instead of weighing them, and counting is the one aggregation every tradition surveyed rejects\n   (the vote-counting result of Hedges & Olkin 1980; the ravens paradox; the Amgoud–Ben-Naim\n   incompatibility of cardinality and quality precedence).\n\n2. **\"Tiers of evidence\" are not one ladder anywhere mature.** The fields that invented the pyramid\n   abandoned the single ladder: Oxford CEBM 2011 has one ladder per *question type*; GRADE treats study\n   design as a *starting point* moved by five named down-grades and three up-grades; every current\n   appraisal tool grades **a result, or a body of evidence for one outcome — never a source by its\n   design**. The code's methodology ladder (meta-analysis 0.70 … expert opinion 0.30) fails on six\n   independent grounds and should not become a strength input. What survives is a labelled default\n   starting point, itself an attackable claim.\n\n3. **Four quantities that mature frameworks keep apart, and Deliberus mostly fuses**: *direction*\n   (which way the evidence points), *weight* (how much independent evidence has been examined),\n   *certainty* (how far the body of evidence can be trusted, after bias, inconsistency, imprecision,\n   indirectness and reporting bias), and the *decision* (what to believe or do, which also needs priors\n   and values). The August 2026 direction/weight split is on the right side of this line; certainty is\n   the missing third axis, and the decision is correctly left to people.\n\n4. **Independence is the variable.** Agreeing reports add up only when their errors are independent.\n   Every working system that combines evidence says so explicitly (STRING's independent channels,\n   ACMG's no-double-use rule, intelligence analysis's circular-reporting checks, capture–recapture in\n   truth commissions). Deliberus has no representation of shared origin, so four claims from one study\n   count four times and three one-word answers from one account count like three trials.\n\n5. **Source reliability and report credibility are separate axes.** Intelligence (the Admiralty\n   code's A–F × 1–6), Schum's evidence science, Walton's schemes, and the persuasion literature all\n   split *is this source generally reliable* from *is this particular report credible*, and split\n   credibility further along a chain: *did they say it, did they believe it, were they positioned to\n   observe it, did it happen*.\n\n6. **To the founder's open question — \"what actually constitutes un-decomposable, more atomic\n   evidence?\"** The research converges on a precise answer, proposed here, unruled: **there is no\n   atomic proposition, but there is an atomic encounter.** Every evidential claim decomposes, but\n   what it finally rests on is not a smaller claim: it is a pointer to a contact with the world — an\n   observation, a measurement, a document span, an utterance — which cannot be decomposed into claims,\n   only *re-encountered* (re-read, re-measured, replicated). Philosophy (Haack's crossword clues),\n   evidence science (Schum's tangible evidence), fact-checking (evidence as a located span in an\n   identified source) and machine-safety work (Bengio's \"factual observation\") arrive at the same\n   shape. In the graph, that is a **locator with provenance**, not a node type.\n\n7. **To the founder's hope that the same evidence should give everyone the same probability** (ledger\n   P2): it holds for *what the evidence says* — the likelihood ratio is independent of anyone's prior,\n   so two people with different starting beliefs can agree exactly on how much a finding favours one\n   hypothesis over another. It does not hold for *what to believe*, where priors, standards of proof\n   and evidence about one's own reliability legitimately differ. The version worth keeping is that\n   **every divergence from shared evidence must be attributable to a named slot** — a prior, a\n   standard, a model, or a piece of evidence — each of which can be a claim. That is also the answer to\n   the fear of \"fuzzy gooey subjectivity\" (P3): subjectivity is allowed only in named, attributable,\n   attackable slots.\n\n8. **Royall's three questions are the cleanest statement of what Deliberus is.** *What does this\n   evidence say? What should I believe? What should I do?* Deliberus answers the first, maps the\n   inputs to the second and third, and answers neither of those for anyone. It is the most economical\n   philosophical defence of \"maps reasoning, does not endorse conclusions\" in the literature.\n\n---\n\n## 1. Where Deliberus stands (the grounding)\n\nFull maps: [`corpus-map-research-docs.md`](what-counts-as-evidence/corpus-map-research-docs.md),\n[`corpus-map-code.md`](what-counts-as-evidence/corpus-map-code.md),\n[`corpus-map-founder-voice.md`](what-counts-as-evidence/corpus-map-founder-voice.md).\n\n### 1a. What the founder has said and ruled\n\n- **The founding conviction** (Session 7, March 2026): *\"I don't want any copouts, I want people and AI\n  agents surfing around in Deliberus to see exactly what the best minds on our planet think in detail\n  and what the best evidence shows - evidence that also is decomposed into its constituent\n  parts/arguments. … What actually constitutes un-decomposable, more atomic evidence is an open\n  question to me, scientifically, rationally, philosophically.\"* And: a meta-analysis PDF *\"should be\n  decomposed as far as possible - maximum granularity that is helpful to the rest of humanity.\"*\n- **The August answer** — *\"there is no atomic evidence, only claims whose decomposition nobody has\n  demanded yet\"* — was Claude's, not his (`scheme-bounded-decomposition-and-evidence-as-subgraph.md`,\n  2026-08-16). His own later correction points the same way: *\"evidence breaks down into a subgraph,\n  and all of the subgraph claims can/should be dealt with in depth\"* (2026-09-14, ledger O2).\n- **Rulings that constrain evidence handling**: the mother-claim pattern (April); fact-checking *\"at\n  least as important as\"* values-checking, facts first (2026-08-18); evidence named among what must be\n  right before launch (*\"the meta layer is the gate\"*, 2026-08-27); fallible self-knowledge\n  (2026-08-25); bare judgments are telemetry and there is no ratify act (2026-08-31); confident\n  housekeeping is the machine's job (2026-09-13); horizon kinds read, never set (2026-09-15).\n- **Open positions**: P2 (same evidence, same probability, \"ideally\"), P3 (the subjectivity fear), P4\n  (the methodology ladder), P7 (does Deliberus ever output a probability), and the evidence-division\n  question he said *\"might challenge the foundation\"* (2026-08-21).\n\n### 1b. What the code actually does\n\n- **Evidence is a position, not a type**: any claim at the tail of a SUPPORTS or ATTACKS edge.\n- **Strength moves through two channels only**: answers to critical questions on incoming\n  scheme-bearing edges, and decomposition children. Answers are **counted** (+0.15 each on a polarity\n  claim, capped at +0.45); QEM is applied twice (edge, then claim), so one fully answered six-question\n  expert-opinion edge lifts its claim only to 0.58, and a one-question scheme such as\n  `lack_of_evidence` can never leave the neutral band on one edge. An answer's content, its own\n  standing, attacks on it, and whether it duplicates another answer do not matter.\n- **Every property a scientist would call evidential quality is stored and inert**: methodology tier,\n  sample size, effect size, confidence interval, study count, heterogeneity, `evidence_type`,\n  `epistemic_status`, `confidence`, edge `strength`, `trust_tier`. User-added evidence is always\n  labelled `anecdotal` and `observed`.\n- **Weight** (`evidential_weight.py`) counts non-neutral inputs; it does not saturate and has no notion\n  of independence.\n- **Missing entirely**: source reliability separate from the claim; source independence; likelihood\n  ratios; sample-size or precision weighting; \"we searched and found nothing\" as data; chain of custody,\n  quote verification, page or timestamp locators; timestamps on edges and answers; a scheme for study\n  evidence as such; typed rebutting versus undercutting; an epistemic status that says \"open\".\n- **Present and sound**: the direction/weight split with no number for the unexamined; weakest-part\n  semantics for required parts; human-ratified validity lapse; non-propagating reported speech; built\n  but unadopted evidence routing with fair-share discounting.\n- **Defects found by reading the code** are listed in `TODO.md` § Evidence layer, separately from\n  the proposals here. All but one were repaired on 2026-09-27; the one that proved to be a\n  strength-semantics decision (a fully answered one-question edge can never leave the neutral band,\n  and one answer moves a lone edge's claim by under 0.0005) is ledger row Q27.\n\n### 1c. Contradictions inside the corpus this research resolves or sharpens\n\n- **Is impact testimony \"weighed as evidence\"?** `what-human-judgment-is-for.md` § 5 says never; the\n  derivability doc and `cheap-to-add-slow-to-matter.md` say yes. § 6 below dissolves the conflict: it is\n  three different edges.\n- **Three unmapped vocabularies of evidence kind** (`evidence_type`; the April four kinds; the\n  methodology tier). § 3 maps them onto the literature's axes: they are *form*, *substance* and *design*\n  respectively, and none of them is *strength*.\n- **Older docs call QEM \"Bayesian\"** (`qbaf-gradual-semantics-research.md`,\n  `inductive-reasoning-research.md` § 9). Corrected in place today with pointers to P5.\n\n---\n\n## 2. What academia means by evidence\n\nFull strand: [`philosophy-of-evidence.md`](what-counts-as-evidence/philosophy-of-evidence.md).\n\n### 2a. The concept has several jobs that pull apart\n\nKelly's map (SEP, \"Evidence\"): evidence is what **justifies belief**, what a rational thinker\n**respects**, a **sign or symptom** of how things are, and a **neutral arbiter** between people who\ndisagree. The last role — public, intersubjective, settling disputes — is the one Deliberus is built\nfor, and the one philosophers doubt evidence can fully play, because *what counts as evidence is itself\ncontested*. The FRI adversarial collaboration on AI risk found exactly that residue among good-faith\nexperts after 25–100 hours each: different **standards** of evidence. The design answer the literature\nsupports is the one Deliberus already leans toward: where evidence cannot be neutral, make the\ndisagreement about what counts *public*, as claims.\n\n### 2b. The quantitative core: evidence is comparative\n\n- **Incremental vs absolute confirmation**: raising a probability is not the same as making it high\n  (Carnap). The badge currently reads as both.\n- **Measures of confirmation disagree** (Fitelson's measure sensitivity). The **log-likelihood ratio**\n  — Good's weight of evidence, Turing's decibans — is the one that *adds* across conditionally\n  independent items, which is the formal warrant for summing in the corroborative channel **and the\n  exact condition under which summing is licit**.\n- **Royall's three questions** (*Statistical Evidence*, 1997): what should I believe (needs a prior),\n  what should I do (needs losses and values), and what does this evidence say (needs neither — the\n  likelihood ratio). Statistics had conflated them.\n- **Severity** (Mayo): data are evidence for H only if the test would probably have *failed* H were H\n  false. Publication bias, optional stopping and data-dredging lower severity without changing the\n  data — so they attach to the *inference*, not the conclusion.\n- **Weight vs balance** (Keynes 1921; Joyce 2005; Skyrms's resilience): how much evidence stands behind\n  a judgment differs from which way it points. Keynes's weight is the amount of *relevant* evidence;\n  Joyce adds specificity; Skyrms reads weight as *resilience* — how little the judgment would move under\n  further evidence.\n- **Total evidence** (Carnap; Good's theorem that free evidence never hurts): a verdict from less than\n  the available evidence is defective even if every step is valid — the formal name for the corpus's\n  *coherent absence*.\n\n### 2c. Aggregation: independence, variety, coherence\n\n- **Agreement multiplies only across independent errors** (Condorcet; Bovens & Hartmann; Easwaran et\n  al.'s \"synergy\"). With a shared origin, agreeing reports can be worth barely more than one.\n- **\"Diverse evidence beats repeated evidence\" is false in general** (Bovens & Hartmann 2003; Claveau\n  2013; Osimani & Landes 2023): two concordant reports from one instrument can be evidence *that the\n  instrument is reliable*. Whether variety helps depends on how sources fail. That is why the corpus's\n  rule — *a known shared source becomes a node, never a blind discount* — is exactly what this\n  literature recommends.\n- **Coherence is not truth-conducive in general** (the Bovens–Hartmann and Olsson impossibility\n  results): a tightly coherent cluster from one source family is not thereby likely true.\n- **Holism** (Duhem–Quine) is real, but blame *is* localisable once auxiliaries are explicit and have\n  independent support — which is the payoff of making implicit premises into claims.\n- **Inference to the best explanation** is comparative: an abductive edge's strength depends on the\n  *set of rival explanations present*, so rivals should be sibling claims and \"no rival proposed\" a\n  coverage gap, not strength.\n- **Evidential pluralism** (Russo–Williamson): causal claims normally need both *difference-making*\n  evidence (correlation, trials) and *mechanistic* evidence. **Evidence for use** (Cartwright &\n  Hardie): \"it worked there\" is not \"it will work here\"; relevance is a property of study + target +\n  argument, i.e. of the **edge**.\n\n### 2d. Is any evidence atomic?\n\nSellars (the Myth of the Given), Quine and modern modest foundationalism agree that **no evidence is\nimmune to challenge** — which supports \"no atomic evidence\" as anti-foundationalism. Two qualifications\nfrom the literature:\n\n- **Haack's crossword** (*Evidence and Inquiry*, 1993): experiential clues are not beliefs and are not\n  infallible, but they are a *different kind of input* from the intersecting entries. A warrant\n  structure that treats clues as just more entries can be completed coherently while ignoring the\n  clues — the coherentist fairy tale. Evidence has an anchor other claims lack.\n- **\"Demanded\" should be \"reasonably challenged\"** (Brandom's default-and-challenge; Rescorla on\n  dialectical regress). If any bare demand triggers decomposition, decomposition becomes a stalling\n  weapon: a strategy-class adversary can demand forever.\n- **Hinges** (Wittgenstein; Pritchard) are commitments that make evaluation possible and are not\n  improved by evidential support — a principled place where decomposition stops being productive, and\n  should not be scored as a copout.\n\n---\n\n## 3. How science and policy grade evidence\n\nFull strand: [`grading-evidence-in-science-and-policy.md`](what-counts-as-evidence/grading-evidence-in-science-and-policy.md).\n\n### 3a. The pyramid, and why its inventors left it\n\nThe classic ordering — systematic review/meta-analysis > RCT > cohort > case-control > case series >\ncase report > expert opinion — is a proxy for one thing: how likely confounding and selection are to\ndistort an *intervention effect*. It was never a general measure of evidential weight. Its successors:\n\n- **Oxford CEBM 2011**: a *table* of ladders, one per question (prevalence, diagnosis, prognosis,\n  treatment benefit, common harms, rare harms, screening), with a footnote moving a study down for\n  bias, imprecision, indirectness or inconsistency and up for a large effect [verified]. For \"how\n  common is X\" the top rung is a local random survey; an RCT is irrelevant. An observational study\n  with a dramatic effect sits beside the RCT.\n- **The \"new pyramid\"** (Murad et al. 2016): wavy lines between designs, and the systematic review\n  removed from the apex and drawn as a **lens** — a reading *of* the studies below, inheriting their\n  quality.\n- **Critiques**: Worrall, Howick, Cartwright, Bluhm (\"from hierarchy to network\" — the closest\n  philosophical statement of a graph view), Stegenga (meta-analysis is malleable: many defensible\n  choices, contradictory conclusions from the same evidence).\n\n### 3b. GRADE, the operative framework\n\n- **Certainty** — high / moderate / low / very low — *that the true effect lies on one side of a\n  threshold*: certainty is always relative to a claim, never a free-floating property of studies.\n- **Starting point by design** (randomised bodies high, observational low; since GRADE guideline 18,\n  non-randomised studies assessed with ROBINS-I start high and are marked down for the bias actually\n  found [verified]).\n- **Five down-grades**: risk of bias, inconsistency, indirectness, imprecision, publication bias.\n  **Three up-grades**: large effect, dose–response, plausible confounding that would *reduce* the\n  effect.\n- **Unit**: the body of evidence for one outcome. The same review can be \"high\" for one outcome and\n  \"very low\" for another.\n- **Certainty ≠ recommendation** (Evidence-to-Decision frameworks: values, resources, equity,\n  acceptability and feasibility are separate rows). In WHO guidelines, 55.5% of strong recommendations\n  rested on low or very-low certainty [verified] — decisions routinely outrun evidence, which is\n  legitimate when other rows justify it and illegitimate when they don't; only a structure that holds\n  both makes the difference visible.\n- **Grading is noisy**: trained GRADE raters reach inter-rater reliability of about 0.66–0.72 against\n  0.27–0.31 for intuitive judgements, but agreement on imprecision and risk of bias can fall to κ\n  ≈ 0.06–0.21 [verified]. Grading is contested ground: the machine should propose and people argue.\n- **GRADE-CERQual** grades *qualitative findings* on methodological limitations, coherence, adequacy of\n  data and relevance — deliberately not on the same ladder as effect estimates.\n\n### 3c. The recurring structure across fields\n\nGRADE, IPCC, EFSA's weight-of-evidence guidance, IARC, the EEF padlocks and the environmental-health\nOHAT method repeat four moves: **separate the lines of evidence by type; appraise each on a few named\ndimensions (bias/reliability, relevance/indirectness, precision/quantity, consistency, reporting bias);\nintegrate by explicit, mostly ordinal rules, often weakest-link; report certainty separately from the\ndirection and size of the effect and from the decision.** No mature framework produces one cardinal\nnumber from a design label.\n\nThree worked separations to borrow:\n\n- **IPCC**: *confidence* (evidence limited/medium/robust × agreement low/medium/high) is orthogonal to\n  *likelihood* (a calibrated probability scale). Readers systematically under-read verbal probability\n  words (Budescu), so words should travel with numbers where numbers exist.\n- **IARC** (2019 Preamble) [verified]: *hazard* (can it?) is separated from *risk* (how much, at what\n  exposure); three streams (human, animal, mechanistic) combine by rules; and **Group 3, \"not\n  classifiable\", is an explicit \"this is open\" status** — the category Deliberus's epistemic-status list\n  lacks.\n- **EEF padlocks** [verified]: the interim rating is the **lowest** of design, detectable effect size\n  and attrition, then marked down for threats — weakest-link-then-downgrade, and \"the ratings take no\n  account of whether the intervention itself was successful\": certainty separate from direction.\n\n### 3d. Meta-analysis, statistics and the replication record\n\n- **Meta-analyses mislead in known ways**: garbage in; redundancy; and poor handling of dependence and\n  publication bias. A 2026 study of 1,000 meta-analyses across ten disciplines (Wu, Duan, Reed & Tipton,\n  *Research Synthesis Methods* 17(1)) found fewer than half of those with dependent effect sizes\n  adjusted for dependence, and fewer than half of those that detected publication bias adjusted for it\n  [verified]. Re-running a meta-analysis's arithmetic is highly reproducible (86% with data or code);\n  re-making its *choices* is not (Hamilton et al. 2026) — Stegenga's malleability, measured.\n- **Vote counting fails**: tallying significant against non-significant studies can converge on the\n  wrong answer as underpowered studies accumulate (Hedges & Olkin 1980) [recalled].\n- **p-values** do not measure the probability that a hypothesis is true (ASA 2016); a \"significant\"\n  p = 0.05 corresponds under broad conditions to a Bayes factor of at most about 2.5 (Sellke, Bayarri &\n  Berger 2001) [recalled] — weak evidence.\n- **Many analysts, one dataset**: 29 teams produced odds ratios from 0.89 to 2.93 on one question\n  (Silberzahn et al. 2018) [verified]. A single study's estimate is a draw from a distribution over\n  defensible analyses.\n- **Replication**: roughly 36–62% of findings replicated in the large psychology and social-science\n  projects [recalled/verified by project]; a high-profile 2023 paper claiming high replicability was\n  itself retracted — replication status needs a status.\n- **E-values** (betting-based, Shafer; Vovk & Wang; Grünwald et al.) multiply across independent studies,\n  average validly under arbitrary dependence, and stay valid under optional continuation — the only\n  mainstream evidence measure designed for accumulating, possibly dependent, sequentially-chosen\n  evidence, which is the graph's situation.\n\n### 3e. The methodology ladder, assessed\n\nThe ladder in `models.py` fails on six grounds: it ignores **question type**; it grades the wrong\n**unit** (a source, not a result or body); it treats **design as a verdict** rather than a starting\npoint; it ranks **meta-analysis as a tier** rather than a reading of its inputs (and double-counts the\ntrials inside it); it puts **expert opinion** on the ladder, which GRADE's 2019 position calls an\ninterpretation of evidence, not evidence; and it has **no place for mechanistic or qualitative\nevidence**. A seventh, internal: a cardinal number (0.45 vs 0.55) claims interval meaning no framework\nclaims, and the numbers are uncited house choices — by the corpus's own sorting test, a commitment two\nusers could disagree about belongs in the graph, not the ontology. **What survives**: a labelled\ndefault starting certainty for intervention-effect claims, stated as a claim anyone can attack.\n\n---\n\n## 4. What the professions of proof teach\n\nFull strand: [`law-intelligence-history-journalism-forensics.md`](what-counts-as-evidence/law-intelligence-history-journalism-forensics.md).\n\nFive professions decide under evidence for a living: law, intelligence analysis, history, journalism\nand forensic science. None of them has a genre ladder at the centre of its practice, and **none of\nthem resolves anything by counting supporting items.** The same four moves recur in all five, and law\nadds a fifth.\n\n### 4a. The four moves every profession makes\n\n1. **Separate the source from the report.** How good a sayer is in general is a different question\n   from whether this statement is true. NATO's Admiralty code grades the two on separate axes (A–F for\n   source reliability, 1–6 for information credibility); Schum separates a source's *competence* from\n   its *credibility*; historians separate *external* criticism (is the document what it claims to be)\n   from *internal* criticism (is what it says true); law separates authentication from weight. The\n   axes are meant to disagree — a reliable source can be deceived this once, an unreliable one can be\n   independently confirmed — and **the disagreeing cases are the most informative ones**. The failure\n   is measured: in about 1,400 US Army field reports, **87% of Admiralty ratings fell on the diagonal**\n   (A1, B2, C3), so in practice the two axes collapsed into one (Baker, McKendry & Mace 1968, via Irwin\n   & Mandel 2019).\n2. **Break credibility into named attributes that can each be attacked.** The hearsay tradition names\n   four testimonial infirmities: perception, memory, narration and sincerity (Tribe 1974). Schum names\n   veracity, objectivity and observational sensitivity for testimony; authenticity, accuracy and\n   reliability for tangible evidence. This is the epistemology behind the hearsay rule, which is **not**\n   \"second-hand information is worthless\" (historians keep hearsay; Gottschalk says so outright). Its\n   actual content is: *a claim whose credibility attributes cannot be interrogated should not be\n   weighted as if they had been.* Cross-examination is the machine for interrogating them, and every\n   hearsay exception names a structural reason why one attribute need not be tested: a statement\n   against interest (sincerity), a routine business record (memory and sincerity), an excited\n   utterance (no time to fabricate). **Answering critical questions is Deliberus's cross-examination**,\n   and an exception-type fact (\"this was a statement against the speaker's own interest\") is itself a\n   claim that argues for a higher starting weight.\n3. **Weigh evidence by how much it discriminates, not by how much of it there is.** Forensic science\n   calls this the likelihood ratio; Heuer's Analysis of Competing Hypotheses calls it *diagnosticity*;\n   Schum calls it inferential force; Federal Rule of Evidence 403 excludes \"needlessly … cumulative\n   evidence\". Heuer's image carries it: a high temperature tells a doctor the patient is sick and almost\n   nothing about which illness, and *\"most of the evidence supporting what you believe is the most\n   likely hypothesis really is not very helpful, because that same evidence is also consistent with\n   other hypotheses.\"* One caution travels with it: diagnosticity is relative to the rival set, so when\n   the rivals are not exhaustive, \"non-diagnostic\" misleads (the zero-sum fallacy, Pilditch, Fenton &\n   Lagnado 2019). And ACH *as a procedure* is weakly supported — in a randomised study analysts trained\n   in it did no better (36% against 33% correct) and produced far more tied rankings (Dhami, Belton &\n   Mandel 2019). **The concept is settled; that particular procedure is not.**\n4. **Count agreement only in proportion to independence.** Langlois and Seignobos, the 1898 manual of\n   historical method, has as chapter headings *\"Agreement of statements — Necessity of proving them to\n   be independent — Perfect agreement not so conclusive as occasional coincidence.\"* Verbatim\n   agreement suggests copying; stemmatics turns that into a method in which **shared errors reveal a\n   common ancestor**, so a family of copies counts once. The canonical failure is Iraq WMD: one source,\n   Curveball, never met by any US officer, supplied about a hundred reports; the intelligence estimate\n   attributed the judgment to \"multiple sensitive sources\"; the daily briefs *\"left an impression of\n   many corroborating reports where in fact there were very few sources\"* (Robb–Silberman Commission,\n   2005). The Commission's record holds **four separate failures**, each with a graph analogue: one\n   source's many reports counted as volume · \"multiple sources\" asserted without independence · a\n   source flagged as a fabricator in May 2002 still used in October, because **the retraction did not\n   propagate** · and a relay that saw itself as *\"just a conduit\"*, so **passing a report on laundered\n   its credibility**. A September 2026 preprint states the machine version: with the evidence fixed and\n   the number of reporting agents raised from 1 to 32, naive aggregation dropped posterior coverage from\n   0.94 to 0.26 — *\"another agent is not another observation\"* (Bara 2026, arXiv:2609.01873, abstract\n   read). Every LLM pass in Deliberus is a relay in exactly this sense.\n\n**The fifth move, law's: the threshold is set by the question and the cost of error**, never by one\nuniversal scale — preponderance in a civil suit, beyond reasonable doubt in a criminal one, different\nburdens on different parties. Carneades formalises this in the same Walton lineage Deliberus already\nuses: **each statement carries its own proof standard**, and acceptability means meeting it. ⚠\n**Correction to a figure that circulates, including in this corpus's grounding notes: the 2007\nCarneades paper (Gordon, Prakken & Walton, *AI* 171) defines three standards — scintilla of evidence,\nbest argument, dialectical validity.** The five-standard version adding preponderance, clear and\nconvincing, and beyond reasonable doubt is Gordon & Walton **2009**, and its exact definitions were not\nre-verified for this pass. Three things from the 2007 paper matter here: a statement can meet a\nstandard **and its complement** at once at low standards, which is the honest difference between\n*acceptable* and *true*; burden is allocated by **premise type** (an *ordinary* premise must always be\nbacked, an *assumption* holds until questioned, an *exception* does not bite until backed), so some\ncritical questions merely ask while others must bring evidence; and Verheij's four roles for critical\nquestions separate those that **undercut** from those that **rebut**. Deliberus today has none of this:\none badge, no standard, and critical questions all of one type.\n\n**Where Deliberus stands against the five**: it implements none of (1), (3), (4) or the fifth, and\nimplements (2) only through Walton critical questions whose answers are counted rather than weighed.\n\n### 4b. Schum's combinations: what two items pointing at one claim can be\n\nThe most important single piece of the strand. Deliberus has one evidential relation, and its\ncontributions add. Schum (2009) shows the same surface pattern — two items bearing on one claim — is at\nleast five different things:\n\n| Pattern | What it is | How it should combine |\n|---|---|---|\n| **Corroborative** | different sources report the **same event** | *saturating*: a second report *\"springs to life to the extent that we find the first source not credible\"*; a perfectly credible first source makes the rest redundant |\n| **Convergent** | different events, same conclusion | additive, and possibly **synergistic** (together they exclude a rival neither excludes alone) |\n| **Cumulative redundancy** | a different event whose force is already contained in the first | adds nothing new — law's own word *cumulative* |\n| **Contradictory** | mutually exclusive events | resolved **on credibility, never on headcount** |\n| **Divergent** | compatible events pointing different ways | needs a joint explanation, not a tally |\n\nAdditive aggregation overcounts the first, third and a collusive variant, and undercounts the synergy.\nThe formal result under the contradiction row is Schum & Kelly (1973): a majority report wins *\"iff\nthe joint reliability of the majority sources exceeds the joint reliability of the sources in the\nminority. Thus, source reliability is more important than the extent of source agreement.\"* So **a\nbadge must never show three-against-one as a majority.** The corpus's own five-case double-counting\ntaxonomy (August 31) is the *structural* half of this; Schum supplies the *semantic* half — same event\nor different events, redundant or synergistic. It also names the repair for those of the census's 70 not-a-part edges that really are evidence (⚠ 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)): three studies of one effect are **corroborative witnesses**\n(saturating), never necessary parts (weakest-link).\n\n### 4c. History: silence, and witnesses in spite of themselves\n\nTwo further contributions from source criticism. **The argument from silence has stated conditions**\n(Langlois & Seignobos): silence proves something only if no document mentioning the event ever\nexisted (*\"if the documents are lost we can conclude nothing\"*), and only if the event is of a kind\nthat could not fail to be recorded. They name its typical abuse too — silence about *\"the secret acts\nof governments and the complaints of the lower classes\"*, the origin of the *good old times* sophism.\nThat is a critical-question set for the argument from ignorance **with a political-bias warning\nattached**: an archive that systematically omits a class of events cannot support negative conclusions\nabout that class. Heuer independently supplies the answer menu for any absence — *not happened, not\nobservable, concealed, or not looked for*. And **Bloch prized \"witnesses in spite of themselves\"** —\na merchant's ledger over a chronicle — because unintended traces carry no motive to mislead: veracity\nis a question only for intended testimony.\n\n### 4d. Journalism: agree on the findings, disagree on the scale\n\nFact-checkers **agree on the facts and disagree on the rating.** Lim (2018) compared PolitiFact and the\nWashington Post on 77 shared statements: agreement was good on outright falsehoods and plain truths and\nfell to 14 of 23 in the ambiguous range, and the discrepancies came from *\"subjective judgment of the\nsignificance of inaccuracies … rather than … whether the statement itself was true.\"* On one Jeb Bush\nclaim both gave **identical reasons** and verdicts two notches apart. Lee et al. (2023) found 749\nSnopes–PolitiFact matches with only **one** genuine conflict once scale granularity and claim wording\nwere adjusted for. So **the disagreement lives in the step that maps a set of findings onto one\nordinal verdict** — the step Deliberus declines to take. The transferable rule: each finding is its own\nclaim with its own *significance* claim (*this omission matters because…*), and a free external\nvalidation target follows — on claims checked by several organisations, a Deliberus decomposition\nshould reproduce their shared findings wherever their ratings differ.\n\n### 4e. Forensic science: validity is measured, wording is dangerous, order matters\n\n- **Validity of a method is measured, not credentialed.** PCAST (2016): a subjective method's\n  validity *\"can only be established through multiple, appropriately designed black-box studies\"* and\n  *\"neither experience nor professional practices can substitute\"* for it. Measured false-positive\n  rates: latent fingerprints **as high as 1 in 306 to 1 in 18** depending on the study; firearms 1 in\n  66 (bound 1 in 46) on a single study, judged insufficient; bitemarks typically above 10%. The law's\n  own version separates validity of the method (FRE 702(c)) from validity of this application (702(d)).\n  This is the professions' strongest argument against a genre ladder: **they grade the specific method\n  on this report by its measured error rate**, not the kind of evidence it is.\n- **Report a likelihood, never a posterior.** ENFSI's guideline requires findings to be reported as\n  *\"support for Hp rather than Hd\"* on a six-band verbal scale (weak 1–10 through extremely strong\n  above a million). The wording exists to block two named errors: the **prosecutor's fallacy**\n  (reading the chance of the evidence given innocence as the chance of innocence) and the **defence\n  fallacy** (dismissing a huge likelihood ratio because the posterior is still modest).\n- **The weak evidence effect.** Evidence described as *weakly* supporting guilt led participants to\n  **lower** their belief in guilt — they inverted it (Martire et al. 2013, about 900 participants),\n  most in verbal conditions. A badge reading *weak support for X* can be heard as evidence against X.\n- **Order changes judgment.** Examiners given irrelevant context reach different conclusions on the\n  same evidence, including contradicting their own earlier identifications (Dror). *Linear sequential\n  unmasking* has the examiner analyse the trace alone, document it, and only then see the reference,\n  with restricted revisions afterwards. PCAST found fingerprint examiners altering the features they had\n  marked after seeing a matching exemplar — the same circularity a claim graph invites when it shows an\n  answerer the current badge before they answer.\n- **The hierarchy of propositions.** One finding has a different force at each level: *sub-source*\n  (the DNA is his) → *source* (the stain is his) → *activity* (he did the act) → *offence* (he is\n  guilty). A DNA match can be strong support at sub-source level and nearly neutral at activity level,\n  because transfer and persistence lie in between. Outside crime the ladder reads *the study measured\n  X → X occurred in the sample → the intervention caused X → the policy is good*. **An edge that jumps\n  more than one level skips exactly the steps that carry the contestable assumptions**, and only the\n  lower levels can hand over to an empirical oracle — which fits the horizon-kind reading.\n\n### 4f. What the professions say not to import\n\nA fixed methodology ladder as weight (every profession grades the report and its chain, not the genre)\n· a single ordinal verdict scale (the fact-checkers' lesson) · ACH's disconfirmation-only integration\nrule (analysts' own both-sides adding was no worse) · and headcount resolution of contradictions\n(refuted formally by Schum & Kelly, and by historians' own maxim *majority does not rule*).\n\n**Contested inside the professions, and kept contested here**: whether legal proof is probabilistic or\nexplanatory (Allen & Pardo against the Bayesians); what explains the refusal to convict on naked\nstatistics; whether verbal scales help or harm lay understanding; whether black-box studies are the\nonly route to validity for subjective methods.\n\nThe concrete proposals from this strand are rows Q22–Q26 in § 8 and the ledger; most of the rest\nalready sits in Q3 (weigh, don't count), Q4 (independence as structure), Q9 (typed attacks and\npremise types), Q10 (absence as a search claim), Q12 (source against report) and Q16 (standards of\nproof).\n\n---\n\n## 5. Formal and computational models\n\nFull strand: [`formal-and-computational-models-of-evidence.md`](what-counts-as-evidence/formal-and-computational-models-of-evidence.md).\n\n- **Rebut, undercut, undermine** (Pollock; ASPIC+). A rebutter is a reason against the conclusion; an\n  undercutter says the premises don't support it *here* (the object looks red, but it is lit by red\n  light); an underminer attacks a premise. **An undercutter should move a claim toward \"uninformed\",\n  not toward \"false\"** — which a model where every attack subtracts strength cannot express.\n- **Critical questions are not homogeneous** (Verheij; Prakken; Carneades): some ask whether a premise\n  holds, some point to an exception, some raise a counter-argument. Carneades adds the burden\n  consequence: an open *assumption* question (is E an expert?) withholds weight until answered; an open\n  *exception* question (is E biased?) costs nothing until someone substantiates it.\n- **Counting vs quality is a proved trade-off** (Amgoud & Ben-Naim 2013/2018; Bonzon et al. 2016):\n  \"fewer attackers wins\" and \"one strong objection beats many weak ones\" cannot both hold. QEM with sum\n  is cardinality-sensitive and fully compensating. The literature says: choose, and state the choice.\n  Carneades compares *strongest* arguments; GRADE rates a body by its best design then downgrades; ACMG\n  caps reuse of the same data.\n- **Subjective logic** (Jøsang): an opinion is (belief, disbelief, uncertainty, base rate). It is a\n  published algebra for exactly the badge's direction/weight split, with three things the badge lacks: a\n  **base rate** per kind of claim, **saturation** of weight by construction, and **fusion operators keyed\n  to an independence judgement** (cumulative for independent sources, averaging for dependent ones).\n  Trust discounting through a chain of sources drives a long hearsay chain toward *uncertainty*, not\n  disbelief. Denœux's cautious rule is the idempotent (min-based) combination for non-distinct evidence\n  — the same dial the weakest-link doc found.\n- **Evidence idioms** (Fenton, Neil & Lagnado): an evidence node has a separate *accuracy* parent;\n  dependency between witnesses is its own idiom; **explaining away** (evidence for one cause lowering\n  another) is a cross-item interaction no edge-local rule can represent.\n- **Truth discovery**: copy detection keys on **shared errors**, not shared truths — independent\n  sources rarely make the *same* mistake (Dong, Berti-Équille & Srivastava 2009); Google's\n  Knowledge-Based Trust separates a wrong source from a wrong *extraction* of a source.\n- **Ontologies**: ECO separates *what kind of evidence* from *who asserted it (human or machine)*;\n  nanopublications split the provenance *of the assertion* from the provenance *of the record*; SEPIO's\n  **evidence line** bundles several items into one argument with its own direction and strength (the\n  natural home for those of the census's 70 not-a-part edges that really are evidence); **discourse\n  graphs** give Evidence its own node type — a particular, past-tense, source-bound finding — distinct\n  from a general Claim. Deliberus can keep one node type and still keep those distinctions: a *scope*\n  (particular vs general), a mandatory *locator* on particular findings, and a *generalises* inference\n  carrying the external-validity questions.\n- **GO evidence codes** are a warning: their designers said the codes describe the type of work, not\n  quality — and users turned them into a quality ladder anyway. Any displayed evidence *type* will be\n  read as a *weight*.\n- **Fact-checking** (FEVER, SciFact, AVeriTeC, FActScore, AIS/ALCE) always treats evidence as **a\n  located span in an identified source, aligned to one sub-claim**. AVeriTeC's evidence is literally\n  question–answer pairs — critical questions arrived at independently — and it has a fourth verdict,\n  *conflicting evidence / cherry-picking*, distinct from *not enough evidence*.\n- **A production system for combining heterogeneous evidence with explicit strength tiers**: clinical\n  variant classification (ACMG/AMP 2015), recast by Tavtigian et al. (2018) as a naive-Bayes model —\n  prior 0.10; \"very strong\" = odds 350, each tier down the square root of the one above (18.7, 4.33,\n  2.08) — and then as points (supporting 1, moderate 2, strong 4, very strong 8; ≥10 pathogenic, 6–9\n  likely pathogenic, 0–5 uncertain) [verified; the benign cut-points differ between sources]. A type of\n  evidence earns a tier only at the diagnosticity it has been **measured** to achieve (ClinGen\n  calibrations); the same data may not be used under two criteria; and \"uncertain significance\" is the\n  default outcome of thin or conflicting evidence. **STRING** combines evidence channels as independent\n  probabilities with the prior removed and restored, and treats *all text-mining evidence as one\n  channel*, so a thousand abstracts repeating one finding cannot multiply.\n\n---\n\n## 6. Testimony, lived experience and normative evidence\n\nFull strand: [`testimony-experience-and-normative-evidence.md`](what-counts-as-evidence/testimony-experience-and-normative-evidence.md).\n\n### 6a. Evidential advantage, not authority\n\nStandpoint epistemology at its most defensible (Wylie 2003), epistemic-injustice theory (Fricker 2007),\nthe introspection literature and the expert-elicitation literature all support **evidential advantage**\n— being in a position makes one more likely to know something specific, to a degree, checkably — and all\nwarn against sliding into **authority**, where a position's say-so closes a question. Táíwò's\nelite-capture critique is, underneath, statistical: *the testimony a platform receives is a\nself-selected sample of the positions it claims to represent*. Medina and Davis add that credibility is\ncomparative: in any ranked or truncated view, lifting one class of testimony lowers everyone else's. The\ncorpus rule \"testimony is evidence for a premise, never a verdict\" is the advantage-without-authority\nposition in software, and it holds up.\n\n### 6b. The strongest evidence *for* impact testimony\n\nGilbert, Killingsworth, Eyre & Wilson (*Science* 2009): people predicted their own reaction to an\nexperience **better from one other person's report of it than from full information about it** — and\nboth they and outside judges expected the opposite. The disability paradox points the same way (people\nliving with serious disability rate their quality of life far above outsiders' expectations). So impact\ntestimony's main evidential job is **correcting other people's imagined versions of a position** — not\nestablishing causes.\n\n### 6c. The fallible-self-knowledge conviction, tested\n\nThe half that says people are **fallible about causes, counterfactuals and satisfiers** is strongly\nsupported (Nisbett & Wilson 1977; choice blindness, Johansson & Hall 2005; affective forecasting,\nGilbert & Wilson). The half that says people are **incorrigible about the felt experience itself** is\nchallenged by Schwitzgebel (*Philosophical Review* 2008): people make gross errors about their *current*\nexperience, at least its fine structure. Defenders reply that coarse present states (\"this hurts\") are\nreliable. A rewording that fits the evidence, **proposed and unruled because it touches a ratified\nconviction**: *authoritative about one's present refusal or endorsement (a speech act that makes its own\nfact), presumptive about one's present coarse state, fallible about the fine structure of experience and\nabout generalisations over time, and weak about causes, counterfactuals and satisfiers.* Choice\nblindness adds a design warning: a person confirming a machine rendering of their position is **weak\nevidence** that the rendering is right — which is why their own words should come first.\n\n### 6d. Impact testimony is three edges, not one\n\nThis dissolves the corpus contradiction:\n\n1. **Testimonial evidence for an empirical premise** (\"this policy costs people in position P in way\n   C\"), through position-to-know or witness testimony. It **moves strength**, subject to its critical\n   questions.\n2. **Input to a normative or weighing premise** (the values row; \"can this be justified *to* people in\n   P?\"). It moves strength only through the weighing scheme's own questions, never as proof of the\n   conclusion.\n3. **Registered testimony**, as reported speech on the non-propagating REPORTS edge: heard, available to\n   the first two roles, moving nothing. Forcing every testimony into an evidential role is itself a\n   documented harm (the South African TRC's four kinds of truth; hermeneutical injustice).\n\nProposed critical questions for positional testimony (from the strand, § 8.2): position (does it give\naccess to *this* claim?); kind of first-person claim (present state, retrospective, cause,\ncounterfactual, satisfier); representativeness; independence of corroborators; sincerity/consistency\nover time; rendering fidelity; scale shift; presentation stability; stake (a statement *against*\ninterest is strong evidence). The proposed testimony tiers survive with one amendment: \"positionally\ncorroborated\" should require **independent** corroboration, strongest from people in the position who\n*otherwise disagree* — the Community Notes logic.\n\n### 6e. Testimony at scale: model the process that produced it\n\nTruth commissions (Peru, Guatemala, Kosovo) treated testimony as indispensable *and* systematically\nbiased as a sample, and resolved it neither by discounting nor by counting but by **modelling the\nprocess that produced the testimony** — capture–recapture across overlapping independent lists — which\nin Peru reversed the picture raw testimony counts suggested about who was responsible for most deaths.\nThis is a method for the corpus's *coherent absence*: the overlap between independent lists is the\nevidence about what is missing.\n\n### 6f. Normative evidence\n\nNormative premises have **reasons**, not evidence in the empirical sense: considerations assessed by\ncoherence (reflective equilibrium), stability under variation in presentation (experimental\nphilosophy's order and framing effects, which expertise does not immunise against), resistance to\ndebunking (Street; Greene — and a debunking argument is an **undercutter**), and reciprocity. GRADE's\nEvidence-to-Decision frameworks show the practical form: type the premise (effect, value, equity,\nacceptability), grade each with the method fitting its type, and make the weighing a visible,\nattributable act. **Inductive risk** (Rudner; Douglas) is the principled bridge: stakes legitimately set\n*how much* evidence is enough, never stand in for evidence.\n\n### 6g. Expert judgment without an expert gate\n\nWeighting experts by **demonstrated calibration on checkable questions** beat equal weighting\nout-of-sample in 26 of 33 studies (Colson & Cooke 2017); in the IDEA protocol, status and experience\npredicted performance poorly. Independence matters: seeing others' estimates narrows the spread without\nimproving accuracy (Lorenz et al. 2011). Calibration can only be measured where ground truth arrives —\nthe oracle gradient again — so this is a route for the fact layer, never for evaluative claims.\n\n---\n\n## 7. Answers to the founder's open questions\n\n**\"What actually constitutes un-decomposable, more atomic evidence?\"** (1h in the voice map.) No\nproposition is atomic: every evidential claim can be challenged and decomposed (Sellars, Quine). But\ndecomposition does not bottom out in a *smaller claim*. It bottoms out in an **encounter** — the moment\nsomething in the world was observed, measured, recorded, said — which the graph can only *point at*,\nwith a locator (source, page, timestamp, span) and a provenance chain. An encounter cannot be decomposed\ninto claims; it can only be **re-encountered**: re-read, re-measured, replicated, re-observed. Claims\n*about* the encounter (the thermometer was calibrated, the witness was present, the transcription is\nfaithful) remain ordinary decomposable claims — lateral support, in Haack's terms. Four traditions\nconverge: Haack's clues, Schum's tangible evidence (\"the report is not the event\"), fact-checking's\nlocated span, and Bengio's factual observation versus communication act. **Proposed restatement of the\ncorpus line** (unruled): *No evidence is immune to challenge, and every evidential claim decomposes on\nreasonable challenge; what evidence finally rests on is not a smaller claim but a located encounter with\nthe world, which can only be re-examined or re-encountered.*\n\n**P2, the same evidence, the same probability.** Vindicated where Deliberus operates, qualified\nelsewhere: agreement on *what the evidence says* (the likelihood ratio, prior-free given a model) is\nachievable and person-independent; agreement on *what to believe* is a long-run ideal (merging-of-opinions\ntheorems), not a finite-evidence requirement. With thin evidence the rational state can be an\n**interval**, and people choosing different points inside it are not irrational (Walley). The FRI\nresidue — different standards for what counts as evidence — becomes representable if a side can mint\n*\"this kind of argument should carry these critical questions\"* as an attackable claim.\n\n**P3, the fear of fuzzy gooey subjectivity.** Answered by *named slots*: every divergence in conclusion\nfrom shared evidence must be traced to a prior, a standard of proof, a model or likelihood, or a piece of\nevidence, each of which becomes an attributable, attackable claim. Subjectivity is not banned; it is\n**located**.\n\n**P7, does Deliberus ever output a probability?** The research favours option (a) with a check: the\nbadge reports *the evidence term* — direction, weight and certainty — and never a posterior, which would\nimport a prior the platform has no right to choose. Where claims resolve against the world, the oracle\nbridge can still test whether strong readings turn out true more often. (This is a recommendation; P7\nstays open.)\n\n---\n\n## 8. The translation: how evidence should enter Deliberus\n\nAll of the following are **proposals**, each a ledger row in Arc Q. They are ordered from the\nconceptual frame to the cheapest structural changes to the larger ones.\n\n### 8a. The frame (Q1)\n\n**Deliberus answers \"what does the evidence say\", maps \"what should I believe\" and \"what should I do\",\nand answers neither of those for anyone.** The badge is the evidence term, not a posterior. This is the\npublic statement of the no-verdict design in its most defensible academic form (Royall), and it\nlocates every normative horizon kind in the second and third questions, where no amount of the first\nsettles things.\n\n### 8b. Weigh, don't count (Q3, Q5, Q9)\n\n- **An answer to a critical question moves its edge through its own standing**, not by existing. The\n  recursion the badge already uses for decomposition children should reach answers: an attacked or\n  refuted answer should weigh less than an unattacked one, and three one-word answers from one account\n  should move nothing beyond one.\n- **Each answer carries a direction and an ordinal strength tier** — supporting, moderate, strong, very\n  strong, on a log-odds scale (the ACMG/Tavtigian shape: points 1 / 2 / 4 / 8). The tier is **proposed**\n  (by the answer's content, by a machine pass, by the contributor) and is itself an attackable claim.\n  Default for a bare answer: the lowest tier. A kind of answer earns a higher default only when its\n  diagnosticity has been **measured** (the ClinGen move), never by fiat — which is how any methodology\n  default could become live without ratifying invented constants.\n- **Counting survives, honestly labelled, as coverage**: \"4 of 6 critical questions addressed\" is a\n  weight-like quantity and should be shown as such, never as strength.\n- **State the aggregation principle** (Amgoud–Ben-Naim): the proposal is quality precedence *within* a\n  channel of dependent inputs, and saturating accumulation *across* independent channels.\n- **Weight becomes Keynesian**: count only independent inputs; let it saturate as the scheme's question\n  space is covered; and add a **resilience** readout — how far the reading would move if its single\n  most influential input were removed or reversed — computable with the existing hinge machinery.\n- **Type attacks**: rebutting (competes with the claim), undercutting (caps the edge, pushes toward\n  \"uninformed\"), undermining (attacks a premise). Type critical questions: *assumption* (open means\n  weight withheld) versus *exception* (open means no cost until substantiated).\n\n### 8c. Independence as structure (Q4)\n\n- **A refuted reason and an objection can be one finding** (added 2026-09-29): refuting *the inspection\n  passed* asserts *the inspection failed*, which can also be filed as an objection. The no-double-use\n  rule below has to reach across a reason and its negation. Ruled 2026-09-30, for now: the refuted reason\n  drops out and the finding counts once, as the objection, which the page offers to file. On the proposed\n  starting-value path a failed reason would also subtract, and the mirror would then have to be matched and\n  counted once (founder: *\"one finding can be written twice\"*): `relevance-and-the-two-factors.md` § 7.\n\n- **Sources are nodes with reliability as attackable claims** — people, labs, instruments, datasets,\n  sponsors, publications. Every report, citation and answer links to its source.\n- **Channels**: inputs sharing an origin form one channel; channels combine as independent, items within\n  a channel do not multiply (STRING; Jøsang's cumulative vs averaging fusion; ACMG's no-double-use).\n- **Shared origin is proposed, never silently discounted**: a `SHARES_ORIGIN_WITH` proposal, with copy\n  detection keyed on **shared errors** (defeated positions, identical phrasings, identical citation\n  errors) rather than shared truths.\n- **Concordant reports from one source update the source's reliability**, not the claim's strength\n  (Bovens–Hartmann).\n- **An assertion is evidence, not a starting value** (added 2026-09-30, proposed): who asserted a claim\n  moves it through the link from its source, weighted by the source's reliability, while the claim's start\n  comes from its kind. Counted into the start, a source's reliability would never reach the claims the\n  reason supports: [starting-values-and-the-problem-of-the-prior.md](starting-values-and-the-problem-of-the-prior.md) § 6.2.\n\n### 8d. Anchors, scope and locators (Q2, Q11)\n\n- **A `scope` on empirical claims**: *particular* (\"in trial T, n = 240, the effect was X\") versus\n  *general* (\"X reduces Y\"). Particular findings **must** carry a source binding and a locator (page,\n  timestamp, span, retrieval date). General claims are supported by particulars through a *generalises*\n  inference carrying external-validity questions (same causal structure? support factors present?\n  population relevantly similar?).\n- **Provenance in two layers** (nanopublications): who found what, by which method, in which source\n  (assertion provenance), separate from which contributor or extractor version wrote this graph object,\n  when (record provenance). Supplies the missing edge and answer timestamps.\n- **Quote fidelity as its own attackable link** (the AIS test, *\"according to P, s\"*): a claim can be\n  wrong because the source is wrong or because the extraction mis-rendered the source (Knowledge-Based\n  Trust), and these need separate variables.\n- **A coherent-but-unanchored detector**: a cluster of claims that support each other strongly while\n  containing no anchored finding is flagged, not rewarded (Haack; the coherence impossibility results).\n\n### 8e. Bodies of evidence and certainty (Q6, Q7, Q8)\n\n- **Three nodes, all Claims**: a **study-result** claim (particular, anchored; its risk-of-bias\n  questions live here) → a **body-of-evidence** claim (the mother; inconsistency, imprecision,\n  publication bias and the three up-grades live here) → the **target** claim someone is actually arguing\n  (indirectness lives on the edge from body to target, because the same body can be direct for one\n  target and indirect for another). Study results are **corroborative** children of the body — never\n  necessary parts — which is also the reading the census gave its 70 not-a-part edges (⚠ 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)). A meta-analysis PDF decomposes into\n  exactly this shape: the founder's \"maximum granularity\" with a known skeleton.\n- **Certainty as a third badge axis**, a *reading* in GRADE's words (high / moderate / low / very low /\n  not yet assessed), always relative to a threshold claim, recomputed when domain claims change, each\n  step linked to the claim that justifies it — so \"why is this moderate?\" is a graph query. Never a\n  number. On surfaces, GRADE's plain verbs: \"probably\", \"may\", \"we are uncertain whether\".\n- **The methodology ladder**: retired as a strength input; retained, if at all, as a *labelled default\n  starting certainty* for intervention-effect claims, **per question type** (OCEBM 2011 as a seed) or\n  better from **design features** (was assignment controlled? did exposure precede outcome? is there a\n  comparison group?), each a claim. Expert opinion leaves the ladder: it is a scheme with its own\n  questions, and what experts report having *observed* is testimony.\n- **Mechanistic evidence as its own stream**, corroborative to the body claim, with its own appraisal\n  questions (EBM+; IARC's integration rules).\n- **Status claims with dates**: replication, retraction, correction, registered-report status, funding\n  source.\n\n### 8f. Absence as data (Q10)\n\nA **search claim**: what was searched for, where and how, when, and an estimate of the search's\nsensitivity. Without the sensitivity it carries a likelihood ratio near one and is shown as such. It\nfixes the stuck `lack_of_evidence` scheme (its closure question's answer carries the sensitivity), gives\nthe publication-bias domain its evidence, and gives coherent absence a measurable form: *no attacks\nfound* counts only in proportion to how exposed a claim was to people with reason to attack it.\n\n### 8g. Source reliability and report credibility (Q12)\n\nKeep the two axes apart (the Admiralty principle), and let a testimony or citation edge decompose on\ndemand into Schum's chain: **did the source say it** (locator, quote fidelity) → **did they believe it**\n(veracity) → **were they positioned to observe it** (objectivity, observational sensitivity) → **did it\noccur**. The first link is REPORTS-shaped, the third is position-to-know; the chain gives chain of\ncustody, quote verification and locators one pattern.\n\n### 8h. Testimony and normative evidence (Q13, Q14, Q17)\n\n- Impact testimony as **three edges** (§ 6d), with the positional critical questions and independent\n  corroboration.\n- **A representation instrument**: for a contested claim, which affected positions have contributed at\n  all — counting only, no model. The positional face of ingestion balance.\n- **Surrogation display**: beside an outsider's forecast of what a state is like, show testimony from\n  people in that state.\n- **The fallible-self-knowledge rewording** (§ 6c) — his call, since it touches a ratified line.\n- **Normative premises typed and graded by their own method**; debunking arguments filed as\n  undercutters; **inductive risk as an explicit bridge claim** when stakes are used to set how much\n  empirical evidence is enough.\n\n### 8i. Standards of proof and named slots (Q15, Q16)\n\n- **A standard of proof is an attributable claim attached to a question by whoever asks it** (\"for this\n  policy decision I need clear and convincing evidence\"). The graph can report whether the current\n  reading meets *that asker's* stated standard — a reading relative to a stated standard, never a\n  verdict.\n- **Named slots for divergence**: when two contributors reach different conclusions from shared\n  evidence, the surface asks where the difference lies — prior, standard, model, or evidence — and mints\n  the answer as an attributed claim.\n- **Evidential-standard claims** (\"arguments of this kind should face these questions\") make the FRI\n  residue representable instead of hidden in a hard-coded ladder.\n\n### 8j. An \"open\" status (Q18)\n\nAn epistemic status that says **the evidence does not decide this** — IARC's Group 3, OHAT's\n\"inadequate\", ACMG's \"uncertain significance\" — distinct from \"no effect\". It pairs with the gap found\nwhen a friend's twenty-four explicit open questions had nowhere faithful to live.\n\n### 8k. The discipline for importing all of this (Q19)\n\nThe corpus's rule for borrowed taxonomies applies with force: **a borrowed category earns its place by\nyielding a move.** GRADE's domains, Schum's credibility attributes, CERQual's components and ACMG's tiers\nall yield moves — critical questions to answer, parameters to argue about — so they enter as questions\nand parameters, each with a does-not-fit value, never as schema cells or as a ladder displayed as\nweight. The GO evidence-code history is the warning: a descriptive type, once displayed, is read as a\nquality rank.\n\n### 8l. Machine passes grading evidence (Q20, Q21)\n\n- **Pass-level calibration is higher-order evidence** about every output of that pass (Christensen): a\n  measured misfiling rate should discount the confidence of all the pass's unexamined outputs, not only\n  those shown wrong. This joins the machine-work watermark proposal.\n- **Consilience as a displayed structural fact**: support arriving from disjoint regions (source\n  clusters, concept clusters, methods with distinct failure modes) is shown, not added to strength.\n\n### 8m. From the professions of proof (Q22–Q26)\n\n- **Typed combinations (Q22).** When two or more evidential edges bear on one claim, a machine pass\n  *proposes* which Schum pattern they form (corroborative, convergent, cumulative, contradictory,\n  divergent), and anyone may contest it. Corroborative inputs saturate; convergent inputs add and may\n  be marked synergistic; contradictions are resolved on the credibility of each side and the display\n  never shows a headcount (§ 4b).\n- **Retraction propagates, relays do not launder (Q23).** When a source is shown to be fabricating or\n  unreliable, every claim whose support routes through it is flagged. A report that passed through an\n  intermediary (including a machine pass) keeps the original source's credibility questions open. Any\n  surface saying \"N items support this\" says from how many independent origins (§ 4a, move 4).\n- **Level-jump flag (Q24).** Evidential edges are typed by how far they jump along the hierarchy of\n  propositions (measured → occurred → caused → ought). A jump of more than one level is flagged as a\n  decomposition invitation, because the skipped steps carry the contestable assumptions (§ 4e).\n- **Badge wording (Q25).** Three rules from forensic reporting: say *support for X rather than Y* and\n  name the rival; near the neutral band never say *weak support* (say *not yet discriminating*, and\n  show the direction beside the weight); never put likelihood and confidence into one phrase, on any\n  surface, including agent-facing ones (§ 4e).\n- **Answer before seeing the badge (Q26).** Where feasible, a contributor records a critical-question\n  answer before seeing the claim's current reading or others' answers, and the record notes what they\n  had seen. This is cheap, protects the independence that Q4 and Q22 depend on, and gives the\n  ratification-tether work an instrument: blind answers against anchored ones (§ 4e).\n\nThe expert-evidence factors of the English Criminal Practice Directions (7.1.2 — among them *where in\nthe range of expert opinion this expert sits* and *whether the expert had complete information*) are\nthe natural critical questions for the expert-opinion scheme. They are filed under Q19's import\ndiscipline rather than given a row of their own.\n\n---\n\n## 9. What this does not change\n\n- **One node type.** Every proposal above uses Claims, edges and properties; the Source node already\n  exists. Nothing requires an Evidence node type.\n- **Propose, never assert.** Tiers, gradings, shared-origin links and certainty steps are proposals and\n  attackable claims; grading reliability data (κ as low as 0.06) say they must be.\n- **Bare judgments stay telemetry.** Calibrated, independent judgments *can* be evidence (Cooke;\n  Easwaran et al.), which gives the telemetry rule a sharper reason and states the condition under which\n  it could change: independence and calibration both measurable, which only the fact layer can supply.\n- **No verdicts.** Certainty and standards of proof are readings relative to stated thresholds.\n- **Decomposition stays maximal to separability**; \"reasonably challenged\" governs *on-demand*\n  decomposition requests, not the machine's own separability floor.\n\n---\n\n## 10. Where these proposals sit in the ruled order of work\n\nThe founder's ruled order puts the four reviews first. This document therefore **adds rows to the ruling\nledger (Arc Q) and changes no code**. The defects found in the code during the grounding pass — which are\nrepairs rather than rulings — are listed in `TODO.md` § Evidence layer — defects found 2026-09-26, so that\nwhichever session next touches the evidence layer sees them first.\n\n---\n\n## Cross-references\n\n[`what-counts-as-evidence/`](what-counts-as-evidence/README.md) (the dossier) ·\n[`scheme-bounded-decomposition-and-evidence-as-subgraph.md`](scheme-bounded-decomposition-and-evidence-as-subgraph.md)\n(evidence as subgraph; the founder's open question) ·\n[`evidence-division-and-the-foundation.md`](evidence-division-and-the-foundation.md) (holism, link never\ndivide) · [`long-form-sources-and-meta-analysis-weighing.md`](long-form-sources-and-meta-analysis-weighing.md)\n(per-claim methodology) · [`strength-layer-audit.md`](strength-layer-audit.md) ·\n[`weakest-link-arithmetic-and-the-merge-hunch.md`](weakest-link-arithmetic-and-the-merge-hunch.md)\n(min, Bayes, double counting) · [`what-human-judgment-is-for.md`](what-human-judgment-is-for.md) (impact\ntestimony) · [`parity-and-the-zone-interior.md`](parity-and-the-zone-interior.md) (fallible\nself-knowledge) · [`truth-graph-evidence-system.md`](truth-graph-evidence-system.md) (the April ladder) ·\n[`session26-open-rulings.md`](session26-open-rulings.md) (Arcs P and Q)\n"}