{"path":"research/what-counts-as-evidence/research-briefing.md","content":"# Briefing for academic research on EVIDENCE — what Deliberus already is\n\nDeliberus is a collaborative argumentation platform (\"Lean for all knowledge and decision-making\").\nOne graph; one node type, **Claim** (types: empirical / normative / definitional / value_premise;\nepistemic_status: observed / causal_hypothesis / prediction / counterfactual / feasibility /\nnormative_assertion; evidence_type: citation / quantitative / anecdotal / none). Typed edges:\nSUPPORTS, ATTACKS, QUALIFIES, REFRAMES, DECOMPOSES_INTO, DEFINES, REPORTS (reported speech,\nnon-propagating). Every inferential edge is classified into a **Walton argument scheme** (expert\nopinion, witness testimony, position to know, cause-to-effect, correlation-to-cause, sign,\nevidence-to-hypothesis, abductive, lack-of-evidence, ignorance, analogy, example, precedent,\nbias, argument from weighing, does_not_fit ...) whose **critical questions (CQs)** are rendered as\nclaims that can be answered. Strength is a **QBAF gradual semantics (Potyka's Quadratic Energy\nModel)**: answers to CQs on an incoming edge move the edge from 0.5; supports add, attacks\nsubtract. Since Aug 2026 the badge separates **direction** (which way evidence points) from\n**weight** (how many examined inputs) after **Keynes's weight of arguments**; an unexamined claim\ncarries no number at all. Decomposition (DECOMPOSES_INTO, \"mother claims\") aggregates children:\n*necessary* parts by **min** (Fréchet bound; weakest-link), *corroborative* by summed offsets\n(≈ Good's additive log-odds). \"Horizon kinds\" type where a line of reasoning currently stops\n(empirical = hands over to evidence; definitional; fittingness; structural; axiom choice;\npermissive zone) — \"the terminus enum is a map of the oracle gradient\": only the empirical leg\nreaches ground truth.\n\nPositions and principles already held (so research should deepen, test, or challenge them, not\nrestate them):\n- \"There is no atomic evidence, only claims whose decomposition nobody has demanded yet.\"\n  Evidence is a *position* in the graph (a claim at the tail of a support/attack edge), and\n  evidence properties (heterogeneity, sample size, methodology tier) should become claims.\n- The founder: a meta-analysis PDF \"should be decomposed as far as possible — maximum\n  granularity that is helpful to the rest of humanity\"; \"what actually constitutes\n  un-decomposable, more atomic evidence is an open question to me, scientifically, rationally,\n  philosophically.\"\n- Founder position (open): given the same evidence, probability should ideally not differ\n  between people; fear of \"fuzzy gooey subjectivity\" moving assessments without decomposition.\n- Machine proposes, never asserts on contested ground; bare judgments (votes, clicks) are\n  telemetry and never move strength; only contributions that argue for themselves do.\n- Fallible self-knowledge: first-person testimony is incorrigible about the felt experience,\n  fallible about causes, counterfactuals and what would satisfy it.\n- \"Impact testimony\" / positional evidence (what something costs from where one stands) is the\n  human contribution machines cannot supply; only people sharing a position can check it.\n- Confirmation holism (Duhem–Quine) acknowledged: evidence bears on claim-webs; \"link, never\n  divide\"; fair-share discounting for shared evidence (built, unadopted).\n- Fact layer first: the dispute structure around a factual claim is the non-commodity product;\n  Deliberus maps reasoning, does not issue verdicts.\n- A methodology ladder exists in code but is unratified and inert: meta-analysis 0.70,\n  systematic review 0.68, RCT 0.65, quasi-experimental 0.55, observational 0.45, case study 0.38,\n  anecdotal 0.32, expert opinion 0.30.\n\nKnown gaps in the current model (research should say what the literature offers for each):\nanswers to CQs are COUNTED not weighed; no source reliability separate from claim; no source\nindependence / shared-origin correlation; no likelihood ratio or diagnosticity; no sample-size or\nprecision weighting; no representation of \"we searched and found nothing\" as data; no chain of\ncustody / quote verification / page or timestamp locator; no edge or answer timestamps; no\nscheme for statistical/study evidence as such; no rebutting vs undercutting distinction typed;\nno epistemic status that can say \"this is open\"; methodology assigned per source not per claim;\nno mechanistic-evidence tier; no representation of publication bias, replication status,\nfunding conflict; no standard-of-proof concept.\n\nOutput expectations for every research agent: a thorough, well-organised report with concrete\ndefinitions, numbers, named frameworks, key authors and dates, and citations (URLs or full\nreferences). Distinguish settled consensus from contested positions. Flag anything you could not\nverify. Close with a section \"Most transferable to a claim-graph argumentation platform like the\none above\" — concrete, not generic.\n\n## Further facts from the grounding pass (2026-09-26)\n\n- Exact arithmetic: each answer to a CQ adds 0.15 to its polarity claim (cap 0.45); QEM is applied\n  at edge then at claim, so one fully answered six-CQ edge lifts a claim only to 0.58; a one-CQ\n  scheme (lack_of_evidence) can never leave the neutral band. Answers' content, source, and\n  independence are ignored; three one-word answers from one account = three RCTs.\n- Weight = count of non-neutral inputs; does not saturate; no notion of independence.\n- 35% of decomposition edges (70 of 197) are evidence relations 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)]; the weakest-part\n  (min) default then applies to what is really corroborating evidence (\"three studies are\n  witnesses, not parts\").\n- Double counting taxonomy already drafted (same claim twice; one evidence node by many paths;\n  parallel edges; partial overlap; distinct claims sharing a source) and reconciliation literature\n  already surveyed (Amgoud similarity-adjusted semantics, Denœux cautious rule, Jøsang cumulative vs\n  averaging fusion, robust variance estimation in meta-analysis, variety-of-evidence results —\n  \"a known shared source becomes a node, never a blind discount\").\n- Positional testimony: proposed tiers (unverified / self-attested with record / positionally\n  corroborated), \"weight rises with tier; nothing gated on tier\"; sole-witness label; testimony is\n  evidence for a premise, never a verdict; simulated perspectives may ask, never testify.\n- Bengio's Scientist AI separates \"factual observation\" from \"communication act\" (someone uttered\n  X); Deliberus's always-mint reported speech + non-propagating REPORTS edge is the same move.\n- Oracle gradient: only empirical claims can bottom out in a checkable fact; evaluative ones never.\n- Contradictions: whether impact testimony is \"weighed as evidence\"; three unmapped vocabularies of\n  evidence kind (evidence_type, April four kinds — testimonial/statistical/logical/observational —,\n  methodology tier); old docs call QEM \"Bayesian\" (later rejected); hinge epistemology (Pritchard)\n  says some commitments are not improved by evidential support.\n- The FRI adversarial collaboration found residual disagreement traced to \"different standards for\n  what counts as evidence\".\n"}