{"path":"research/strength-calculus-literature/computational-argumentation.md","content":"# The strength calculus in computational argumentation\n\n**Date**: 2026-09-28 · **Type**: literature review; every formula quoted below was read in the primary\nsource, and the code was checked at commit `76da237` · **State**: research, nothing here is ruled ·\n**Scope**: computational argumentation (gradual semantics, weights on links, structured argumentation\nand critical questions, machines and crowds as judges). Bayesian epistemology, confirmation theory,\nprobabilistic argumentation and informal logic are the sibling report in this directory; they appear\nhere only where a computational paper builds on them.\n\n**The question** (founder, 2026-09-28): *\"I want to make conscious choices about such crucial design\nand epistemic-calculus considerations and want to deeply consider everything academia has produced\nthat bears on these decisions generally and specifically about these tidbits.\"* The ruling it serves\n(ledger Q33): a link's effect on the claim it supports or attacks is the supporting claim's own\nstrength times the link's relevance. The eight open decisions, D1–D8, are laid out in § 3.\n\n**Reading key.** **DIRECT**: the publication answers or constrains the decision. **ADJACENT**: it\nbears on the decision through a neighbouring construct. **BACKGROUND**: context only. **UNVERIFIED**:\nnot checked against the source in this pass. Decisions a publication does not touch are not listed.\n\n---\n\n## 0. The QEM claim: half right\n\n`relevance-and-the-two-factors.md` § 0 item 2 says Potyka's QEM *\"computes a target's energy as each\nsupporter's own strength multiplied by the link's weight.\"*\n\n**The supporter's own strength: right.** From Potyka's tutorial (arXiv:1811.12787, Definition 2),\nwhich restates the KR 2018 model: *\"the energy at j is defined as E_j = Σ_{i∈Sup_j} s_i −\nΣ_{i∈Att_j} s_i\"*, with impact *\"h(x) = max{x, 0}² / (1 + max{x, 0}²)\"* and dynamics *\"ds_j/dt =\nw(j) − s_j + (1 − w(j)) · h(E_j) − w(j) · h(−E_j)\"*, where s_i is the supporter's current strength.\nEach supporter enters with its own strength, raw, not as an offset from 0.5.\n\n**The link's weight: wrong. QEM has no link weights.** Its framework is *\"A = (A, w, R, S), where A\nis a finite set of arguments, w : A → [0, 1] is a weight function and R and S are binary relations\non A called attack and support\"* (same source, Definition 1). Weights sit on arguments; attack and\nsupport are plain relations, so every link counts in full.\n\n**Where the product does exist.** Later work multiplies a supporter's (or attacker's) strength by a\nweight on the link:\n\n| Publication | Weight on the link | Combination, quoted | Semantics |\n|---|---|---|---|\n| Potyka, AAAI 2021 | real-valued; negative means attack | *\"α_a := Σ_{(b,a)∈E} w(b,a) · s_b\"*, then *\"s_a := φ_l(ln(β(a)/(1−β(a))) + α_a)\"* | logistic (\"MLP-based\") |\n| Amgoud & Doder, AAMAS 2019 | on attacks only, in [0, 1] | *\"v_a = f(σ(a), g(h(π((b1,a)), v_b1), …))\"*, one choice being *\"h_prod(x1, x2) = x1x2\"* | h-categorizer, max-based and others |\n| Yin, Potyka, Rago, Kampik & Toni, 2025 | on attacks and supports, *\"w : R− ∪ R+ → [0, 1]\"* | *\"A^σ_α = {σ(β) · w((β, α)) \\| β ∈ R−(α)}\"*, the same for supports | QE, DF-QuAD, Euler-based |\n| Egilmez, Martins & Leite, 2013 | on attacks, set by votes | product (in Amgoud & Doder's reconstruction) | shown to have several solutions |\n| Amgoud, Doder & Lagasquie-Schiex, 2024 / 2026 | on attacks, themselves attackable | an attack's own strength combined with its source's | higher-order; product excluded from their uniqueness result |\n\nYin et al. read the weight exactly as the ruling does. Changing base scores *\"can be seen as\nadjusting the apriori belief in an argument, whereas the latter* [changing edge weights]\n*corresponds to adjusting the relevance of an argument for a discussion.\"* **Their edge-weighted QE\nis the closest published counterpart of the ruling.** Three differences remain: Deliberus uses\noffsets from 0.5 with a floor and an explicit no-standing state where they use raw strengths;\nDeliberus's weight is argued, proposed or presumed, and moved by answers, where theirs is given;\nDeliberus computes one pass where they iterate to a fixed point.\n\nMossakowski & Neuhaus (2018) are credited by both Potyka 2021 and Yin et al. with edge-weighted\ngraphs, but the version read here (arXiv:1807.06685v2) allows only −1, 0 or 1 in its incidence\nmatrix. Whether a later version admits real weights is UNVERIFIED.\n\n**The sentence as it should read** (this report edits no other file): *\"QEM feeds each supporter's\nown strength into the target's energy and has no link weight; edge-weighted variants published\nsince (Potyka 2021; Yin et al. 2025, who read the weight as relevance) multiply that strength by a\nweight on the link.\"* The following sentence, *\"The March build replaced both with one number per\nlink\"*, should say it replaced the supporter's own strength: there was no link weight in QEM to\nreplace. `support-semantics-and-the-conjunction-problem.md` had already stated the divergence\ncorrectly.\n\n---\n\n## 1. What the corpus already has\n\n- `docs/research/qbaf-gradual-semantics-research.md`: the choice of QEM over DF-QuAD, Euler-based and\n  h-categorizer (two errors, § 4).\n- `docs/research/support-semantics-and-the-conjunction-problem.md`: necessary support (Nouioua &\n  Risch), and the correct statement that the build fed edge strengths where QEM feeds supporters'\n  degrees, as offsets.\n- `docs/research/weakest-link-arithmetic-and-the-merge-hunch.md`: minimum versus product as t-norms,\n  the padding defence, similarity-adjusted semantics (§ 9b), the argued requirement cap (§ 9d).\n- `docs/research/strength-layer-audit.md`: one pass versus fixed point, cycles measured, additive\n  energy.\n- `docs/research/the-irit-harvest.md`: Shapley and relation attributions, similarity between\n  arguments.\n- `docs/research/what-counts-as-evidence/formal-and-computational-models-of-evidence.md`: ASPIC+,\n  Carneades, Verheij, modular semantics, ArgRAG, the axioms that bear on counting (three errors,\n  § 4).\n- `docs/research/relevance-in-the-corpus-sweep.md`: six senses of relevance; lists the weighted-link\n  literature as missing (covered here).\n- `docs/research/relevance-and-the-two-factors.md`: the ruling and the build.\n  `docs/research/kialo-deep-dive.md`: Kialo's impact votes.\n- Code: `deliberus/relevance.py` (every named choice), `deliberus/graph/qbaf.py`\n  (`argument_edge_contribution`), `deliberus/graph/badge_inputs.py` (`attach_flow_inputs`, the\n  counting rule), `deliberus/extraction/schemes.py` (`CQ_BEARS_ON`: 118 templates, 81 connection,\n  26 premise, 11 rebuttal), `deliberus/relevance_judge.py`, `deliberus/evidential_weight.py`.\n\n---\n\n## 2. The publications\n\n### A. Gradual semantics: how strengths combine\n\n**1. Potyka, N. (2018). Continuous dynamical systems for weighted bipolar argumentation. KR 2018.**\nDefinitions quoted from Potyka, N. (2018). *A tutorial for weighted bipolar argumentation with\ncontinuous dynamical systems and the Java library Attractor*. arXiv:1811.12787.\nQEM (formula in § 0): the raw sum of supporters' strengths minus attackers', pushed through a\nquadratic impact away from the argument's base weight. Guaranteed to converge on acyclic graphs; on\ncyclic graphs *\"convergence in general remains an open question\"*, though *\"so far no diverging\nexample has been found and tests with large randomly generated bipolar graphs show that the\nstrength values converge in many cases.\"*\nGrades: D2 DIRECT (the raw convention the offset departs from) · D3 DIRECT (acyclic guarantee,\ncyclic open) · D1 ADJACENT (supporter's strength only, no second factor).\n\n**2. Potyka, N. (2019). Extending modular semantics for bipolar weighted argumentation. AAMAS 2019,\n1722–1730. arXiv:1809.07133.**\nEvery semantics splits into an aggregation function and an influence function. Iteration converges\nwhen the product of their Lipschitz constants stays below 1 (the contraction principle), but\n*\"convergence guarantees derived from the contraction principle are bought at the expense of\nopen-mindedness\"*; continuizing a semantics (computing the limit of a continuous-time process instead\nof repeated jumps) removed the divergences observed. QE appears as QE(κ), where κ sets how\nconservative the influence is near zero.\nGrades: D3 DIRECT (the general convergence test) · D1 ADJACENT (κ is a dial on how much small\ncontributions count).\n\n**3. Mossakowski, T. & Neuhaus, F. (2018). Modular semantics and characteristics for bipolar\nweighted argumentation graphs. arXiv:1807.06685.**\nSeparates the neutral value from the minimum. In a domain such as [−1, 1], *\"attacks or supports by\narguments with the neutral acceptability degree will have no effect\"*, and an argument below neutral\nis unacceptable. Offers *sum* (*\"attacks with negative degree are effectively supports and vice\nversa\"*) and *sum-pos* (*\"only parents with positive degrees are taken into account\"*), and leaves\nthe choice open: *\"One open question is whether the support of an unacceptable argument has no\neffect or should be counted against the supported argument.\"* Proves that *\"no sum-based semantics\ncan converge for all graphs.\"*\nGrades: D2 DIRECT (Deliberus's offset is this domain; its floor is sum-pos) · D3 DIRECT.\n\n**4. Amgoud, L. & Ben-Naim, J. (2018). Evaluation of arguments in weighted bipolar graphs. IJAR 99,\n39–55. doi:10.1016/j.ijar.2018.05.004.**\nEuler-based semantics, *\"s_i = 1 − (1 − w(i)²) / (1 + w(i) · exp(E_i))\"* (as quoted in Potyka's\ntutorial, equation 3), with the same energy as QEM; and the bipolar principles, among them\nNeutrality (worthless attackers and supporters change nothing), Weakening and Strengthening.\nGrades: D2 ADJACENT (neutrality at 0, where Deliberus puts it at 0.5) · D8 BACKGROUND.\n\n**5. Rago, A., Toni, F., Aurisicchio, M. & Baroni, P. (2016). Discontinuity-free decision support\nwith quantitative argumentation debates. KR 2016.**\nDF-QuAD: supporters aggregate by probabilistic sum, *\"aggΠ(A, S) = Π_{x∈A}(1 − x) − Π_{x∈S}(1 − x)\"*\n(as quoted by Yin et al.), so one supporter at strength 1 makes the others irrelevant (saturation).\nNo link weights.\nGrades: D8 ADJACENT (saturation is a built-in cap on accrual) · D1 BACKGROUND.\n\n**6. Baroni, P., Rago, A. & Toni, F. (2019). From fine-grained properties to broad principles for\ngradual argumentation: A principled spectrum. IJAR 105, 252–286. doi:10.1016/j.ijar.2018.11.019.**\nWith: (2018). How many properties do we need for gradual argumentation? AAAI 2018.\ndoi:10.1609/aaai.v32i1.11544.\nReduces 33 properties from the literature to two principles, balance and monotonicity, each strict\nor non-strict.\nGrades: D2 ADJACENT (the floor breaks strict monotonicity on the lower half of the scale) · D6\nADJACENT (monotonicity in the strength of evidence, which counting ignores).\n\n**7. Potyka, N. & Booth, R. (2024). Balancing open-mindedness and conservativeness in quantitative\nbipolar argumentation. KR 2024, 597–607. doi:10.24963/kr.2024/56.**\nThe cost of summing: *\"since αΣ is unbounded, even arguments with marginal strength, say 0.01\ninstead of 1, can significantly impact other arguments when they occur in large numbers.\"* Proposes\nbounds on how far parents may pull an argument (absolute and relative conservativeness); DF-QuAD\nalready satisfies relative conservativeness, QE does not.\nGrades: D8 DIRECT (padding by many weak links, and two published bounds) · D1 ADJACENT.\n\n**8. Munro, Y., Bloch, I. & Lesot, M.-J. (2026). Aggregative semantics for quantitative bipolar\nargumentation frameworks. arXiv:2603.06067.**\nAggregates attackers and supporters by separate functions before combining with the argument's own\nweight: *\"imposing φR = φS is a restrictive condition.\"*\nGrades: D8 ADJACENT (supporters and attackers may accrue by different rules) · D2 ADJACENT.\n\n### B. Weights on links: the second factor\n\n**9. Potyka, N. (2021). Interpreting neural networks as quantitative argumentation frameworks. AAAI\n35(7), 6463–6470. doi:10.1609/aaai.v35i7.16801. arXiv:2012.05738.**\nEdge-weighted QBAF, *\"w : E → R\"*, a negative weight meaning attack; the weighted sum and logistic\ninfluence in § 0. Linear time on acyclic graphs; on cyclic graphs it converges *\"if all arguments in\nQ have at most P parents, the weight of all edges is bounded from above by W and we have W · P <\n4\"*. Near the base score the influence is first-order, unlike QEM's quadratic. The update adds\nweighted strengths in log-odds, which is the shape of a naive-Bayes update with independent\nevidence (this report's reading; Potyka does not claim it).\nGrades: D1 DIRECT (product inside a sum: the published form of the ruling) · D3 DIRECT (smaller\nweights, stronger convergence guarantee).\n\n**10. Amgoud, L. & Doder, D. (2019). Gradual semantics accounting for varied-strength attacks.\nAAMAS 2019, 1270–1278. HAL hal-02325853.**\nAn evaluation method ⟨f, g, h⟩: h combines an attack's weight with its attacker's strength, g\naggregates the attacks, f applies them to the argument. The choices listed for h include\n*\"h_prod(x1, x2) = x1x2\"*, *\"h_prod,α(x1, x2) = x1^α x2\"*, the Hamacher product and *\"h_min(x1, x2)\n= min{x1, x2}\"*; the class H used to extend existing semantics *\"includes the functions presented\nin Table 1 except hmin\"*. New principle, Attack-sensitivity: *\"the stronger the weight of an\nattack, the greater its impact on the targeted argument\"*, guaranteed when h is strictly increasing\nin the weight, which min is not once the weight exceeds the attacker's strength. Neutrality covers\n*\"worthless attackers … nor worthless attacks\"*.\nGrades: D1 DIRECT (product versus minimum, with the principle that separates them) · D2 ADJACENT.\n\n**11. Yin, X., Potyka, N., Rago, A., Kampik, T. & Toni, F. (2025). Contestability in quantitative\nargumentation. arXiv:2507.11323 (venue UNVERIFIED).**\nWeights in [0, 1] on attacks and supports, combined with QE, DF-QuAD and Euler-based semantics by\nmultiplying strength by weight (§ 0); weights read as relevance. Asks how to change weights to reach\na desired strength, guided by gradient-based relation attributions (G-RAEs): the sensitivity of a\nclaim to each link's weight.\nGrades: D1 DIRECT (closest published counterpart of the ruling) · D7 DIRECT (relevance as the thing\npeople contest) · D3 ADJACENT (iterates to a fixed point; cites cyclic cases that fail to converge).\n\n**12. Amgoud, L., Doder, D. & Lagasquie-Schiex, M.-C. (2024). Higher-order argumentation\nframeworks: Principles and gradual semantics. IJCAI 2024, 3224–3231. doi:10.24963/ijcai.2024/357.**\nExtended as (2026). Gradual semantics for weighted higher-order argumentation frameworks. JAIR 86,\narticle 31. doi:10.1613/jair.1.21469.\nAttacks carry weights and can themselves be attacked, so an attack's strength is computed like an\nargument's (their Equality principle evaluates the two in the same way). Local Equality gives an\nattack and its source equal importance, which with a strictly increasing aggregation requires a\nsymmetric combination. Their uniqueness theorem admits the minimum and two modified averages;\nproduct is out: the condition *\"rules out certain functions such as h7(x, y) = x × y, which can be\nused by the semantics of (Amgoud and Doder 2019) but not in the context of wHO-AFs.\"* Values settle\n*\"in fewer than 20 iterations, even for large frameworks.\"*\nGrades: D1 DIRECT (once relevance is itself argued, their uniqueness result favours minimum over\nproduct) · D3 DIRECT · D7 DIRECT (argued relevance is a higher-order structure, and this is its\nsemantics).\n\n**13. Egilmez, S., Martins, J. & Leite, J. (2014). Extending social abstract argumentation with\nvotes on attacks. TAFA 2013, LNCS 8306, 16–31. doi:10.1007/978-3-642-54373-9_2.**\nWith: Amgoud, L., Bonzon, E., Correia, M., Cruz, J., Delobelle, J., Konieczny, S., Leite, J.,\nMartin, A., Maudet, N. & Vesic, S. (2017). A note on the uniqueness of models in social abstract\nargumentation. arXiv:1705.03381.\nAttack weights come from votes on the attacks and combine with the attacker's strength by product\n(evaluation method ⟨f_comp, g_psum, h_prod⟩ in Amgoud & Doder's reconstruction; the original was not\nread in this pass). Per Amgoud & Doder, the resulting equations *\"may have several solutions\"*.\nGrades: D7 DIRECT (the published attempt to set link weights by voting, and its failure mode) · D3\nADJACENT.\n\n**14. Leite, J. & Martins, J. (2011). Social abstract argumentation. IJCAI 2011.**\nVotes on arguments set their base strength. Known here from secondary descriptions only.\nGrades: D7 BACKGROUND.\n\n**15. Amgoud, L., Ben-Naim, J., Doder, D. & Vesic, S. (2017). Acceptability semantics for weighted\nargumentation frameworks. IJCAI 2017, 56–62. doi:10.24963/ijcai.2017/9.**\nWeighted h-categorizer (all attackers summed), weighted max-based (the strongest attacker only),\ncard-based (the number of attackers first): the quality-versus-quantity choice as three semantics.\nGrades: D8 ADJACENT.\n\n**16. Amgoud, L., Bonzon, E., Delobelle, J., Doder, D., Konieczny, S. & Maudet, N. (2018). Gradual\nsemantics accounting for similarity between arguments. KR 2018. HAL hal-01872590.**\n*\"If x1 and x2 are fully similar, only one of them is considered in the evaluation of a since the\nother is redundant. However, if they are partially similar, the redundant part should not be counted\ntwice.\"* Principle: Redundancy Freeness.\nGrades: D8 DIRECT (the published remedy for accrual inflated by near-duplicates).\n\n**17. Dunne, P. E., Hunter, A., McBurney, P., Parsons, S. & Wooldridge, M. (2011). Weighted argument\nsystems: Basic definitions, algorithms, and complexity results. AIJ 175(2), 457–486.\ndoi:10.1016/j.artint.2010.09.005.**\nAttack weights with an inconsistency budget, for extension semantics. Not read in this pass.\nGrades: BACKGROUND throughout.\n\n### C. Structured argumentation: what a critical question attacks, and who must prove it\n\n**18. Prakken, H. (2010). An abstract framework for argumentation with structured arguments.\nArgument & Computation 1(2), 93–124. doi:10.1080/19462160903564592.** With: Modgil, S. & Prakken,\nH. (2014). The ASPIC+ framework for structured argumentation: a tutorial. Argument & Computation\n5(1), 31–62. doi:10.1080/19462166.2013.869766.\nThree attacks: undermining (on a premise), undercutting (on the defeasible inference), rebutting (on\nthe conclusion). The basis of `CQ_BEARS_ON`'s premise, connection and rebuttal.\nGrades: D5 DIRECT.\n\n**19. Verheij, B. (2003). Dialectical argumentation with argumentation schemes: An approach to legal\nlogic. Artificial Intelligence and Law 11, 167–195. doi:10.1023/B:ARTI.0000046008.49443.36.**\nCritical questions play four roles, not three: *\"criticizing a scheme's premises\"*; pointing *\"to\nexceptional situations in which the scheme should not be used\"*; *\"conditions for a scheme's use\"*;\npointing *\"to other possible arguments relevant for a scheme's conclusion.\"* His example of the\nthird role is a relevance question, Walton's ad hominem CQ2: *\"Is the allegation that a is a bad\nperson relevant to judging a's argument α?\"* The first and fourth kinds *\"are not specific for a\nscheme.\"*\nGrades: D5 DIRECT (the connection bin merges roles two and three) · D4 ADJACENT (exceptions and\nconditions differ in who must show them).\n\n**20. Gordon, T. F., Prakken, H. & Walton, D. (2007). The Carneades model of argument and burden of\nproof. AIJ 171(10–15), 875–896. doi:10.1016/j.artint.2007.04.010.**\n*\"Carneades models critical questions as additional premises of an argument, with a different type\nof premise, called assumptions and exceptions, for each kind of question.\"* If the questioner should\nbear the burden of production, the question is an exception; if the arguer should, an assumption.\nAssumptions hold until questioned, exceptions bite only once shown, ordinary premises hold only once\nsupported, and the scheme *\"allows answers to be assumed for critical questions which have yet to be\nasked.\"* Worked example, expert opinion: credibility and being based on evidence are assumptions;\nreliability and consistency with other experts are exceptions; *\"E is an expert in the subject\ndomain S containing the proposition A\"* is an ordinary premise.\nGrades: D4 DIRECT (presumption and burden set per question) · D5 DIRECT (a burden axis beside the\nattack axis) · D6 ADJACENT (proof standards decide when a premise holds).\n\n**21. Prakken, H. (2005). A study of accrual of arguments, with applications to evidential\nreasoning. ICAIL 2005, 85–94. doi:10.1145/1165485.1165500.** Follow-up: (2019). Modelling accrual\nof arguments in ASPIC+. ICAIL 2019, 103–112. doi:10.1145/3322640.3326703.\nThree principles for any account of accrual: an accrual is sometimes weaker than its elements,\nbecause reasons are not always independent (witnesses *\"more likely to confirm each other's\nstatements when these statements are false than when they are true\"*); an accrual makes its\nelements inapplicable; and *\"when an individual reason or argument turns out to be flawed, it does\nnot take part in the accrual.\"* Hence *\"in general the strength of an accrual cannot be calculated\nfrom the strengths of its elements.\"*\nGrades: D2 DIRECT (a refuted supporter drops out; it does not reverse) · D8 DIRECT (accrual is not a\nsum; dependence) · D6 ADJACENT (flawed evidence must not be counted).\n\n**22. Prakken, H. (2021). Philosophical reflections on argument strength and gradual acceptability.\nECSQARU 2021, LNCS 12897, 144–158. doi:10.1007/978-3-030-86772-0_11.**\nSeparates inferential, dialectical and rhetorical strength, and warns that *\"trade-offs between\nattacks and supports, or even to regard supporters as strengthening the supported argument, may make\nsense in the statement interpretation but may not make sense in the inferential-structure\ninterpretation.\"*\nGrades: D1 ADJACENT · D6 ADJACENT (Deliberus is a hybrid, its claims statements and its links\ninferences, so the calculus should say which reading each factor follows).\n\n**23. Prakken, H. (2024). An abstract and structured account of dialectical argument strength. AIJ\n335, 104193. doi:10.1016/j.artint.2024.104193.**\nDialectical strength as the number of ways an argument can be successfully attacked in expansions of\nthe framework; most strength principles in the literature fail for it.\nGrades: D6 ADJACENT (a principled sense of counting: routes of attack, not items of evidence).\n\n**24. Cohen, A., Parsons, S., Sklar, E. I. & McBurney, P. (2018). A characterization of types of\nsupport between structured arguments and their relationship with support in abstract\nargumentation. IJAR 94. doi:10.1016/j.ijar.2017.12.008.**\nFour kinds of support between structured arguments (subargument, premise, conclusion, intermediate),\ncited by Prakken 2021 as the reason \"support\" has no single meaning.\nGrades: D5 ADJACENT.\n\n**25. Rago, A., Vasileiou, S. L., Tran, S., Toni, F. & Yeoh, W. (2025). A methodology for\nincompleteness-tolerant and modular gradual semantics for argumentative statement graphs. KR 2025,\n500–511. doi:10.24963/kr.2025/49.**\nEach premise literal is first evaluated dialectically by any QBAF semantics, then the premises are\ncombined by a *\"premise aggregation function\"*; *\"arguments with partially specified premises can\nplay a meaningful role\"*, and a statement is not rejected merely because its support is missing.\nGrades: D5 DIRECT (premise strength computed once, then combined: the acceptability factor done\nstructurally) · D4 ADJACENT.\n\n**26. Jedwabny, M., Croitoru, M. & Bisquert, P. (2020). Gradual semantics for logic-based bipolar\ngraphs using T-(co)norms. ECAI 2020. doi:10.3233/FAIA200166.**\nCombines an argument's components by t-norms and t-conorms. Known here from Rago et al. 2025's\ndescription.\nGrades: D1 ADJACENT.\n\n**27. Rapberger, A., Russo, F., Rago, A. & Toni, F. (2025). On gradual semantics for\nassumption-based argumentation. KR 2025, 512–522. doi:10.24963/kr.2025/50.**\nStrengths for assumptions rather than whole arguments. On synthetic frameworks their method converged\nin 90% of scenarios, a QBAF baseline in about 70%.\nGrades: D3 ADJACENT (convergence is not a given in practice) · D5 ADJACENT.\n\n**28. Thieyre, J., Al Anaissy, C., Beynier, A., Destercke, S., Maudet, N. & Vesic, S. (2025).\nUncertainty in quantitative bipolar argumentation frameworks. ECAI 2025, FAIA 413, 1679–1686.\ndoi:10.3233/FAIA250995.**\nBase weights as intervals derived from sparse votes (imprecise Dirichlet model), tested on Kialo\ndebates; restricted to acyclic graphs because *\"most of the online debate platforms are designed in\na way that constrain graphs to be acyclic.\"*\nGrades: D4 ADJACENT (ignorance kept as an interval, a cousin of Deliberus's no-number state) · D7\nADJACENT · D3 ADJACENT.\n\n### D. Explaining and contesting a strength\n\n**29. Kampik, T., Potyka, N., Yin, X., Čyras, K. & Toni, F. (2024). Contribution functions for\nquantitative bipolar argumentation graphs: A principle-based analysis. IJAR 173, 109255.\ndoi:10.1016/j.ijar.2024.109255.**\nRemoval-based, Shapley-based and gradient-based measures of how much one argument contributed to\nanother's strength.\nGrades: D7 ADJACENT (a contribution mixes both factors and depends on the whole graph, so it cannot\nserve as a relevance input without circularity).\n\n**30. Yin, X., Potyka, N. & Toni, F. (2024). Explaining arguments' strength: Unveiling the role of\nattacks and supports. IJCAI 2024, 3622–3630. doi:10.24963/ijcai.2024/401.**\nShapley attribution over links (already in `the-irit-harvest.md`).\nGrades: D7 ADJACENT.\n\n**31. Leofante, F. et al. (2024). Contestable AI needs computational argumentation. KR 2024,\n888–896. doi:10.24963/kr.2024/83.** Position paper.\nGrades: BACKGROUND.\n\n### E. Machines and crowds as judges of relevance\n\n**32. Freedman, G., Dejl, A., Gorur, D., Yin, X., Rago, A. & Toni, F. (2025). Argumentative large\nlanguage models for explainable and contestable claim verification. AAAI 39(14), 14930–14939.\ndoi:10.1609/aaai.v39i14.33637.**\nA language model generates supporting and attacking arguments and rates their base scores zero-shot;\nDF-QuAD computes the claim. Weights on arguments, none on links.\nGrades: D7 ADJACENT (machine-set factors, on the acceptability side only).\n\n**33. Zhu, Y., Potyka, N., Hernández, D. et al. (2025). ArgRAG: Explainable retrieval augmented\ngeneration using quantitative bipolar argumentation. NeSy 2025. arXiv:2508.20131.**\nQE over retrieved evidence. *\"Initializing base scores without prior knowledge (NP) outperforms\nusing retriever scores (RS), indicating that retriever confidence may not align with the actual\nrelevance or trustworthiness of retrieved content.\"* The choice of semantics (Euler, QE, DF-QuAD)\nmade only minor differences.\nGrades: D7 DIRECT (a similarity score used as a relevance prior did worse than no prior).\n\n**34. Gorur, D., Rago, A. & Toni, F. (2025). Can large language models perform relation-based\nargument mining? COLING 2025, 8518–8534.**\nBest micro-F1 82 for support versus attack; 68 once \"no relation\" is a third answer.\nGrades: D7 DIRECT (the relation-or-not decision is where the machine is weakest, and the relevance\njudge's gate is that decision).\n\n**35. Wachsmuth, H., Naderi, N., Hou, Y., Bilu, Y., Prabhakaran, V., Alberdingk Thijm, T., Hirst, G.\n& Stein, B. (2017). Computational argumentation quality assessment in natural language. EACL 2017,\n176–187.**\nDefines local relevance (*\"A premise of an argument is relevant if it contributes to the acceptance\nor rejection of the argument's conclusion\"*) apart from local acceptability and sufficiency.\nKrippendorff's α for local relevance: .47, for the most agreeing pair of annotators.\nGrades: D7 DIRECT (the human agreement ceiling for relevance) · D1 ADJACENT (relevance and\nacceptability as separate dimensions, which is the ruling's split).\n\n**36. Mirzakhmedova, Gohsen, Chang & Stein (2024). Are large language models reliable argument\nquality annotators? RATIO 2024, LNCS, 129–146. doi:10.1007/978-3-031-63536-6_8.**\nLocal relevance: α among human annotators 0.41 (experts) and 0.36 (novices); among repeated model\nruns 0.59–0.98; agreement between models and humans *\"moderate\"*.\nGrades: D7 DIRECT (a model agrees with itself far more than with people: consistency is not\nvalidity).\n\n**37. Durmus, E., Ladhak, F. & Cardie, C. (2019). The role of pragmatic and discourse context in\ndetermining argument impact. EMNLP-IJCNLP 2019, 5668–5678. doi:10.18653/v1/D19-1568.** With Kialo's\nown guidance (Kialo Edu help, \"About Voting\").\n47,219 Kialo claims with five-level impact votes, cast with the full argument in view. Kialo defines\nthe vote: *\"A claim's impact is the result of its veracity and relevance.\"*\nGrades: D7 DIRECT (the largest crowd dataset of link-like judgements measures both factors\ntogether, never relevance alone).\n\n---\n\n## 3. The design space, decision by decision\n\nFor each decision: what is built, the options and who backs them, which principles each keeps or\nbreaks, and the failure to watch.\n\n### D1. How the two factors combine\n\n**Built**: product, relevance × (acceptability − 0.5) (`relevance.link_contribution`).\n\n| Option | Published in | Keeps | Breaks |\n|---|---|---|---|\n| Product | Potyka 2021; Amgoud & Doder 2019; Yin et al. 2025; Egilmez et al. 2013 | both factors always matter; a fully relevant link passes the supporter on exactly (Amgoud & Doder require h(1, x) = x) | the uniqueness guarantee once the weight is itself argued (Amgoud, Doder & Lagasquie-Schiex 2026); a votes-on-attacks product can have several solutions. Both bite only for a fixed point on cycles. |\n| Minimum | Amgoud & Doder 2019; the 2026 uniqueness family | symmetric, idempotent, unique once argued | attack-sensitivity: once relevance exceeds acceptability, arguing relevance higher changes nothing, and the reverse |\n| Other t-norms (Hamacher, Łukasiewicz) | Amgoud & Doder 2019; Jedwabny et al. 2020 | between the two | — |\n| Averages that are zero when either factor is zero | the 2026 uniqueness family | unique once argued | a half-relevant link from a fully accepted claim passes on 0.75, an odd meaning for \"relevance\" |\n\nTwo things the literature adds that the ruling did not weigh:\n\n1. **A minimum needs its two scales made comparable; a product does not.** The product is taken on an\n   offset in [−0.5, 0.5]. A minimum would compare relevance in [0, 1] with that offset, so it would\n   need a stated mapping, for example min(relevance, 2 × offset). No publication covers offsets.\n2. **The product meets QEM's quadratic impact.** Near neutral, h(E) is about E², so one link moves\n   its target by about 0.5 × relevance² × (acceptability − 0.5)². On the built constants: a link the\n   judge would call *settles* (relevance 1) from a fully accepted claim moves a neutral claim to\n   0.60; at the presumed relevance 0.5, to 0.53; a *moderate* link from a claim at 0.7 moves it by\n   0.005; four fully accepted *settling* links are needed to reach 0.9. In the published\n   edge-weighted QE, a single supporter nobody has examined, at base 0.5 on a full-weight link,\n   already moves its target to 0.60. This is also why the 12 live links that start from a claim with\n   standing move nothing measurable: small offsets get squared. The published alternatives respond\n   in proportion near the base: the logistic influence (Potyka 2021), DF-QuAD's linear influence, or\n   QE with a smaller κ (Potyka 2019).\n\n### D2. Raw or offset; does an unexamined supporter add; does a weakened one subtract\n\n**Built**: offset from 0.5; a supporter nobody has examined has no standing and adds nothing; a\nsupporter below 0.5 adds nothing (`WEAK_SUPPORTER_SUBTRACTS = False`).\n\n- **The mainstream semantics** (QEM, DF-QuAD, Euler-based, logistic, edge-weighted QE) work in [0, 1]\n  with the neutral point at the minimum, 0. An argument nobody has examined carries its base score\n  and passes it on; a supporter can only raise; a supporter at 0 is inert (Neutrality). Deliberus\n  departs twice: its neutral point is 0.5, and an unexamined claim carries no number at all.\n- **The offset has a published home**: Mossakowski & Neuhaus's domain, where neutral and minimum\n  differ. The built floor is their *sum-pos*; switching `WEAK_SUPPORTER_SUBTRACTS` on gives their\n  *sum*, under which *\"attacks with negative degree are effectively supports and vice versa.\"* They\n  leave the choice open.\n- **The argumentation tradition decides it through accrual**: a flawed reason *\"does not take part in\n  the accrual\"* (Prakken 2005). It drops out; it does not reverse. That backs the built choice.\n  Probability backs the opposite (sibling report). The adversarial reason in `relevance.py`, a\n  deliberately weak supporter refuted to drag a claim down, is not analysed in any computational\n  paper found in this pass.\n- **The no-standing rule** has two published cousins, neither numeric: Carneades's ordinary premise,\n  which holds only once supported, and Thieyre et al.'s interval of ignorance.\n\nPrinciples: the floor is monotone but not strictly (weakening a supporter from 0.45 to 0.05 changes\nnothing), so strict monotonicity (Baroni et al.) fails on the lower half; subtraction restores it\nbut turns a refuted supporter into an attacker.\n\nFailures to watch. **A refuted supporter reads as unexamined.** Its link's displayed strength lands\nexactly on 0.5, and `evidential_weight.count_evidential_inputs` counts only inputs away from 0.5, so\n\"somebody examined this supporter and it failed\" adds no weight, and a claim whose supporters all\nfailed can revert to UNEXAMINED. **Standing has to originate somewhere**\n(`relevance-and-the-two-factors.md` § 6): in the literature it always originates in a base score,\nset by votes (Leite & Martins), a model's rating (Freedman et al.), retrieval similarity (Zhu et al.,\nwhere it did worse than a flat prior), or intervals from sparse votes (Thieyre et al.).\n\n### D3. Fixed point or one pass\n\n**Built**: one recursive pass, memoized; a claim already on the current path is read as having no\nstanding; depth limit 8 (`badge_inputs.attach_flow_inputs`, `compute_badge_recursive`).\n\n- **Acyclic graphs**: every semantics here converges, and one pass in topological order equals the\n  fixed point (Potyka 2018; Rapberger et al.). Platform-oriented work often assumes acyclic graphs\n  outright (Thieyre et al.).\n- **Cyclic graphs**: QEM has no general guarantee; *\"no sum-based semantics can converge for all\n  graphs\"* (Mossakowski & Neuhaus); a contraction is sufficient but costs open-mindedness (Potyka\n  2019), and continuizing removed the divergences seen; the logistic semantics converges when\n  W · P < 4 (Potyka 2021); the higher-order uniqueness theorem excludes product; on synthetic cyclic\n  frameworks, 70–90% of runs converge (Rapberger et al.).\n- **Weights help.** The Lipschitz constant of a weighted sum is the sum of its weights. Derived here\n  from Potyka 2019's contraction principle, not a published theorem: QEM's steepest slope around a\n  neutral base is about 0.32, so iterating the argument channel converges to one answer whenever no\n  claim's incoming relevance adds up to more than about 3 (each argued relevance claim adds up to\n  0.5 to that sum). The weakest-part ceiling on parts has slope 1, so it needs the part graph to\n  stay acyclic, as it currently is.\n\n**A failure found in the code, with no counterpart in the literature: the result can depend on where\nthe walk starts.** The memo is keyed by claim alone, yet a claim's value is computed with whatever is\non the current path cut off and with supporters beyond the depth limit dropped. Where several claims\nshare one memo (the hinge, the truth-graph layer, the map's tag endpoint in `api.py`), a claim first\nreached through a cycle or near the depth limit keeps its truncated value for every later caller, and\nwhich claim comes first is the order the database returns rows in. Latent\ntoday (no two-cycles among live argument links; the largest upstream region is 41 claims). Published\nsemantics avoid it by being fixed points, or by being defined only on acyclic graphs.\n\n### D4. What an unexamined link presumes, and who carries the burden\n\n**Built**: relevance presumed 0.5 when the source asserted the link, a person wrote it or a split cut\nit; 0 for a machine's guess between other people's claims; an argued relevance claim or a confident\nmachine proposal replaces the presumption; unanswered questions move nothing.\n\n- **The textbook presumes full relevance**: without link weights, every link counts at weight 1 (QEM,\n  DF-QuAD, Euler-based). The edge-weighted papers read here take weights as given; none sets a\n  default.\n- **Carneades presumes per question, by burden.** An assumption holds until someone asks; an\n  exception bites only once someone shows it; an ordinary premise holds only once supported.\n  Verheij's conditions of use and exceptions are the same split.\n- **In Carneades's terms, the built design treats relevance as an assumption with a fixed partial\n  weight and acceptability as an ordinary premise.** The pairing is coherent. What it cannot\n  express: asking an assumption-type connection question changes nothing until someone answers it\n  with evidence, where Carneades drops the presumption at the moment of asking; and an exception-type\n  question needs the same evidence as any other.\n- **Presuming by who wrote the link has no precedent found.** The nearest is trust-based semantics,\n  which weights arguments (not links) by their source (da Costa Pereira, Tettamanzi & Villata, IJCAI\n  2011; not read in this pass).\n\n### D5. Sorting the critical questions\n\n**Built**: premise (moves acceptability), connection (moves relevance), rebuttal (acts on the\ntarget); 118 templates sorted by hand, 26 / 81 / 11 (`schemes.CQ_BEARS_ON`).\n\n- ASPIC+ backs the three bins. Verheij splits the middle one into exceptions (the scheme should not be\n  used here) and conditions of use (the scheme applies only if this holds, his example being a\n  relevance question), and Carneades gives those two opposite burdens. The connection bin therefore\n  merges two kinds of question the literature keeps apart, and that is exactly the axis D4 needs.\n- **Divergences from Carneades's worked expert-opinion example**, each defensible and each a shift of\n  burden: whether the expert's field covers the claim is an ordinary premise there and a connection\n  question here, so it is presumed rather than proven; reliability is an exception there and a\n  connection question here; consistency with other experts is an exception there and a rebuttal\n  here.\n- **Premise answers act on one link.** `argument_edge_contribution` moves acceptability *\"as this link\n  uses it\"*, so an answer that the supporting claim does not hold lowers it for this link only. In\n  ASPIC+ an undermining attack hits the premise wherever it is used, and Rago et al. 2025 likewise\n  evaluate each premise once, then combine.\n- Verheij's fourth role, other arguments for or against the conclusion, is *\"not specific for a\n  scheme\"*: such answers are independent arguments, which suggests filing them as their own links\n  rather than as a term inside this one.\n\n### D6. Answers weighed by their standing, or counted\n\n**Built**: a question's pole moves 0.15 per supporting evidence item, capped at 0.45\n(`badge_inputs.EVIDENCE_SHIFT`); the items' own standing is never read. Ledger Q3's direction is\nruled toward weighing; its arithmetic is open.\n\nThe literature is uniform: an attacker's or supporter's influence is its computed strength, and when\ncritical questions are modelled as arguments (ASPIC+ undercutters and underminers, Carneades\npremises, the generated arguments of Freedman et al.) they are weighed like any other argument. The\ncounting rule breaks three published principles: Neutrality (a refuted evidence item still adds\n0.15), strict monotonicity in the evidence's strength (Baroni et al.; Reinforcement in Amgoud &\nDoder), and Prakken's rule that flawed arguments do not accrue. The one principled sense of counting\nfound is Prakken 2024's dialectical strength, which counts ways to attack, not items of evidence.\nOnce answers are weighed, how several answers on one pole add up becomes a D8 question.\n\n### D7. Who sets relevance\n\n**Built**: argued (a relevance claim whose computed strength becomes the reading) outranks a\nconfident machine proposal from a second model family, which outranks the presumption; never voted.\n\n| Source of the weight | Published | What it teaches |\n|---|---|---|\n| Given, edited to contest | Yin et al. 2025 | gradient attributions price each weight's leverage: the natural generalisation of the hinge to \"argue this relevance\" |\n| Argued: the link's strength computed from arguments about it | Amgoud, Doder & Lagasquie-Schiex 2024 / 2026 | the published semantics of argued relevance; it covers attacks on links, not supports |\n| Votes on the link | Egilmez et al. 2013 | the product form gave several solutions |\n| Votes on the claim, sparse | Leite & Martins 2011; Thieyre et al. 2025 | sparse votes need intervals, not points |\n| A model rating arguments | Freedman et al. 2025 | used for base scores, never links |\n| Retrieval similarity | Zhu et al. 2025 | worse than a flat prior |\n| A model classifying the relation | Gorur et al. 2025 | the no-relation decision is the weak spot (micro-F1 82 falls to 68) |\n| People rating relevance | Wachsmuth et al. 2017; Mirzakhmedova et al. 2024 | α .36–.47: people disagree about relevance about as much as about any dimension of argument quality |\n| Crowd impact votes | Durmus et al. 2019; Kialo | measure veracity and relevance together |\n\nArguing relevance through a claim whose strength is computed like any other claim's is the Equality\nprinciple of the higher-order line. The literature offers no validity check better than human\nagreement, and a model's agreement with itself is not evidence of validity (Mirzakhmedova et al.).\n\n### D8. Accrual, padding and Goodhart\n\n**Built**: contributions summed across links, then QEM. Defences already in place: the floor (only\nexamined claims above neutral count), zero presumed relevance for machine-guessed links, the\nrelevance judge's gate, the weakest-part ceiling on parts.\n\n| Aggregation | Published in | Padding | Accrual |\n|---|---|---|---|\n| Sum (QEM, Euler-based, logistic, h-categorizer; built) | Potyka 2018; Amgoud et al. 2017 | many weak links add up without bound (Potyka & Booth) | full |\n| Strongest only | weighted max-based (Amgoud et al. 2017); TAC (Potyka & Booth) | immune | none |\n| Probabilistic sum | DF-QuAD (Rago et al. 2016) | saturates | diminishing |\n| Bounded by the parents | Potyka & Booth 2024 | bounded | partial |\n| Discounted for similarity | Amgoud et al. 2018 | near-duplicates count once | full for distinct reasons |\n| A separate rule per polarity | Munro et al. 2026 | set per side | set per side |\n\nPrakken's first accrual principle names what no sum handles: reasons that are not independent accrue\nto less than their sum. The hole left in the built design: a person can still add many modestly\nrelevant supporters, each examined just above neutral, and the sum rewards their number. Shared\norigin between supporters is not represented on the argument channel; the similarity finding in\n`weakest-link-arithmetic-and-the-merge-hunch.md` § 9b is not wired, and the fair-share discount in\n`deliberus/evidence_routing.py` (built, not adopted) covers one fact routed to several parts, not\nseveral supporters with one source.\n\n---\n\n## 4. Where the corpus is wrong or missing something\n\n**Wrong** (none of these files was edited):\n\n1. `docs/research/relevance-and-the-two-factors.md` § 0 item 2: QEM has no link weight; § 0 above\n   gives the corrected sentences and answers the doc's own \"under verification\" note.\n2. `docs/research/qbaf-gradual-semantics-research.md` line 30: the Euler-based formula is\n   `1 − (1 − w²)/(1 + w·e^E)`, not `1 − (1−w)²/(1+w·exp(E))`; the version in the doc does not return\n   the base weight when E = 0. Line 45: *\"Convergence empirically validated on 3,000+ random graphs\"*\n   is stronger than the primary source, which says strengths *\"converge in many cases\"* and calls\n   general cyclic convergence *\"an open question\"*; the 3,000 figure is UNVERIFIED. *\"Satisfies ALL\n   desirable properties\"* is overtaken by Potyka & Booth 2024: QE fails relative conservativeness,\n   which DF-QuAD satisfies.\n3. `docs/research/what-counts-as-evidence/formal-and-computational-models-of-evidence.md` § 1.4:\n   Verheij's paper is in *Artificial Intelligence and Law* 11 (2003), 167–195, not *Argumentation*\n   17 (the file's own reference list has the right venue); and the three-way summary drops Verheij's\n   third role, conditions of use, which is where relevance questions sit.\n4. The same file's ArgRAG entry, repeated in `docs/research/relevance-in-the-corpus-sweep.md` line\n   426: ArgRAG's reported best configuration uses uninformed base scores; retriever scores were\n   tested and did worse. The base scores do not *\"come from a model (relevance/stance confidence)\"*.\n5. `docs/research/scheme-bounded-decomposition-and-evidence-as-subgraph.md` line 126: *\"critical\n   questions map to undercutting attacks\"* holds for exceptions only; premise questions map to\n   undermining and counter-argument questions to rebutting (ASPIC+, Verheij, Carneades).\n6. `docs/research/kialo-deep-dive.md`: Kialo defines an impact vote as *\"the result of its veracity\n   and relevance\"*, which the doc does not say. It matters because Kialo data cannot calibrate\n   relevance on its own.\n7. `deliberus/graph/qbaf.py` header: *\"Switching to the literature-standard raw sum is a one-line\n   change at NEUTRAL_POINT\"* no longer holds. `relevance.link_contribution` keeps its own neutral\n   point, the floor and the no-standing rule, all of which also encode the offset.\n\n**Missing**:\n\n- The edge-weighted line (§ 2 B). The corpus sweep's \"not in the corpus\" list can point here.\n- Carneades's per-question burden and Verheij's conditions of use: the corpus cites both, and neither\n  reaches the build (D4, D5).\n- The quadratic damping near neutral, and the gap it opens between the judge's scale and the strength\n  scale (D1).\n- A refuted supporter reading as unexamined in the evidential weight (D2).\n- The dependence of the memoized pass on where the walk starts (D3).\n- Premise answers acting on one link rather than on the claim (D5).\n- Any representation of shared origin between supporters on the argument channel (D8).\n- Not read in this pass, and worth reading before the calculus is ruled: Gordon & Walton,\n  \"Formalizing balancing arguments\" (COMMA 2016, doi:10.3233/978-1-61499-686-6-327), for Carneades's\n  weighing functions; the extension-based varied-strength line (Martínez, García & Simari;\n  Coste-Marquis et al.; Cayrol, Devred & Lagasquie-Schiex); surveys of higher-order frameworks with\n  supports as well as attacks on links (UNVERIFIED which survey is current); da Costa Pereira,\n  Tettamanzi & Villata (2011) on trust-based semantics.\n\n---\n\n## 5. Questions only the founder can decide\n\n1. **Combining.** Should either factor be able to cap the other (a minimum: a link is only as strong\n   as its weaker factor), or should both always count (a product, as built and as the edge-weighted\n   literature does)?\n2. **A supporter examined and found weak.** Should its link do nothing (as built, and as Prakken's\n   accrual principle says), count against the claim (as probability says, and Mossakowski & Neuhaus's\n   *sum*), or do nothing while still counting as evidence somebody examined? *(Ruled 2026-09-30, for now: the third option, do nothing while still counting as examined; `relevance-and-the-two-factors.md` § 7.)*\n3. **Where standing comes from.** Should a supporter nobody has examined keep lending nothing, or\n   carry a base score as in every numeric semantics, and if a base score, drawn from what: the\n   methodology ladder, a located encounter with the world, a presumption?\n4. **Burden per question.** Should connection questions be sorted into those whose mere asking\n   removes the presumed relevance until answered (Carneades's assumptions) and those that need\n   evidence before they bite (exceptions)?\n5. **Premise answers.** Should an answer that the supporting claim does not hold weaken that claim\n   everywhere it is used, rather than on one link?\n6. **How far one link may move a claim.** Is QEM's near-silence around neutral wanted (a *settling*\n   link from a fully accepted claim reaches 0.60 at best), or should strength respond in proportion,\n   as the logistic and linear influences do?\n7. **Cycles.** One pass, a fixed point, or iteration over cyclic regions and one pass elsewhere; and\n   may the answer ever depend on which claim the computation started from?\n8. **Accrual.** Should many modest supporters add up without bound, be bounded by the strongest, or\n   be discounted for shared origin?\n9. **Rebuttal answers.** Should an answer that raises an independent case against the target be\n   filed as its own attack link rather than folded into this link?\n\n---\n\n**See also**: `relevance-and-the-two-factors.md` · `relevance-in-the-corpus-sweep.md` ·\n`weakest-link-arithmetic-and-the-merge-hunch.md` · `support-semantics-and-the-conjunction-problem.md`\n· `the-irit-harvest.md` · `what-counts-as-evidence/formal-and-computational-models-of-evidence.md` ·\nthe sibling report in this directory (Bayesian and informal-logic literature).\n"}