{"path":"research/the-irit-harvest.md","content":"# The IRIT Harvest\n\n**Date**: 2026-08-31 · **Type**: targeted survey, founder-commissioned mid-arc: *\"Should we maybe\njust take a beat and explore their papers for anything else that could be highly useful for us?\"*\n— after the weakest-link pass found that the corpus had cited the Amgoud school's axiom papers for\nyears without meeting their similarity line. **Read-depth caveat**: abstracts plus key sections\nread this session, not full papers; verify before citing outward. Three hauls, each with a named\nDeliberus consumer, plus one licensing result.\n\n---\n\n## 1. Strength explanation is a solved problem — Shapley attribution, with our own QEM author on the follow-on\n\n**The find**: Amgoud, Ben-Naim & Vesic, \"Measuring the Intensity of Attacks in Argumentation\nGraphs with Shapley Value\" (IJCAI 2017) axiomatize **contribution measures** — how to divide an\nargument's strength-loss fairly among its attacks — and prove a uniqueness theorem: **the Shapley\nvalue is the only measure satisfying the axioms.** It detects *worthless* attacks (no impact) and\n*redundant* ones (same loss). The follow-on is ours by lineage: Yin, **Potyka** & Toni, \"Explaining\nArguments' Strength: Unveiling the Role of Attacks and Supports\" (IJCAI 2024) — **Relation\nAttribution Explanations (RAEs)**: Shapley over ALL edges (direct and indirect, attacks AND\nsupports) in quantitative bipolar frameworks, with an **efficiency guarantee** (the per-edge\ncontributions sum exactly to the strength's deviation from the base score) and a probabilistic\napproximation algorithm for tractability. Their related-work section gives the field's taxonomy of\ncontribution functions (Čyras et al. 2022): removal-based, **gradient-based — which is our hinge**,\nand Shapley-based.\n\n**The consumer**: the badge's missing \"why this number\" layer. The hinge answers *what-if* (change\nthis part, how much does the root move); an RAE answers *why-now* (which edges are responsible for\nthe current number, summing exactly to the total) — complementary instruments, and the second one\ndoes not exist in Deliberus. Natural surfaces: the badge details panel, the triage feed (rank by\n|contribution| to show which edge is doing the work), and the synthesis citation gate (cite the\nhigh-attribution edges). Filed in TODO as a proposed instrument, unruled.\n\n## 2. Principled structural similarity between arguments — the seven-key judge's theory, pre-built\n\n**The find**: Amgoud & David, \"Measuring Similarity between Logical Arguments\" (2019→) plus the\ncompiled-arguments extension (ECSQARU 2021) and David's first-order generalization (2024/25).\nAxiomatized similarity for STRUCTURED arguments: a claim-plus-premises pair scored by a\n**support-similarity and conclusion-similarity split** (a σ dial between them), Jaccard-family\nmeasures over supports **normalized to clause form first** (\"compiling\" — because {p} and {p∧p}\nare equivalent with empty intersection), model-based comparison for conclusions, under principles\n(Maximality, Symmetry, Substitution, Syntax-Independence, Monotony, Dominance, Minimality). Their\nmotivation is *our* motivation: unaccounted similarity between arguments corrupts gradual-semantics\nevaluation. And the first-order paper names the hybrid hook explicitly: keep the principled\nstructure on top, swap the leaf-level predicate-equality for semantic similarity (embeddings) —\nstructure governs, embeddings only ground the leaves.\n\n**The consumers**: (a) the seven-key pair classifier's machine side — once claims carry\ndecomposition subtrees, a candidate pair is exactly a support/conclusion pair, and this is the\nprincipled non-embedding similarity for it; (b) the novelty dial's input (§ 9b of the weakest-link\ndoc) gains a structural option beside embeddings; (c) \"compile before comparing\" is the\nnormalize-then-match rule (the CDCR result) stated in logic — and rhymes with the mint-time\nbirth-certificate identity. Consequence worth naming: **the deeper the corpus decomposes, the more\nof claim-similarity becomes structural rather than textual** — the reuse-flywheel's\nidentity-decidability condition improving as a side effect of ordinary contribution.\n\n## 3. Analogical-argument semantics — the analogy daemon's strength story, already researched\n\n**The find**: a dedicated IRIT line on evaluating analogical arguments — Amgoud & Prade 2012\n(preliminary formal view), Amgoud, \"Evaluation of Analogical Arguments by Choquet Integral\"\n(ECAI 2020), and brand-new Amgoud, Hanocq & Lagasquie-Schiex (2026): premise-sensitive semantics\nwhere **the dispersion of attackers across premises matters, not just their number or strength**.\nThe 2020 paper distills philosophers' four strength considerations for analogies — **Relevance,\nNumber, Diversity, Disanalogy** — shows extension semantics violate them (an analogy is always\nattackable; binary rejection is wrong), and rebuilds three gradual semantics with the **Choquet\nintegral**, the aggregation function built for interacting contributors (synergies and\nredundancies among shared features). Their older AIJ foundations paper states the licensing\nprinciple plainly: for analogical arguments *every distinct dissimilarity weakens the analogy*, so\ncounting semantics fit there while being wrong elsewhere — **semantics choice is\nargument-type-relative**.\n\n**The consumers**: (a) **the analogy daemon** — its evaluate-analogy-claims pass has a ready-made\nprincipled semantics, and the four considerations map suspiciously well onto the four-check\ntransfer test (Number≈Recurrence, Relevance≈Mechanism identity, Disanalogy≈Adversary/Substrate) —\nthe corpus's home-grown instrument re-derived by the field; (b) **per-scheme aggregation**: we\nalready choose aggregation per relation type (the Munro/Bloch/Lesot license); the IRIT result\nextends the license one level — analogy-scheme edges may deserve counting-flavored semantics while\nothers keep quality-precedence, a per-SCHEME layer atop the per-relation one; (c) the **Choquet\nintegral** joins § 9b's reconciliation family as the general interaction-aware aggregator (the\nnovelty dial and fair-share are special cases of a capacity).\n\n## 4. The licensing result, for the record\n\nThe Amgoud school's foundations paper (AIJ) makes explicit what Deliberus has been doing by\ninstinct: no semantics should satisfy all desiderata, because applications and argument types\nimpose different constraints — the choice of aggregation is *itself* application-relative. That is\nthe per-relation channels, the per-scheme prospect above, and the de-bake template's philosophy,\nstated by the axiomatists.\n\n## 5. Who these people are — and what to take on a forest walk\n\n**The Toulouse school (IRIT, CNRS)**: Leila Amgoud is the central figure — two decades of\naxiomatizing argument strength, similarity and analogy — with Jonathan Ben-Naim, Srdjan Vesic,\nVictor David, and Marie-Christine Lagasquie-Schiex (who co-initiated gradual semantics in 2005),\nplus the Paris collaborators Bonzon, Delobelle, Konieczny, Maudet, Doder. **The London QBAF\nschool (Imperial)**: Francesca Toni leads the computational-argumentation group and founded\nImperial's explainable-AI centre (general chair of IJCAI-ECAI 2026); Nico Potyka (QEM's author)\nand Antonio Rago (DF-QuAD) are her orbit. **Stanford NLP**: Christopher Manning (GloVe,\ntree-recursive networks, the SNLI inference corpus, 2024 IEEE von Neumann Medal); Bill MacCartney\ndid natural logic as his doctoral student before industry. **Philosophy/epistemology**: John\nPollock (d. 2009, the OSCAR defeasible-reasoning project), L. Jonathan Cohen (d. 2006, Oxford),\nBovens & Hartmann (Bayesian epistemology), Claveau and Landes (variety of evidence). **Informal\nlogic**: Douglas Walton (d. 2020), James Freeman, Stephen Thomas. **Evidence law**: Ronald Allen\n(Northwestern) and Michael Pardo, with Kevin Clermont (Cornell) as the fuzzy-logic dissenter.\n**Belief functions**: Thierry Denœux (UT Compiègne). **Subjective logic**: Audun Jøsang (Oslo).\n**Social choice**: Nicolaus Tideman (Virginia Tech) — ranked pairs, the clones criterion, and a\nlifelong Georgist. **Meta-analysis**: Larry Hedges, Elizabeth Tipton, James Pustejovsky.\n\n**Verified forest-walk listening** (checked 2026-08-31):\n\n- **Excited Utterance ep. 161 — Ronald Allen**, \"Minimal Rationality and the Law of Evidence\"\n  (Feb 2025). The evidence-law podcast (host Edward Cheng, himself in the burden-of-proof\n  literature); Allen is the relative-plausibility school in person. Short episodes; the back\n  catalogue likely holds Pardo and Clermont too.\n- **TWIML — Christopher Manning**, \"Language Understanding and LLMs\" (May 2024, fresh off the von\n  Neumann Medal): linguistics vs LLMs, whether they learn human language structure, reasoning.\n  Longer and more biographical: **The Gradient ep. 41 — Manning** (2022), his whole arc from\n  computational linguistics through GloVe and attention.\n- **Exploring Liberty with Barend — Nicolaus Tideman**, \"Geoism: The Needed Moral Revolution\"\n  (Dec 2024, on Spotify): the clones-criterion inventor on land, liberty and public\n  decision-making. Georgism-flavored rather than voting theory, and charming for it.\n- **Francesca Toni**: no consumer podcast surfaced — but her IJCLR 2021 keynote\n  \"Argumentation-Based Explainable AI\" (YouTube) is a listenable overview of the whole QBAF\n  programme, and her Imperial inaugural \"machine arguing\" talk is online.\n- Already consumed on earlier walks, for completeness: **Ruth Chang** (parity — the ruling\n  followed the listen) and **Sara Walker** (assembly theory, Lex Fridman). Amgoud, Denœux,\n  Jøsang, Potyka: nothing podcast-shaped surfaced in this pass — conference talks only.\n\n## 6. Frontier-matched listening — media mapped to the decisions actually pending (founder-commissioned, verified 2026-08-31)\n\nThe founder asked which corpus-referenced researchers sit closest to the OPEN decisions, and what\nlistening would \"jog the mental cogwheels.\" Mapping runs decision → thinker → verified episode.\nAll names verified corpus-present before inclusion (Mercier in 17 docs, Kahan 18, van Gelder 14,\nFishkin 13, Ostrom 10, Community Notes 19, Landemore 4, Tang 5, Megill 2, Bail 4).\n\n| Pending decision | Best cogwheel-jogger | The verified media |\n|---|---|---|\n| **The merge-eval sitting** (seven-key labels; per-kind sameness bars; is human ratification even the right ask) | **Hugo Mercier** — the argumentative theory: reasoning evolved for producing and EVALUATING arguments interactively, solo reasoning is biased by design — the deep warrant for propose-then-evaluate over solo verdicts | Sean Carroll's **Mindscape #233**, \"Reasoning and Skepticism\" (Apr 2023) |\n| **The up-transmission whisper** (element-wise vs holistic strength) | **Ronald Allen** — law's live version of exactly this: prove per element, or compare whole stories | **Excited Utterance #161**, \"Minimal Rationality and the Law of Evidence\" (Feb 2025) |\n| **Workshop participant pick + curiosity-based matching** (pending founder call) | **Dan Kahan** — the curiosity result in his own voice, with tribal psychology as the failure mode being matched around | **You Are Not So Smart #122/#146** (with Lilliana Mason, tribal psychology; his science-curiosity paper is in the show notes) |\n| **The live election test + co-present session design** | **Chris Bail** — the measured warning for the null arm: stepping outside the echo chamber made people MORE polarized; flags vs reasons is the variable | **Opinion Science #39** (2021) or **Democracy Works** (Jul 2024), both on *Breaking the Social Media Prism* |\n| **Bridging feed ruling + what production consensus-finding feels like** | **Audrey Tang** — Polis run inside a government: no reply button, agenda crowdsourced, only cross-cluster supermajorities count; the no-verdict architecture's closest living cousin | **80,000 Hours #120** (Feb 2022; the Polis walkthrough starts 13:48) · bonus: **Colin Megill on The Futurists**, \"Computational Democracy\" |\n| **Condition (d) / the constellation review / why the machine holds the hand** | **Tim van Gelder** — the argument-mapping generation's own post-mortem: the tools demanded a skill (his windsurfing water-start analogy) and never got uptake — the exact failure class the inversion principle exists to escape | YouTube: \"**Argument Mapping & Intelligence Augmentation**\" (2013, 22 min) and the longer Future Salon talk |\n| **The epistemic constitution gap** (Ostrom's participation-in-rule-making; who designs the process) | **Hélène Landemore** and **James Fishkin** — open democracy and deliberative polling; her \"citizens on top, experts on tap\" is the ratification philosophy at institutional scale | **The Ezra Klein Show**, \"A Radical Proposal for True Democracy\" (Feb 2021); fresher: **Democracy Innovators** (Mar 2026). Fishkin: **Future Hindsight** (Jul 2025) |\n\nSequencing suggestion for one long walk: Mercier → Tang → van Gelder. Mercier reframes what human\njudgment is FOR right before the eval-format decisions; Tang is the closest running system to the\nratification end-state; van Gelder is the sobering ancestor — the man who built argument mapping\ntools for twenty years explaining, with great honesty, why they never took off.\n\n---\n\nCross-references: [weakest-link-arithmetic-and-the-merge-hunch.md](weakest-link-arithmetic-and-the-merge-hunch.md)\n§ 9b (the reconciliation family this extends) ·\n[soft-canonical-clustering-and-reversible-merge-semantics.md](soft-canonical-clustering-and-reversible-merge-semantics.md)\n(the seven-key eval the similarity measures serve) · [the-analogy-daemon.md](the-analogy-daemon.md)\n(consumer of § 3) · [strength-layer-audit.md](strength-layer-audit.md) (the badge the RAE would\nexplain) · [qbaf-gradual-semantics-research.md](qbaf-gradual-semantics-research.md) (the QEM\nchoice; Potyka).\n"}