{"path":"research/session41-when-a-reason-fails.md","content":"# Session 41: When a reason fails\n\n**Date**: 2026-09-28 (evening) to 2026-09-29 · **Type**: two literature reviews landed and corrected\nthe corpus; the split panel's future became a measurement plan; the first calculus question was put\n(and ruled the next day, for now: § 9); an interactive page was built to answer the founder's questions about it; a\nharvest of two community threads was handed over for mapping.\n\nContinues session 40 (`session40-two-factors-and-a-day-of-rulings.md`). Its through-line: **the first\ncalculus question turned out to rest on a question nobody had asked, what the starting point of a\nclaim means.**\n\n---\n\n## 1. The literature reviews, and what they corrected\n\nThe review the founder asked for on 2026-09-28 landed in two halves, both in\n`strength-calculus-literature/`: computational argumentation (37 publications graded against the\neight open calculus decisions) and Bayesian epistemology with informal logic (about 60 sources).\n\n- **A claim of ours was half wrong.** The session-40 synthesis said the published model already\n  multiplies a supporter's strength by a weight on the link. Potyka's QEM feeds each supporter's own\n  strength, raw, and **has no link weights**; its weights sit on arguments. The product comes from\n  later edge-weighted variants, the closest to the ruling being Yin, Potyka, Rago, Kampik & Toni\n  (2025), who read the weight as relevance. Corrected on all six surfaces that repeated it.\n- **Seven other corpus errors**, all fixed the same evening: the Euler-based formula, an overstated\n  convergence claim, Verheij's venue and his dropped fourth role for critical questions (conditions of\n  use, where relevance questions sit), ArgRAG's base scores in two docs, critical questions as\n  undercutters only, what a Kialo impact vote measures, and a stale comment in `qbaf.py`.\n- **A filing corrected by the Bayesian half**: `relevance-and-the-two-factors.md` § 2 had filed\n  *is the expert credible* under acceptability; the build's sort and the Bayesian reading file it under\n  connection.\n- **The Bayesian headline**: the ruled rule is Jeffrey conditioning, exactly, when relevance means Δ,\n  how much likelier the claim is if the supporter holds than if it fails (Christensen 1999; Godden &\n  Zenker 2018). And the machine judge's prompt measures sufficiency, not Δ, so **its readings are not\n  to be written** until the founder has said what relevance means.\n- **Three code findings, none deployed**: a refuted supporter lands exactly on neutral and so reads as\n  unexamined; the memoized one-pass walk can keep a truncated value, so a result can depend on where\n  the walk started (latent); premise answers weaken a supporter on one link only.\n\n## 2. The judge, on a second model family\n\nAfter the founder refreshed the `claude` login, the dry run over the 12 links that matter today ran on\nClaude Sonnet throughout: 12 of 12 confident, **6 strong, 4 moderate, 2 weak**, none irrelevant, and\nno disputed filing. The number no longer piles up in the middle (the fallback trial had put 8 of 12\nat 0.5). Nothing was written. Detail: `relevance-and-the-two-factors.md` § 5.\n\n## 3. The split panel: from a question to a measurement\n\nThe research on his question (`split-review-and-the-founders-intentions.md`) found his line unchanged\nsince 2011, *automate as far as possible, people at the residue*, and his own test of 2026-09-21: a\npanel belongs in the UI only if the machine cannot be trusted to do the work alone. His trust in the\nsplitter was conditional (*\"IFF it works reliably\"*) and **had never been measured by a person**: no one\nhad touched any of the 37 machine-cut parts, and a second model called about two thirds genuine.\n\nThe proposal was a sitting: he judges those parts in one go, and the panel goes if nine in ten or more\nhold up, his own 2011 bar for promoting a validator. His question in reply:\n\n> *\"Sounds like a great idea. But. Hasn't a sitting like that been created already? What\n> sittings/reviews have been created thus far?\"*\n\nIt had, nearly. **Three review surfaces existed and not one had a single answer** (measured\n2026-09-29: both answer tables empty): the parts review (built 2026-09-22, 60 links of every kind, 11\nof them machine-cut), the sameness sitting (built 2026-09-21, 120 claim pairs under the seven keys), and\nthe split panel itself. A fourth, a quick pass over claims flagged as bundled, was designed and never\nbuilt. After *\"we'd still keep the parts-review, but it's separate? or?\"*, the answer was one page with\na second batch, and his ruling:\n\n> *\"Ok go, sounds wise\"*\n\nBuilt the same evening: a sitting named `splits-1` holds every machine-cut part and nothing else (37\nlinks, 8 of them blind, so the anchoring effect of seeing the machine's reading can be measured), and a\nlink answered in either batch counts in both, carrying the blind flag it was answered under. The app's\nproduction-deploy check stopped Claude deploying it, so it waits on a branch built from the live commit\n(`deploy/splits-sitting`), which keeps the held link-strength rule out of production.\n\n## 4. Should a refuted reason pull its claim down?\n\nThe Bayesian report disagreed with the build on one point. His question:\n\n> *\"Why was this not built already? Don't change anything yet, just explain.\"*\n\nIt was Claude's call during the build, marked as his to decide: with subtraction on, anyone could attach\na weak \"supporter\" to a claim they oppose and refute it. Asked which voices say what:\n\n- **For pulling it down**: probability theory's complementary evidence (if a reason's truth makes a\n  claim likelier, its falsity must make it less likely) and conservation of expected evidence (a check\n  cannot be expected to help only one way); the Jeffrey form moves the claim both ways. The report adds\n  a cost of the one-sided rule: examining a supporter can then only help, a free bet for either side.\n- **Against**: Prakken on accrual, where a flawed reason *\"does not take part\"*; the mainstream\n  argument-strength models (QEM, DF-QuAD, Euler-based), where supporters only raise and a worthless one\n  changes nothing; Mossakowski & Neuhaus define both and leave the choice open.\n\n**The reconciliation Claude proposed, which the rest of the session built on**: the Jeffrey form is\nexact only when the claim's starting point already priced the reason in. Deliberus's start means\n*nothing examined*. So for testimony or an argument someone offered, properly referenced Bayes returns\nthe claim to where it started when the reason fails, which is the drop-out rule; only a failed test,\nwhere the claim could have failed, pushes it below. Two more points: the graph already has a home for\nevidence against, an objection, so subtracting as well can count one finding twice; and fixing\nrelevance alone does not make subtraction safe, since every claim is measured from 50/50 however it\nwas invented. A worry that authors often knock down weak objections inside their own texts is not\nborne out: of 36 attacks inside one source, one is answered inside that source (measured 2026-09-29).\n\nAfter *\"Right. Ok, ask me the first calculus question. Ultrathink\"*, the question went to him as: when a\nreason is refuted, should its claim fall below where it started? Proposed: no, keep the build's rule,\nwith two repairs (a refuted reason still counts as examination; a failure that is itself evidence gets\noffered as an objection). His answer, **explicitly not a ruling**:\n\n> *\"Keep it, with the two repairs. Seems more intuitive/truthful to me. No... actually I'm not sure.\n> Because if the implication is that there's SOME truth value to a claim if it's in the middle then\n> that'd be misguiding if there's no good and relevant reasons whatsoever for believing it's true. But\n> ofc all of this depends on whether it's been examined deeply and exhaustively and it all depends on\n> how you see/judge *relevance*. Right? Btw I'm curious what is an 'objection' vs an attack link/claim,\n> and what are these in relation/comparison to a 'refuted reason'? Can they simply be negations of each\n> other? How do we handle this potential mirroring?\"*\n\nThree things came out of it, none ruled:\n\n1. **The start has three meanings**: nothing examined, nothing stands (every reason checked failed),\n   and unsettled (reasons on both sides balance). Only the last is today's `UNSETTLED`. Whether\n   *nothing stands* should lean against a claim depends on how thorough the look was.\n2. **An objection is the claim at the far end of an attack link**, and refuting a reason often asserts\n   its negation, which can itself be filed as an objection. One finding can then sit in the graph\n   twice.\n3. **A third rule answers the mirroring**: a reason and its negation as one question with two weights,\n   what the claim gains if the reason holds and what it loses if it fails. A failed test weighs; a\n   failed witness or argument does not. The critical questions' yes and no poles already have this\n   shape. It counts the finding once only if the system recognises the mirror.\n\n## 5. The page\n\nHe asked to understand before ruling: *\"I want to fully absorb and understand all this. Have not yet\ngrasped the full picture good enough to make this call. Make a visualization to explain all of this.\nKeep it very very short and concise and pedagogical\"*. An inline comparison came first; then *\"Let's\nbuild a better visualization based on an example in a web page to get a better sense of what we're\ndiscussing, please. Iterate on it until it's very well-polished and pedagogical.\"*\n\nThe result is **When a reason fails** (claude.ai/artifact/BbEi36mTifgXJ2avwMYQHz, private): one claim\nabout a bridge, with a test (the inspection), a witness (the neighbour), a straw man planted to be\nknocked down, and the failed inspection written as an objection. Five cases or free settings; all three\nrules update side by side, each marking where it gets the case right or wrong; three takeaways answer\nhis questions. Built in the site's own palette, with claims in a reading face. Iterations fixed the\nphone order (the results now follow the controls), a result label that wrapped badly, and a missing\nflag for the straw man under the build's rule. The numbers are illustrative.\n\n**A method note for the calculus choices still to come**: he wanted a worked example before the\nquestion, not after, and the inline sketch was too compressed to do that job.\n\n## 6. A harvest handed over\n\nThe founder: *\"Astra has finished the work on the Borderland Talk APs, finally.\"* The second of the two\ncommunity threads is now fully verified and is being drafted into a private map, following the method\nof the first. Everything about it stays private: `TODO.md` § Borderland and the compartment's own files.\n\n## 6b. One line added to always-loaded memory\n\nAt the end of the session one line went into the project's always-loaded instructions, beside the strength entry:\na refuted supporting claim adds nothing to the claim it supports, on purpose and still unruled, so no session\nshould switch on subtraction before the founder rules. Graded *predicted* rather than demonstrated: no session\nhas yet flipped the switch, but the Bayesian report framed the build as disagreeing with probability theory, so a\nsession reading it cold could take the one-sided rule for a bug. The founder, asked whether to add it:\n*\"Yes, add it\"*.\n\n## 7. Considered and rejected\n\n- **Deploying from master** to ship the review batch: it would ship the held link-strength rule before\n  his first calculus choice.\n- **A new review page for the machine-cut parts**: the parts review already asks the right question;\n  a second batch on it costs a filter.\n- **Subtracting for every refuted reason**: see § 4.\n- **A per-link mark for test-like reasons**: more machinery, and the same double count whenever someone\n  also files the objection.\n- **The inline sketch as the explanation**: it compressed the three meanings of the start and the\n  mirroring into rows he could not interrogate.\n\n## 8. Open loops, each with its home\n\n- The first calculus question: ruled 2026-09-30, for now (§ 9; `relevance-and-the-two-factors.md` § 7).\n- The machine-cut batch's deploy: done 2026-09-30 with `master` (§ 9).\n- The sitting itself, once deployed: same section.\n- The judge's readings, held until relevance is defined: `TODO.md` § Relevance.\n- The three code findings: `TODO.md` § Relevance.\n- The second community map: `TODO.md` § Borderland.\n\n## 9. Continuation, 2026-09-30: the ruling, and the question under it\n\n**The deploy hold was lifted.** Founder: *\"Just keep deploying continuously after we make changes. I'll\nmost likely be ruling about a lot of epistemic calculus stuff in the near future so no need to hold off\non deploying just because things are in flux, I'm thinking\"*, then, refining it: *\"I just meant that we\ndon't have to hold off on this epistemic calculus change in particular. I still just want us to deploy\nwhen it's considered and feels natural, we're confident nothing will break, et cetera. So, like, not\nafter each and every change, I guess...\"* `master` went live the same morning (`65ab563`), carrying\nthe two-factor rule and the machine-cut batch; the rule now stands in `CLAUDE.md` where the hold was.\n\n**The recommendation**, asked for as *\"Think extremely deeply about this ... Then come back to me with\nyour recommendations. Wait for my approval before changing anything\"*: keep the rule as built, and rule\nthe principle that failure counts against a claim only through a contribution that says why (a finding\nas an objection, a need as a requirement claim, a thorough search as an argument from ignorance), with a\ndisplay state for *nothing stands*. Asked why probability disagrees, he said *\"I still don't fully\nunderstand the different approaches to all this and the deeper reasons why 'probability' says something\nelse rather than what your recommendation is...\"*, and after the explanation ruled: *\"but we'll go with\nyour recommendations on this for now, and then we can always poke holes in that later if we feel like\nit.\"*\n\n**Then he poked the first hole, the same hour**: the strengths are meant to become the graph's best\nguesses, computed from subgraphs, and *\"one finding can be written twice\"* is a matching problem, not an\nargument. It corrected the explanation Claude had given (both rules start at the same value; they differ\non what an unchecked reason is worth, a coin flip or nothing until shown) and showed that the sign\nquestion is the starting-value question: with honest starting values per claim, a failed reason would\nsubtract its starting value times its relevance, which is probability proper and his two weights.\nRecorded as ledger Q34 (the ruling) and Q35 (the path, open), with the reasoning in\n`relevance-and-the-two-factors.md` § 7.\n\n**A method note, again**: the explanation that landed was the one built on what an unchecked reason is\nworth, not the one built on double counting. His objections came from imagining the product a year on,\nwhich is the test to run on any calculus choice before putting it to him.\n\n## Cross-references\n\n`relevance-and-the-two-factors.md` (§§ 5–6) · `strength-calculus-literature/computational-argumentation.md`\n· `strength-calculus-literature/bayesian-and-informal-logic.md` · `split-review-and-the-founders-intentions.md`\n· `the-parts-review.md` · `session26-open-rulings.md` (rows B7, Q33) · `docs/ux-principles.md` (the\n`UNSETTLED` ruling) · `what-counts-as-evidence.md` § 8c · `session40-two-factors-and-a-day-of-rulings.md`.\n"}