{"path":"research/relevance-and-the-two-factors.md","content":"# Relevance, and why a link's effect has two factors\n\n**Date**: 2026-09-28 · **Type**: synthesis of a full corpus sweep, checked against the code · **State**:\nRULED 2026-09-28 (ledger Q33; founder: *\"About the decision: your proposal is blessed, make it so.\"*). The\nliterature review he asked for landed the same day (`strength-calculus-literature/`). What a failed reason\ndoes was ruled 2026-09-30, for now (§ 7, ledger Q34); how the two factors combine, what an unexamined link\npresumes, and fixed point versus one pass are still open. The direction on weighing answers is ruled\n(ledger Q3).\n\n**The question, verbatim** (founder): *\"It feels rather obvious that the answers should move the link's\nstrength in proportion to how well supported it is. But here we also - Just like in the case of normal\nsupport or attack links from a mother claim, or from subclaims to a main claim, you know the standard\ncase - we need to consider the *relevance* as well... search for everything that has been documented\nand philosophized and have decided in the past in the corpus please about relevance and how we could\ndetermine it and let it affect the strength of a link from a claim/subgraph I guess\"*\n\nThe full sweep, verbatim: [relevance-in-the-corpus-sweep.md](relevance-in-the-corpus-sweep.md).\n\n---\n\n## 0. The answer in five lines\n\n1. **The standard case does not have relevance either, and it has less than that.** For an ordinary\n   support or attack link, the target's strength moves only by the link's own number, computed from\n   its counted critical-question answers. **The supporting claim's own strength never reaches the\n   target** (`qbaf.py`: each incoming edge contributes `edge_strength − 0.5`; `hinge.py`: support and\n   attack *\"enter as local energy on each node, not as strength flow\"*). Relevance exists in code only\n   on part-of links, as the weight of the weakest-part cap.\n2. **The textbook model has the first factor; later work added the second.** Potyka's QEM, the\n   published model the strength layer follows, feeds each supporter's own strength into the\n   target's energy and has no link weight. Edge-weighted variants published since (Potyka 2021;\n   Yin, Potyka, Rago, Kampik & Toni 2025, who read the weight as relevance) multiply that strength\n   by a weight on the link. The March build replaced the supporter's own strength with one counted\n   number per link; the corpus calls it a deliberate divergence, and no ruling on it is recorded.\n   *(Corrected 2026-09-28: this item first said QEM itself weights links. It does not; its weights\n   sit on arguments only. Primary-source quotes: `strength-calculus-literature/computational-argumentation.md` § 0.)*\n3. **\"Relevance\" means six different things in the corpus.** The two that matter here are *bearing*\n   (does the premise move the conclusion, and how much) and *dependence* (how much a whole needs a\n   part). The only machine relevance score built measures dependence; the question asks about bearing.\n4. **The classic frame is not in the corpus, and it fits exactly**: Johnson and Blair's acceptability,\n   relevance and sufficiency. Acceptability is the supporter's own strength; relevance is the link's\n   bearing; sufficiency is whether the reasons together are enough (the corpus's *lift* question).\n5. **Ruled**: a link's effect is the supporter's own strength times the link's relevance, with\n   relevance living on the link, proposed by a machine, argued by people, and never voted.\n\n## 1. What the corpus holds\n\n**Decided.** The requirement-weighted cap on parts (2026-09-17, *\"Build the designed completion.\"*). The\nfounder's direction on who judges relevance (2026-09-19): *\"ideally the RELEVANCE (which I guess all\nthis is about) is judged provisionally by the machine/LLM-passes or whatever and then humans can\nchallenge it but I'm wary of saboteurs/spammers adding bogus parts that aren't actually relevant\"*,\nwith attack-proofness left open (ledger O13). The relatedness gate that decides, yes or no, whether\nthe automatic linker writes a link at all (2026-09-22). Votes are telemetry since 2026-08-31, so\nrelevance cannot be voted; it must be argued.\n\n**Built.** The cap on part-of links, weighted by a *requirement weight* (presumed 0.5, argued through a\nrequirement claim, 1.0 when tagged necessary). A machine relevance score of *\"how much the whole\ndepends on this part\"*, run on all 197 part links (mean 0.33; ⚠ 84 of those 197 were the linking step's premise links, stored backwards and since retired, so this is a mean over a pooled set: [auto-connect-upgrade.md](auto-connect-upgrade.md) § 2c). A switch meant to let that score lower\nthe cap, **wired to nothing**: no caller passes the score in, so turning it on would change nothing.\nFor support and attack links: the yes/no gate, and critical questions that ask about relevance\n(expert field match, relevant differences, whether a character attack bears on truth), counted like\nevery other question.\n\n**Philosophised.** *Relevance strength* as a contestable property of the link (`drowning-in-claims.md`\n§ 8). Relevance as a gate plus a magnitude (law's split between relevance and probative value, in the\nevidence dossier). Relevance on the link, never on either claim (Cartwright and Hardie, GRADE's\nindirectness). Critical-question answers as parameters of a likelihood ratio (Hahn and Hornikx). The\nearliest statement of the two factors is the March prediction-markets doc: *\"strength S reflects both\nP's probability and the conditional relevance P(C|P)\"*. And the 2013 prototype had an *Irrelevant*\nvote on every link, used for a quarter of all votes.\n\n## 2. The ruled design: two factors, each with an honest home\n\n| Factor | What it asks | Where it lives | How it is set |\n|---|---|---|---|\n| **Acceptability** | how well supported is the supporting claim itself | the supporter's own reading, computed recursively | its own supports, attacks and answered questions |\n| **Relevance** | if the supporter holds, how much does that bear on the target | a weight on the link: a gate (does it bear at all) and a magnitude | proposed by a machine of a second model family, with an abstention; argued by people through the link's connection questions and requirement-style claims; never voted |\n| **Sufficiency** | are the reasons, together, enough | the whole claim | the open *lift* question, unchanged by this proposal |\n\n**The critical questions sort themselves into the two factors.** Questions about the premise (did\nthe expert actually assert it, is the premise acceptable) are acceptability, and belong to the\nsupporter's own reading. Questions about the connection (is the expert credible and in the right\nfield, are there relevant differences, does this follow if the premise holds) are relevance, and\nset the link's weight. *(Corrected 2026-09-28: this paragraph first filed \"is the expert credible\"\nunder acceptability. The build's sort, `CQ_BEARS_ON`, files it under connection, and so does the\nBayesian reading, since credibility governs whether the assertion carries over to the conclusion:\n`strength-calculus-literature/bayesian-and-informal-logic.md` § 5.)* Today\nboth kinds are counted into one number, which blends the two factors the question asks to separate.\n\n**Answers weigh by their own standing** (ledger Q3, agreed in direction), and each answer acts on the\nfactor its question belongs to.\n\n## 3. What it would change, and the risks\n\n- **Every strength in the graph moves**, since supporters' strengths start to flow. Most links would\n  still read *not yet examined*, because few questions are answered.\n- **Cycles**: strength flowing along support links needs the fixed-point computation the published\n  model uses; measured in August, three two-step cycles and one three-step cycle in about 3,950 links.\n- **Brigading**: relevance claims can be propped up by many accounts, a strategy-class weakness named\n  and undefended (ledger D8, D9). Payoff removal comes before detection: a relevance proposal starts\n  at the machine's modest reading, and raising it must be argued.\n- **Same-hand bias**: a relevance judge from the extraction model's family grades its own family's\n  links; the corpus already recommends a second family.\n- **Do not reuse the dependence score as bearing.** It would discount a corroborating study for having\n  siblings, penalising redundancy that the novelty dial already handles.\n\n## 4. Defects the sweep found (repairs, filed in `TODO.md`)\n\n- The relevance switch for the cap is wired to nothing, while its own comment says switching it on\n  re-scores every capped claim.\n- Docs and a test say the strength layer reads `support_semantics`; it reads `support_interpretation`,\n  so the \"197 of 197 untagged\" measurement queried a property nothing reads.\n- The hinge computation omits the coverage verdict the badge uses, so the two disagree on invented\n  splits.\n- The endpoint for arguing a part's requirement has no screen.\n- Requirement claims are always typed normative, though many are empirical or definitional.\n- One worked example calls the same relation *relevant* in one doc and *similar* in another.\n\n## 5. What was built (2026-09-28, the day of the ruling)\n\n**State: built, tested, measured on the live graph; live since 2026-09-30.** It went out before the\ncalculus choices below, on the founder's word (*\"we don't have to hold off on this epistemic calculus\nchange in particular\"*); the defaults stay as built until he rules on each.\n\n- **`deliberus/relevance.py`** holds the ruling and every named choice. An argument link now\n  contributes `relevance × (acceptability − 0.5)`, signed by polarity (`qbaf.argument_edge_contribution`).\n  A question's own pole keeps the counted rule: its answers already reach the target through the\n  question on the argument link, and letting the pole's own standing flow as well would count the\n  same answer twice.\n- **Acceptability** is the supporting claim's own computed strength when it has been examined\n  (`badge_inputs.attach_flow_inputs`, walked recursively like the decomposition channel), moved by\n  the answers to the link's premise questions. An unexamined supporter contributes nothing.\n- **Relevance** comes from an argued relevance claim, else a confident machine reading, else a\n  presumption by writer (0.5 for the source, a person or a split; 0 for a machine's guess between\n  other people's claims), and the link's connection questions move it from there.\n- **The critical questions are sorted** in `schemes.CQ_BEARS_ON`: premise (acceptability),\n  connection (relevance), rebuttal (acts on the target directly), by ASPIC+'s undermining,\n  undercutting and rebutting, adapted to a graph whose links start from one explicit claim. Most\n  land on connection, because Walton's questions mostly probe implicit premises.\n- **People argue relevance** at `POST /claims/{id}/links/{source}/relevance`, which mints an\n  attackable relevance claim (type left blank: nothing has checked whether it argues fact,\n  definition or value) and a non-propagating `ARGUES_RELEVANCE` link.\n  *(2026-10-07, proposed and unruled: what a written-out step adds, its rule, facts and joined\n  names, would be the grounds of this claim, since a rule drawn as a second premise needs a rule of\n  its own: [philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md) § 3; ledger AD22.)*\n- **The machine tier** is `deliberus/relevance_judge.py` with `scripts/propose_link_relevance.py`:\n  the no-answer asked first as its own gate, then direction, then a five-step scale (settles,\n  strong, moderate, weak, marginal) mapped to numbers in one table; bearing judged rather than\n  truth; the model that actually answered recorded on the edge (`client.last_served_model`).\n  A reading that runs the other way from the link as filed is kept as a disputed filing and never\n  used. Dry run by default; writing to the live graph waits on the founder.\n- **The first trial (dry run, the 12 links that matter today) changed the judge three ways.** The\n  Claude engine's login had expired, so the run fell through to a Gemini fallback, the same family\n  that extracted the links, while the label said Claude: the script now refuses to fall through,\n  and every reading records its serving model. Asked for a number with three anchors, the judge\n  answered 0.5 on eight of twelve, so the number became a named scale. And on one link filed as an\n  attack its own reasoning described support, so direction is now asked outright and a\n  disagreement is recorded as a candidate misfiling rather than silently used.\n- **The second trial (dry run, same 12 links, 2026-09-28 evening, after the founder refreshed the\n  login) ran on Claude Sonnet for all twelve**, a second family to the Gemini extraction: 12 of 12\n  confident, **6 strong · 4 moderate · 2 weak**, no link read as irrelevant, none at the scale's\n  ends, and **no disputed filing** (the attack link that the fallback read as support was read as\n  a moderate attack). The two weak readings are the instructive ones: a claim that minimum-wage\n  laws price workers out bears only weakly on the claim that most economists *believe* so, since a\n  survey result depends on much besides the mechanism; and the coherence of loss of control bears\n  only weakly on there being *several* routes from it to catastrophe, since it secures a\n  precondition and nothing more. Nothing was written; `--write` still waits on the founder.\n- **The hinge now computes through the badge** (`hinge.hinge_analysis`), over a region that\n  includes supporters as well as parts. Its old decomposition-only model had drifted from the\n  badge (the sweep's tension fifteen); `propagate` remains only as the pure model its tests pin.\n\n**Measured on the live graph the same day**: of 722 claims with an incoming link or part, **one**\nchanges. Only 12 of 509 argument links start from a claim that has standing, and none of those\nmoves its target by more than 0.001, because almost nothing in the graph has been examined yet and\nthe counted answers are small (ledger Q3, whose arithmetic is still open). The one change is the\njustice-and-equality claim from the first dogfood run: a connection question on its link was\nanswered, but the value premise the link starts from has no standing, so under the ruling nothing\nreaches the claim. Its badge now says so instead of claiming nothing was scored.\n\n**A consequence for the descent, worth his eye**: a connection question on a link whose supporter\nhas no standing can no longer move anything, so the worth-asking price goes quiet on it and the\ndescent will offer premise questions, or the supporter itself, first. Relevance is only ever\nrelevance of something.\n\n**The choices made to build it, each one named constant, all his to revisit**:\n\n| Choice | As built | Why this default |\n|---|---|---|\n| Does a weakened supporter push its target the other way? | No (`WEAK_SUPPORTER_SUBTRACTS = False`); **ruled 2026-09-30, for now (§ 7)** | With it on, anyone could attach a weak \"supporter\" to a claim they oppose and refute it, dragging the claim down by a straw man. Off removes the payoff instead of detecting the move. Probability theory says the opposite when a link's conditional probabilities are known. |\n| What does an unexamined link presume? | 0.5 for the source, a person or a split; 0 for a machine's guess | The same modest, rebuttable presumption a part's requirement weight carries. |\n| How do the two factors combine? | Product | The ruling's wording. Pollock's weakest link (minimum) is the main alternative. |\n| Fixed point or one pass? | One pass, a cycle cut at the claim already on the path | The live argument links hold no two-cycles; the largest upstream region is 41 claims. |\n| Which question bears on what? | `CQ_BEARS_ON`, sorted by hand | A handful are judgement calls between premise and connection. |\n\n## 6. Open consequences the build exposed (not yet ruled)\n\n- **Standing has to originate somewhere.** With the offset convention (an unexamined claim adds\n  nothing), supporters' standing flowing (this ruling) and answers weighed by their own standing\n  (ledger Q3's direction), a graph whose leaves are all neutral would never move at all: every\n  claim's standing would come only from other claims' standing. Today it originates in the counting\n  rule (0.15 per piece of evidence on a question's pole, 0.3 per link into a part, votes on parts).\n  Replacing counts with weighed standing therefore needs a grounding: base scores (the methodology\n  ladder was designed as exactly that and no badge reads it), located encounters with the world\n  (`what-counts-as-evidence.md`), or presumptions of the Carneades kind. This is where Q3, this\n  ruling and the evidence arc meet, and it is his to settle. (2026-09-30: the sign of a failed\n  reason turned out to be the same question, what a claim is worth before anyone examines it; § 7.)\n- **The two channels disagree about unexamined claims.** A supporter with no standing now lends\n  nothing, while a decomposition child with no standing still enters its parent at a count-based\n  prior (two unscored supports lift it to 0.632), so structure alone can still manufacture strength\n  through parts. The same treatment for both is the consistent choice; changing the parts channel\n  moves every decomposed claim, so it is not taken here.\n- **Question ordering changed with it.** A connection question on a link whose supporter has no\n  standing is now priced at zero, so the descent offers premise questions first. Single-question\n  sensitivity cannot see that answering a premise question and a connection question together\n  would move the claim; that interaction is a known limit of one-at-a-time pricing.\n\n- **The judge measures a different number from the one the ruled rule needs** (found 2026-09-28 by\n  the Bayesian half of the literature review). The rule is Jeffrey conditioning when relevance means\n  Δ, how much more likely the target is if the supporter holds than if it fails. The judge's prompt\n  asks how far establishing the supporter would move the target and whether it would settle it,\n  which measure the direct change and sufficiency, not Δ. Which relevance number stays valid also\n  depends on how a target's reasons combine, which the graph does not record. So readings taken with\n  today's prompt should not be written until the founder has chosen what relevance means\n  (`strength-calculus-literature/bayesian-and-informal-logic.md` §§ 2.1–2.2, 6).\n\n\n- **A decision model beside the judge** (proposal, 2026-09-29): a gate at link creation, a direction\n  census over all argument links, four conditional questions per link so Δ can be computed rather than\n  asked, and a veto on writes. It never writes a reading itself, so every value reaching the strength\n  layer still carries the reasoned judge's grounds. The design-independent half: **store the two\n  conditionals rather than one relevance number**, so a later ruling on what relevance means needs no\n  re-sweep. `jev-and-the-judgement-layer.md` § 6.\n\n- **A written-out step is where this rule meets the weakest-part rule** (2026-10-08, founder\n  question). The rule under a link is either a second premise, where the August reasoning makes the\n  weaker of the two govern, or the link's licence, where this product applies; the two give\n  different strengths. And the shape of a step (ledger AD23) proposes the record the review says a\n  relevance number needs. Run on the code, unruled:\n  [the-squirrel-and-the-strength-calculus.md](the-squirrel-and-the-strength-calculus.md).\n\n- **The first calculus question, as put to the founder 2026-09-29: when a reason is refuted, should\n  its claim fall below where it started?** (Ruled 2026-09-30, for now: § 7. What follows is the\n  record of how it was put.) The reasoning, which took a detour worth keeping:\n  - **The Bayesian case for subtraction rests on the reference point.** Jeffrey's form moves the\n    claim by Δ·(q − q0), exact when the claim's baseline already prices the reason in at q0. The\n    graph's neutral point means *nothing examined*, and a link from a claim with no standing adds\n    nothing, so the graph cannot tell \"no reason yet\" from \"a reason at 50/50\". For testimony or an\n    argument someone offered, proper Bayes returns the claim to its no-reason level when the reason\n    fails, which is what drop-out does; subtraction pushes it below. For a test the claim could have\n    failed (a trial, an inspection), proper Bayes does push it below, and drop-out loses that.\n  - **The graph already has a home for evidence against: an objection.** If refuted reasons also\n    subtract, the same failure can count twice, once as the refuted reason and once as the objection\n    someone files, and the double count throws nothing. Drop-out gives evidence against one home,\n    argued and attributed, which is the 2026-08-31 principle that human force enters through\n    contributions that argue for themselves.\n  - **Subtraction cannot be made safe by relevance alone.** A relevant straw man still pays, because\n    every claim is measured from 50/50 however it was invented, and the graph has no honest starting\n    plausibility to replace it. (A worry that authors often knock down weak objections in their own\n    texts is not borne out in the graph: of 36 attacks inside one source, one is itself answered\n    inside that source, measured 2026-09-29.)\n  - **Proposed repairs**: a refuted reason still counts as the claim having been examined (today it\n    lands exactly on neutral and `evidential_weight.count_evidential_inputs` reads it as unexamined);\n    and when a refuted reason's failure itself counts against the claim, the page offers the move to\n    file it as an objection, with its reason. The one-way bet on supporters that drop-out leaves open\n    belongs to the accrual question, decided separately.\n  - **Considered and rejected**: subtracting for every link (above); a per-link mark for test-like\n    reasons that subtract (more machinery, and the same double count whenever someone also files the\n    objection).\n  - **The founder's answer, 2026-09-29, explicitly not a ruling**: *\"Keep it, with the two repairs.\n    Seems more intuitive/truthful to me. No... actually I'm not sure. Because if the implication is\n    that there's SOME truth value to a claim if it's in the middle then that'd be misguiding if\n    there's no good and relevant reasons whatsoever for believing it's true. But ofc all of this\n    depends on whether it's been examined deeply and exhaustively and it all depends on how you\n    see/judge *relevance*. Right? Btw I'm curious what is an 'objection' vs an attack link/claim, and\n    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?\"* What it opens, unruled: (a) the start point\n    carries three meanings the display must keep apart (nothing examined; nothing stands, because\n    every checked reason failed; unsettled, because reasons on both sides balance), and whether\n    *nothing stands* should lean against a claim depends on how thorough the look was, the\n    argument-from-ignorance condition; (b) an objection is a claim at the far end of an attack\n    link, and refuting a reason often asserts its negation, which can be filed as an objection, so\n    one finding can sit in the graph twice; (c) a third rule answers the mirroring directly: treat a\n    reason and its negation as one question with two weights, what the claim gains if the reason\n    holds and what it loses if it fails, so the finding counts once (a failed test weighs, a failed\n    witness or argument does not). The ontology already has this shape in the yes and no poles of a\n    critical question. It depends on recognising the mirror, which the keeper's *opposes* answer\n    could do. Worked through interactively in the page *When a reason fails*\n    (claude.ai/artifact/BbEi36mTifgXJ2avwMYQHz, private).\n\n**Measured 2026-10-08, and what it means for likeness.** No relevance claim exists on the live\ngraph, so every link's relevance is still presumed or zero. Of 1,618 live support and attack links\nbetween claims taken from texts or people, 921 were drawn by a machine between texts (871 by the\nlinking step, 50 by the cross-source premise pass) and carry no kind of argument, and 58 more from\nextraction and from people carry none either. No standing question hangs on a link without a kind,\nand the badge does not read it. On the links that do carry a kind, the questions about likeness\n(is the example representative; is there a relevant difference; is the past case alike in the\nrespect that matters) are filed as questions about the connection. So an answer to *alike in what\nmatters* is an answer about relevance. [what-a-rule-rests-on.md](what-a-rule-rests-on.md) § 2 and § 7; counts from the script in its\nfolder.\n\n**A proposal that meets this ruling (2026-10-08).** Ledger rows AD7 and AD17, written on 2026-10-07 after\nfour ways of having a machine judge a link failed in a day, say that no machine reading may weigh a link.\nAs worded they would undo the proposed tier above, and neither names it. The measurements behind them are\na reason to put the judge's question a second way, and count how many answers change, before it writes.\nWhich governs is the founder's to rule: [what-a-rule-rests-on.md](what-a-rule-rests-on.md) § 14.2.\n\n## 7. Ruled: what a failed reason does (2026-09-30, for now)\n\n**The ruling** (ledger Q34), founder: *\"but we'll go with your recommendations on this for now, and then\nwe can always poke holes in that later if we feel like it.\"*\n\n**The principle: a reason that fails drops out; failure counts against a claim only through a\ncontribution that says why.** `WEAK_SUPPORTER_SUBTRACTS` stays off. Three contributions can carry a\nfailure against a claim, each half-built:\n\n1. **A finding against it, as an objection.** When a claim that refutes a reason also bears on that\n   reason's claim (the inspection *found cracks*), the page offers to file it as an objection; the\n   keeper's *opposes* answer can rank the offers. The finding then counts once, attributed and\n   attackable.\n2. **A need, as a requirement claim** (*C requires P, because …*), which caps C at `1 − w·(1 − strength\n   of P)` (`support_semantics.requirement_weight`). Built for parts, where a constituent is presumed\n   half-needed (0.5); for reasons the presumed weight is 0, so a failed reason bites only once somebody\n   argues the claim needs it.\n3. **A thorough search that found nothing, as an argument from ignorance**: the catalogue's `ignorance`\n   scheme (question `search_adequate`) and `lack_of_evidence` (question `absence_meaningful`). Offered as\n   a move on the new display state, never proposed as settled by a machine, since judging whether a\n   search was thorough is coherent absence, which is human-reserved. This is the home for the founder's\n   *\"all of this depends on whether it's been examined deeply and exhaustively\"*.\n\nTwo repairs ride with it: a refuted reason counts as examined (`evidential_weight.count_evidential_inputs`\ncounts only inputs away from 0.5 today), and a third display state beside unexamined and unsettled,\n**nothing stands (N checked)**, so the middle never implies some truth where every checked reason failed,\nwhich was his objection of 2026-09-29. **State: ruled; all four pieces unbuilt** (`TODO.md` § Relevance).\n\n**Why, as recommended.** Read as the argued case, the badge puts the burden on a reason: it earns credit\nonce shown, and a failed one loses only what it earned (Carneades's ordinary premise; Prakken's accrual,\nwhere a flawed reason *\"does not take part\"*; reinstatement in every gradual semantics; denying the\nantecedent is a fallacy). A finding against the claim then has one home, the objection. Polarity stays\nlegible through chains: under subtraction a refuted objection raises its claim while the map still draws\nit among the challengers, and one wrong refutation deep in a chain becomes evidence against every\nancestor. And the loss side already has a home: relevance in Bayes is Δ = sufficiency + necessity − 1,\nand the ontology splits it into the gain on the link and the loss in a requirement.\n\n**Consequence for relevance.** Under this rule relevance means the gain: how much a reason adds if it\nholds, against having no such reason. The judge's question asks exactly that (*\"if the first claim were\ntrue beyond doubt ... how far?\"*), so the objection in § 6 that it measures sufficiency rather than Δ\napplies only to a subtracting rule. Writing its readings (`--write`) still moves live strengths and needs\nhis approval.\n\n**His challenge, the same hour** (ledger Q35, open), verbatim: *\"But... every claim in Deliberus will\nultimately have a subgraph which is equivalent to the already-guess in the probability/Bayes perspective -\nno? This has always been my intuition for Deliberus...\"*; *\"Yes, but remember that 'one finding can be\nwritten twice'\"*; *\"So Deliberus 'middle' is neutral ground? What does 'neutral ground' translate to in\nBayesian/probability terms?\"*; *\"I'm thinking that if I imagine myself one year hence, one year down the\nline, I'm sitting in front of my computer and using Deliberus through whatever UI or agentic interface. It\nwill already have best guesses for all the claims in the graph database. And those best guesses are based\non the subgraph and the evidence and the reasoning underneath and the flow of truth energy or whatever we\ncall it. Right?\"*\n\n**What it corrected.** Claude had explained the difference as \"probability's start contains a guess,\nDeliberus's start contains none\". Imprecise: both rules start every claim at the same value, and both give\nan unchecked reason no effect. They differ on what an unchecked *reason* is worth. Probability treats it as\na coin flip at its prior, so learning it false is news and lowers the claim; drop-out treats it as worth\nnothing until shown, so its failure removes only what it added. His intuition stands: the strengths are\nmeant to become the graph's best guesses, computed from subgraphs.\n\n**Neutral ground in probability terms** is the indifference prior, 50/50 with no evidence. It is coherent\nfor one yes-or-no question alone and incoherent across claims that exclude each other: three rival\nexplanations cannot all be 50%. The straw man lives in that gap, since an invented reason gets a 50% prior\nfor free.\n\n**The path that reconciles them: honest starting values** by kind and origin, which is the grounding\nledger Q3 already needs (base scores from the methodology ladder, located encounters with the world,\npresumptions; § 6, *Standing has to originate somewhere*). With a starting value `s` for a reason and\nrelevance Δ, Jeffrey's rule gives a gain of `(1 − s)·Δ` if the reason is shown true and a loss of `s·Δ` if\nshown false: the founder's two weights, derived from one relevance and one starting value rather than\njudged separately. An invented straw man starts low and costs little; a reason its own source asserts\nstarts higher and costs what it should; with calibrated starting values, checking a reason is zero-sum in\nexpectation, so piling on reasons stops paying. A finding that is both a failed reason and a filed\nobjection must then be matched and counted once (his *\"one finding can be written twice\"*). Until starting\nvalues exist, drop-out stays the default, which in effect gives every reason a starting value of nothing.\nIf the path is taken, relevance becomes Δ and the judge needs the necessity side too, which argues for\nstoring both conditionals now (§ 6, the decision model).\n\n**His follow-up, the same hour**: *\"Well... hmm, just thinking out loud here... but the honest prior for each claim should be the estimated truth/evidence value, which is calculated in Deliberus, no? It depends on its whole subgraph, and then that truth-guess in its turn is used to argue for other claims, and so the truth-value-calculations/estimations are bubbling up at every turn, no? Isn't this how the graph epistemic calculus was MEANT to work? Am I getting something wrong?\"* The bubbling is right and is live since 2026-09-30: a claim's computed strength is its acceptability wherever it supports another claim, recursively. The one refinement: the computed value is the result, the best guess after the claim's own subgraph, while a starting value is what the claim is worth before it. They cannot be the same number, or the claim's evidence would count twice, once as its start and once as its evidence. The honest start can still come from the graph: from where the claim came from, who asserted it and how reliable they have been, which the proposal to make sources nodes with attackable reliability would put in the graph (ledger Q4, `what-counts-as-evidence.md` § 8c). Only a claim with no origin at all would then sit at the indifference point. So the starting-value path runs through sources as nodes.\n\n**And** *\"What does 'an unchecked reason' mean in probability theory? Is it assumed that it's a RELEVANT reason as it relates to the main claim?\"* No. *Unchecked* means only that the reason's own truth is still uncertain, held at its starting value. Relevance is a separate factor: a failed reason moves the claim by its starting value times its relevance, so the failure of an irrelevant reason moves nothing under either rule, and the straw man bites only where a reason is both relevant and presumed 50/50.\n\n**Refined the same day** ([starting-values-and-the-problem-of-the-prior.md](starting-values-and-the-problem-of-the-prior.md) § 6.2, § 6.8, after research into how the problem of the prior\nis handled elsewhere): *origin* above names three things. The claim's **kind** sets its start; **who asserted\nit** is evidence rather than start (the report is not the event), carried by sources as nodes, and counting\nit into the start would stop it reaching the claims the reason supports; **who attached it as a reason**\ndecides what its failure means, because a reason attached from outside may have been chosen to fail, and a\nrefutation that was bound to be found says almost nothing. So honest starts alone do not make a straw man\ncheap, and drop-out stays right wherever reliance is unknown, not only until starts arrive (ledger Q35,\nQ36).\n\n**Considered and rejected**: subtracting for every failed reason from 50/50 (the straw man pays);\nsubtraction only for reasons the target's own source asserted (a link's meaning would depend on who wrote\nit; that objection is weaker than it looked, since relevance and requirement weights are already presumed\nby provenance, and the selection argument gives provenance a reason to matter, so the question returns to\nthe founder as ledger Q36 rather than being reopened here); a separately judged loss weight per link (on the\nstarting-value path the loss weight is derived).\n\n**Still open**: how several reasons add up (accrual), where the drift toward the side that files more\nreasons lives until starting values arrive; and the starting values themselves.\n\n## Cross-references\n\n[relevance-in-the-corpus-sweep.md](relevance-in-the-corpus-sweep.md) (the evidence for every line\nabove) · [weakest-link-arithmetic-and-the-merge-hunch.md](weakest-link-arithmetic-and-the-merge-hunch.md)\n§ 9d–9e · [drowning-in-claims.md](drowning-in-claims.md) § 8 ·\n[what-counts-as-evidence.md](what-counts-as-evidence.md) · [what-a-link-strength-means.md](what-a-link-strength-means.md)\n· [strength-layer-audit.md](strength-layer-audit.md) ·\n[pass-quality-and-the-choice-of-model.md](pass-quality-and-the-choice-of-model.md) (2026-10-05: how far\na link bears is a question of degree that no two models answered alike, so the judge's gate and\ndirection can be given to cheap models and its step to none until the question changes) ·\n[listening-for-the-calculus-and-sturdy-microtasks.md](listening-for-the-calculus-and-sturdy-microtasks.md)\n(2026-10-05: Fitelson argues on Mindscape 315 that argument strength needs two numbers, one of them\nrelevance; § 6 there says how his two differ from the two here) ·\n[how-far-a-premise-is-from-its-claim.md](how-far-a-premise-is-from-its-claim.md) (2026-10-07: no support or\nattack link on the live graph has had its relevance read or argued; a far premise and a weak one would\nget the same low number, so the proposal there is that a cross-text link carries the claim it rests on,\nminted as its relevance claim when somebody answers) ·\n[what-a-link-rests-on.md](what-a-link-rests-on.md) (2026-10-07: four machine readings of whether a link is a\nreason moved with the question, so the proposal is that no machine reading ever sets a link's bearing; a\nmachine's link stays at none until a text states its further claim or a person argues it)\n"}