{"path":"research/priors-as-standing-work.md","content":"# Priors as standing work: what a machine pass may write when it seeds a claim's answers\n\n*2026-10-24. An all-angles pass on his question. Ledger Arc BC (unruled). Tests:\n[priors-as-standing-work/cases.py](priors-as-standing-work/cases.py); research report, verbatim:\n[priors-as-standing-work/research-report-2026-10-24.md](priors-as-standing-work/research-report-2026-10-24.md).*\n\n## 0. The answer first\n\n**Most of it is already ruled, and the part that is new splits in three.**\n\n1. **The descent is ruled.** On 25 August he ruled that weighing questions be auto-opened *and tentatively\n   responded to in the graph as faithfully as possible according to our reading of the source*, with recursive\n   decomposition after every live test and, leaning, in normal use, capped by what is worth asking\n   ([kartpaus-auto-enrichment-design.md](kartpaus-auto-enrichment-design.md)). AG6 (ruled 2026-10-10) lets the\n   machine write deeper steps under quality control, AG12 makes each sentence it writes a claim nobody asserts,\n   R8 makes such work standing, and AX3 (ruled 2026-10-23) has every stance written down to its grounds. What is\n   left is building it, which he set behind the theory today (*\"revise all microtasks in light of recent rulings\n   … after theory is more set in stone\"*).\n2. **\"A sensible prior\" names three different numbers, and only one of them is a prior.** For a seeded answer:\n   its **start** (how likely a claim of its kind is before its own reasons), the **evidence of its assertion**\n   (who said it), and its link's **relevance** (how far it bears). A pass may write the first, as a claim with a\n   source; the second comes from where the answer was found; the third is read, never set by the seeding pass.\n3. **Robin's research harness is right about where rates come from and too coarse about what it writes**, as\n   he judged. Research turns up two kinds of thing: a rate for a kind of claim, and evidence about this very\n   claim. Folding both into one researched prior counts the claim's own evidence twice. The finer grain: the\n   harness writes claims (the rate, with its source; the membership of the claim in that kind; the evidence as\n   reasons), and code computes the prior from them, as the date check reads *today is Tuesday* and files the\n   reading as a reason rather than as a start.\n\nThe test run found two latent defects that a seeding pass would have set off at once, both fixed today (§ 5).\nAnd the research found the cost no strength rule prevents: a pre-filled answer steers the person who reads it,\nwhatever it weighs (§ 4.5).\n\n## 1. His words\n\n> *\"Could we populate Deliberus with sensible priors as a pass of standing work, whenever new claims arrive, or\n> even after QC questions are posed, we could automatically recursively decompose and mint preliminary starting\n> answers and their reasons/premises and make sure they get a sensible prior based on what kind of claim it\n> is... Right? Read up on this. I think I mentioned the idea during my talk with Robin yesterday or the day\n> before.\"* (2026-10-24)\n\n> *\"This was also discussed in the recent walkthrough on burden of proof and priors, read that too\"*\n\n> *\"Robin anyhow had the idea that priors could be set from a machine pass that does some research online using\n> a harness or some such. Similar to the checking of \"today is Tuesday\" by built-in date/time verification in\n> Deliberus, but with more steps. I think this is an interesting but too coarse grained idea though.\"*\n\nEarlier, typed 2026-10-21, on the fallback values: *\"that's an excellent machine pass to run regularly and\nperhaps even as a daemon, to saturate the graph with meaningful real content\"*; 2026-08-30: *\"Best priors can be\njudged according to the fractal priors kind of reasoning, right, but represented within Deliberus as\nwell-evidenced analogies/claims that express analogies, simply.\"* The walkthrough of 23 October\n([../listening/2026-10-23-who-must-prove-it.md](../listening/2026-10-23-who-must-prove-it.md), part three)\npredicted his answer from his record: *the priors should come from meta-studies and earlier analyses; they'll\ndepend on the kind of claim*. The transcribed parts of the Robin conversation\n(`~/Projects/swhisper-work/audio_in/deliberus-2026-10-23/`) hold prediction markets and common knowledge, not this\nidea; it was in a part not transcribed, as his message says.\n\n## 2. What the corpus holds\n\n| Angle | Home | State |\n|---|---|---|\n| Eager recursive descent, pre-answers from the source, draft tier, show as questions | `kartpaus-auto-enrichment-design.md` | ruled 2026-08-25 (live test mandatory, normal use leaning) |\n| The machine may write deeper steps, quality-controlled | ledger AG6; `quality-control-for-deeper-steps.md` | ruled |\n| A machine's sentence is a claim nobody asserts, weighing nothing until engaged | AG12; `machine_sentences.py` | ruled, built |\n| Standing work, three or four calls per rebuilt pass | R8; `machine-work-as-a-standing-property.md` § 7 | ruled, keeper built |\n| Depth of machine on machine capped by measured error | T1 | ruled |\n| A start is a claim from the claim's reference class; who asserted it is evidence; a model never sets a prior | `starting-values-and-the-problem-of-the-prior.md` §§ 0, 6, 7 | open (Q35, rec 1 and 8) |\n| Where a base rate exists the prior is the base rate, else 0.5 | Q46 recommendation 4; `badge_inputs.set_prior` | ruled; 0 claims carry one |\n| *Extraordinary* is a start; three burdens | `burden-of-proof-and-priors.md`; AZ1, AZ3 | open |\n| An unbacked person's answer at the lowest tier (Tavtigian) | Q3; `weighing-answers-and-the-badge-threshold.md` | built 2026-10-24 |\n| Every stance written to its grounds by the passes there are | AX3; `where-a-shared-doubt-sits.md` § 6c | ruled |\n| Pooling several kinds; empirical Bayes; robust versus prior-dependent | `starting-values…` §§ 6.6, 10.3 | open |\n\nNot covered, researched today (report verbatim beside this note): machine predictions as a capped kind of\nevidence in clinical genetics; bot-seeded encyclopedias and what became of them; pre-filled answers and the\npeople who review them.\n\n## 3. One link, written out\n\nA link on the live graph, from the minimum-wage pair:\n\n- `claim_d84d6ea4d137`: *A 1978 survey published in the American Economic Review found that 90 percent of\n  economists agreed that the minimum wage increases unemployment among low-skilled workers.*\n- supports, under the expert-opinion scheme, `claim_943b6d928963`: *There has been a long-standing agreement\n  among economists regarding the general negative effect of minimum wages on low-skilled employment.*\n- Its six latent questions, as rendered today: *The expert cited is credible as a source in the field this\n  claim belongs to* · *The expert's field is the specific one this claim belongs to* · *The expert actually\n  asserted what is attributed to them* · *The expert is personally reliable and trustworthy on this subject* ·\n  *The claim is consistent with what other experts in the field assert* · *The expert's assertion rests on\n  evidence rather than impression*.\n\nWhat a seeding pass would write for the fifth question, and where each number lives:\n\n- **Found in the source**: whatever the source text says about other economists is the author's claim, filed as\n  an answer with extraction provenance and the author attributed (the source's own *because*, Q36). It weighs as\n  the author's answer, and one source is one origin (Q4), however many of its sentences answer.\n- **Found by research outside the source**: later surveys of economists disagree with 1978 (Claude's recall,\n  unverified here: the IGM panel's split on a federal minimum of $9, and the Card and Krueger study of 1994). That\n  is evidence about this very claim, so it enters as an attack on the yes, with its citation, and is never part\n  of any start.\n- **Made by the machine**: a sentence the machine writes because no text states it (*economists' agreement has\n  weakened since the 1990s*) is a claim nobody asserts (AG12). It is a place to disagree and weighs nothing.\n- **The start**: what kind is *a reported survey finding about expert agreement*? No measured rate exists for it.\n  Pooled to its nearest measured kinds, it starts at 0.5, which is the honest result for most claims (§ 4.2).\n\n## 4. Theory\n\n### 4.1 Three numbers, one word\n\nA Bayesian reading separates them and Deliberus already stores them apart: the start (`Claim.start`, through a\nreference-class claim), the assertion's evidence (the Q3 tier for a person's answer, extraction's presumption for\na source's), and relevance (`relevance.read_relevance`: argued, a confident reading shrunk by the judge's record,\nor the writer's presumption). His *sensible prior based on what kind of claim it is* is the first. The danger in\nthe proposal is the drift from the first to the other two: a pass that gives a seeded answer *a sensible\nweight* has graded its own homework, and does so at machine volume.\n\n### 4.2 What a kind can tell today\n\nThe kinds a claim carries in the graph are its type (empirical 1,969, normative 416, definitional 160, value\npremise 115) and its epistemic status (observed 1,239, causal hypothesis 494, normative assertion 494, …). None of\nthese has a measured base rate. The kinds with rates are narrow and come from fields that count outcomes: a\nhypothesis in a top psychology journal is true about 9% of the time, a drug entering trials is approved 13.8% of\nthe time, a large project finishes on time and budget 8.5% of the time (`starting-values…` § 10.1). So a pass that\ngives *every* claim a start by kind would give 0.5 to nearly all of them today; the real work is finding which\nnarrow reference class a claim belongs to and the rate for that kind. That is research, and it is where Robin's idea earns\nits place.\n\n### 4.3 Robin's harness, at the finer grain\n\nWhat research returns has two shapes, and they go to different places:\n\n| Research finds | It is | It goes |\n|---|---|---|\n| *About 9% of hypotheses in top psychology journals are true (Dreber et al. 2015)* | a rate for a kind | a claim with its source; the start is computed from it |\n| *This claim is a hypothesis tested in a top psychology journal* | membership in a kind | a claim of its own, attackable (a strategic contributor will pick the class that suits them) |\n| *A replication of this very study failed* | evidence about this claim | an attack, with citation; never in the start |\n\nA start that one researched number sets from all three counts the claim's own evidence twice, once in the start\nand again when someone attaches the replication. That is the coarseness he heard. Written as claims, the start is\narithmetic over contestable parts (pooling across kinds, `starting-values…` § 10.3), and his 2026-08-30 sentence\nfalls out: a kind is an analogy (*this claim is like those*), represented as a claim that can be argued. The date\ncheck is the same move at its smallest: the system reads the date and files *today is Tuesday* as a reason with an\noracle's provenance; nothing about Tuesday enters a start.\n\n### 4.4 An answer's own kind is its question\n\nFor a seeded answer the most useful kind is not its type but the question it answers, with its pole: *the\nexpert actually asserted what is attributed to them, yes*. Over time the graph holds how often examined answers\nto that template held, and how much a person's yes to it moved what it guarded. That is how clinical genetics set\nits tiers. The ACMG rules of 2015 capped computational predictors at the lowest tier, counted all of them as one\nline of evidence because they share inputs (agreeing tools were wrong together: 2.8% to 25.8% *false concordance*\nfor the best trios, Ghosh et al. 2017), and only after calibration against outcomes let four tools reach *Strong*,\nfor 1.4–1.7% of variants (Pejaver et al. 2022). The calibration also shows that the lowest tier's size depends on\nthe prior it is read from: at Tavtigian's prior of 0.10 the supporting factor is 2.08, at Pejaver's 0.0441 it is\n2.41, which from 0.5 gives 0.675 against 0.706 (recomputed in `cases.py`). Deliberus's lowest tier uses the first.\nSo Q3's clause *a kind of answer earns a higher default only when its diagnosticity is measured* has a worked\nprecedent, and *several machine passes agreeing are one witness* is the canon line already resident.\n\n### 4.5 Weighing nothing does not mean harming nothing\n\nThe strength rules keep a machine's answer from moving a number. They do not keep it from moving the person\nwho reads it, who then gives the answer that counts. Measured outside: pre-highlighted labels raised full\nagreement among five annotators from 8% to 38% (Schroeder et al. 2025, 410 annotators); a wrong LLM label lowered\npeople's accuracy even when the model was right over 80% of the time (Wang et al. 2024); reviewers caught 82% of\nsurface errors and 31% of errors needing conceptual judgment (Beck et al. 2025), and answers to critical\nquestions are the conceptual kind; clinical advice right 80–90% of the time still switched 6–11% of correct\ndecisions to wrong ones (Goddard et al. 2012). The kartpaus ruling already shows a machine pre-answer as a\nquestion (*our reading of the source suggests — confirm or correct?*); the evidence says that is load-bearing,\nand that the person's answer box should open empty, with the machine's reading folded behind it.\n\n### 4.6 Volume\n\nThe six expert-opinion questions on one link, answered once each, mint six claims; each answer carries its own\nscheme and questions. From the 2,511 extracted claims, a first level is on the order of 15,000 claims, a second\n90,000, a third half a million (rough, `cases.py`). The worth-asking stop rule is a weak brake on fresh material\n(it prices 595 of 1,198 open questions as worth asking, `quality-control-for-deeper-steps.md` § 4.7), and T1 caps\nhow deep machine work may build on itself. The precedent outside: Lsjbot made 99.1% of the Cebuano Wikipedia's\n5.4 million articles, which 257 active people now keep and which are rarely improved; the Swedish Wikipedia,\nwhere Lsjbot wrote about half the articles, fell from about 3.8 million to 2.5 million in cleanups from 2017 to\n2022; Rambot's 30,000 English town stubs of 2002 were largely expanded, in a wiki with a large community.\nSeeding is paid for by whoever maintains what was seeded. In Deliberus that is the standing passes, which is why\ntheir error rates come before their volume.\n\n## 5. Tests on the live strength code\n\n`cases.py` builds the minimum case (an extracted support link under the expert-opinion scheme, one premise\nquestion) in a throwaway graph and adds answers of different origins. Before today's fixes:\n\n| Case | Strength of the supported claim | Weighed answers | Badge shown |\n|---|---|---|---|\n| no answer | none (unexamined) | 0 | no |\n| one person's yes | 0.501 | 1 | no |\n| the source's own yes (extraction, no account) | 0.501 | 1 | no |\n| a machine's yes | none | 0 | no |\n| a machine's yes with a confident relevance reading on its link | **0.501** | **1** | no |\n| ten machine yeses, each with a confident reading | **0.542** | **10** | **yes** |\n| ten source-quoted yeses | 0.542 | 10 | yes |\n| a stated prior of 0.8, one unexamined link | **0.800, CREDIBLE** | 0 | no |\n\nTwo defects, both latent (no claim carries a start, no link carries a confident reading), both set off by the\nfirst run of any seeding pass, fixed today with tests in `tests/test_answers_weighed.py`:\n\n1. **A machine's answer weighed once a confident relevance reading sat on its link**, against AG12. The answer\n   weighing now skips an answer of machine provenance unless a person has argued its relevance.\n2. **A stated prior made a claim read as examined.** Examination was read against the claim's prior, while link\n   strengths sit on a scale whose neutral is 0.5, so an unexamined link at 0.5 counted as evidence beside a start\n   of 0.8 (the unit mismatch the near-miss rule warns of). It is now read against 0.5. A claim with a start and\n   nothing examined reads UNEXAMINED.\n\nLeft as it is, for his ruling: ten yeses quoted from one source each count, because an answer with no account\ncounts as its own (§ 6, BC2). The full suite passed after both fixes (3,581 passed).\n\n## 6. Recommendations (proposed, unruled; ledger Arc BC), in order of dependency\n\n1. **BC1, a seeding pass writes a start and nothing else numerical.** When a pass gives a seeded claim *a\n   sensible prior*, it writes only the claim's prior, as a reference-class claim with its source; it never sets an\n   answer's weight or a link's relevance, which come from where the answer was found and from the relevance\n   reading. This is `starting-values…` recommendations 1 and 8 applied to the seeding pass.\n2. **BC2, a seeded answer is filed by where it was found.** Found in the source: the author's answer, attributed,\n   extraction provenance, and one source counts once per question (Q4's one origin, decided per question now\n   rather than waiting for sources as nodes). Found in another text: that text's claim, with its citation. Made by\n   the machine: a claim nobody asserts (AG12), weighing nothing until a person engages.\n3. **BC3, the research harness writes claims, never a prior.** Robin's idea at the finer grain: a rate claim\n   with its source, a membership claim, and evidence about this claim as reasons; code computes the prior by\n   pooling. Like the date check, an oracle's or a search's finding is filed as what it is.\n4. **BC4, an answer's kind is its question and pole, and a kind earns above the lowest tier only when measured.**\n   The graph records, per question template, how examined answers held and how far a person's answer moved what\n   it guarded; until a template has a record, every answer to it sits at the lowest tier, and several machine\n   passes agreeing count as one (the ACMG precedent).\n5. **BC5, a seeded answer is shown as a question with the source's words, and the answer box opens empty.** The\n   kartpaus display rule, with the reason now measured: a pre-filled answer steers the person, and conceptual\n   errors are the ones accepted.\n6. **BC6, the order.** The theory in rows BC1 to BC5, then his revision of every pass's instructions (named\n   today), then the seeding pass as standing work, depth capped by T1, three or four calls per rebuilt pass (R8),\n   passes off in production until he says otherwise.\n\n## 7. What only he decides\n\n- BC1: whether *sensible prior* means the start only (Claude's pick), or may also set a seeded answer's weight.\n- BC2: whether a source's answers to one question count once.\n- Whether normal-use seeding runs at all before a person has read the first level (the kartpaus lean; T1).\n\n## 8. What a kind is, and what makes a claim extraordinary (later the same evening)\n\nHis words, in three messages:\n\n> *\"Good idea in a way but that's not the only thing that determines the kind of claim it is. Which is the only\n> thing that should determine its prior. The model pass calibration/evaluation is not reasonable to reduce the\n> whole claim kind base rate to, in order to arrive at a prior. The claim says what it says and before ANYTHING\n> else it has a prior, for anyone reading it wherever it appears. But hey, what does \"kind\" mean when we're\n> talking about priors exactly? Is there a hierarchy?\"*\n\n> *\"When I introspect about how I go about judging this in my mind's eye, I guess my mind immediately tries to\n> work out all the implications of what would change if that claim were true. How strange would it be that it\n> were true?\"*\n\n> *\"don't call it a \"start\" please be exact and call these priors. At least here in the sessions.\"*\n\n**The prior is read from what the claim says, for anyone, wherever it appears; who wrote it never enters\nit.** The remark this answered had proposed one factor per writer (how often its links turn out right). That\nfactor belongs to whether a link holds, the presumption on a link's relevance, never to a claim's prior.\n\n**A kind is a reference class**: claims like this one whose truth rate somebody has counted (a hypothesis\ntested in a top psychology journal, a drug entering first human trials). Kinds form a hierarchy and also\ncross: nested reference classes pool, the narrow reference class's rate pulled toward the broad kind's by how little data it has, and\ncrossed kinds with no data on their overlap stay a range (`starting-values-and-the-problem-of-the-prior.md`\n§ 10.3). Which kinds a claim belongs to is itself a claim, open to attack. The grants session's ProbLog trial\nthe same evening combined two reference classes by noisy-or, which assumes the two are independent\n(`tools-for-checking-inferences/trials.py`): the choice pooling makes in the open, made silently.\n\n**His introspection is a second road to the prior: not how often claims of this kind are true, but how much\nelse would have to give if this one were.** It is Hume's *greater wonder* and Quine's *minimal mutilation*\nof the web of belief. Where a kind with a counted rate exists, the two roads meet: claims of that kind are\nrarely true largely because they would overturn much that is well established.\n\n**In Deliberus the implications are not part of the prior, because they are claims and links.** A new claim\nthat contradicts established claims is attacked by them, and those attacks weigh by the established claims'\nown strength. Folding the same clashes into the prior as well would count them twice. So *extraordinary*\ncomes apart into two inspectable things: a low prior from the claim's reference class, and strong established claims it\ncontradicts, written as attack links. H. pylori shows why the split matters: the claim it contradicted\n(*stress causes ulcers*) was widely held and weakly supported, so the clash weighed less than it felt.\n\n**This is the most valuable thing a seeding pass could write**, which brings the evening back to his first\nquestion: for a new claim, what else would have to be true or false if it held, each written as a claim with\nits link. The weight then comes from claims people and texts already back, never from the machine. Proposed\nas ledger BC7, open.\n\n## 9. Who says what kind a claim is, whether every premise must be mapped, and one home per consideration (later still)\n\nHis questions, in two messages: *\"But for priors - how do we, in practice, assert what kind a claim is? Is that\ndone by code or a model and how can we rely on it enough? Is this somehow a solved problem that I've just not\nyet understood?\"* and, on relevance, priors and implications: *\"they BOTH have to represent a super-finegrained\nunderstanding of any and all premises affecting the truth of the claim that could jeapardize its eventual\nstrength, right? … All tenets must be mapped and understood, hypothetically, for relevance to be accurately\njudged, and for implications to be as well, and to some extent for the prior, ideally. Or am I confused\nhere?\"* Then: *\"So then would regular claim-argumentation and relevance-claim-argumentation and\nimplication-argumentation in Deliberus overlap and cause duplicated points of entry for interaction with the\ngraph? … But maybe it's not a problem because we can treat it the way we would a rule/claim that recurs in\nmany different arguments … And same for shared-roots of evidence and testimony or polls in the case of\npol-sci experts vs prediction markets etc.\"* He ruled BC1 and BC7 together and gave the one-level direction\nrecorded under BC6.\n\n### 9a. Saying what kind a claim is: half solved, half the oldest open problem\n\nThe problem has two halves (Paul 2026, `starting-values-and-the-problem-of-the-prior.md` § 10.2). **How rates\nfrom several kinds combine is solved**: partial pooling, plain arithmetic in code. **Which kinds a claim belongs\nto is the reference class problem**, open in philosophy since Venn and Reichenbach and handled in every working\nfield by making the choice visible rather than by solving it. For Deliberus, four routes, in order of trust:\n\n1. **Code reads the kind where the claim's structure states it**: a study design (GRADE starts certainty by\n   design), a statistical generalization's stated proportion (AT8), a claim's type and epistemic status. No model call.\n2. **A model proposes membership in a named kind, with the source of the kind's counted rate**, written as a\n   membership claim anyone can attack (BC3). It never writes the rate or the prior; code pools.\n3. **Pooling makes a wrong narrow reference class cheap**: a kind with little counted data is pulled toward its broad kind,\n   so a mistaken membership moves the prior by little.\n4. **The robust or prior-dependent reading shows where the choice matters** (`starting-values…` § 6.6):\n   recompute the claim under each reasonable kind, and say so where the conclusion flips.\n\nSo reliance is bounded rather than assumed: the prior's influence shrinks as evidence on the claim arrives, it\nbites only where evidence is thin, and there the badge's weight (the weight beside the number, Q41, ruled) says\nhow thin. Membership proposals are a machine pass like any other, proven one case at a time first (the\nprinciple under Y9).\n\n### 9b. Must every premise be mapped? No for calibration, yes for resolution\n\nA rational Bayesian reads a prior or a relevance as an expectation **given what is known**, averaged over what is\nnot. Unmapped premises do not make that estimate wrong; they make it coarse. The forecasting literature's\ndecomposition of a score (Murphy 1973) separates two goods: **calibration**, whether things said at 70% happen\n70% of the time, which a coarse reference class can already deliver; and **resolution**, how well the estimates\nseparate the cases that hold from the ones that fail, which only finer structure buys. His invention case: *of\ninventions that work as a prototype, what share fail at scale for a reason nobody foresaw* is a kind with a rate,\nso the unforeseen quirk is priced on average before anyone names it. Naming it (`what-relevance-is.md` § 3b: full\nrelevance, near-zero doubt, until the decade raised the doubt) is what lets the graph tell this invention from\nthe ones that work. So he is not confused about the aim: the fine-grained graph is how resolution grows. The\npoint he can drop is that a coarse estimate is inaccurate; it is honest if its weight is shown.\n\n### 9c. One consideration, one place, many uses\n\nHis hunch is the corpus's shape already: **a claim has one address and many uses** (AL1, one home per issue;\nAL4, the rule is referenced, never restated). The three kinds of argument he names are not three stores but\nthree places one claim can stand:\n\n| Where a claim stands | Example | How it counts |\n|---|---|---|\n| a reason for or against a claim | *stress causes ulcers* attacking *H. pylori causes most ulcers* | its own strength × that link's relevance |\n| a ground or defeater of a link | *the gate is closed on public holidays* attacking *today being Tuesday bears on the gate being open* | lowers that link's relevance claim |\n| an implication (BC7) | the same *stress causes ulcers* | an ordinary attack link; no channel of its own |\n\n**An implication is not a new kind of argument.** *If H. pylori causes most ulcers, stress does not* is a link,\nand the established claim enters as an attack, so BC7 adds claims and links, never a third mechanism. Which of\nthe first two places an objection goes is AL6's question (*granted the reason, does the objection still\nstand?*). Duplication arises in exactly three ways, each with a home already: **the same consideration in two\nplaces** (AL6 picks one); **one root behind several reasons**, the polls and pundits who read the same polls,\nthe experts and markets who read the same experts (the independence reading Q40, the claim between Q42, a\nreported belief written to its basis AX3, sources as nodes Q4); and **a coarse number that also contains what a\nlink already says** (§ 8's double count).\n\n**The rule that ties §§ 8 and 9 together: a consideration counts once. While unwritten it lives in a coarse\nestimate (a prior from the reference class, a relevance reading); once written as a claim it leaves the estimate and counts\nthrough its links.** So a prior and a relevance reading are each made given what is already written under the\nclaim or the link, and the relevance judge must read a link with its written defeaters in view. Proposed as\nBC8; ruled 2026-10-24 and made resident in the Principles the same night (ledger BC8).\n\n## 10. How a consideration counts once in practice, and what goes where (later still)\n\nHe ruled BC8: *\"Great, ruled, I like this direction a lot. But \"must review existing text\" how exactly? Tell me\nmore. And then walk me through an example please.\"*\n\nHis words that set the rest of this section, from the same message (verbatim, kept because §§ 10c to 10f answer them):\n\n> *\"What happens if a 12-year old submits \"that doesn't matter\" as a rebuttal/attack on a claim, instead of clicking in\n> to argue about the relevance-claim? How do we deduplicate stuff like this? But the purpose of the separated-out\n> relevance-argumentation I guess is to put a place in our object model/ontology for that act/interaction/type of\n> argumentation/disagreement, the same as when the splitter splits messy longer sentences into their parts, so anyone\n> can argue/address each part in its own place. Any PART of the relevance can also be argued in one place, then. And\n> any part of the implications of a strange new hypothesis submitted, can be argued separately. This is something I've\n> always strived for, and it's a guiding light.\"*\n\n> *\"Deliberus needs to be as fine-grained and Vulcan and autistic as possible behind the scenes, it needs to build up\n> this mind like a crystallized lattice in the DB/brain/engine itself... but ALSO be extremely user-friendly and have a\n> super intuitive and non-cluttered and mentally soothing and psychologically savvy user interface and experience.\"*\n\n> *\"A positive claim and a negation have to be marked as mirror images of each other somehow, or merged/clumped up,\n> somehow, I feel\"*\n\n> *\"These parts are important, but they live, I've realized, on a higher plane of abstraction from the parts of\n> Deliberus that we've been and are in the midst of setting in stone … In order for the what-if-simulations and\n> everything attunement/worldview-lens-related to work correctly/optimally, we need all the foundations settled and\n> stable. At least this is a hunch I have, feel free to prove me wrong.\"* (on attunement, worldview lenses and what-if)\n\n> *\"We should start to collect input from people using a little (+) button integrated in the new map view … because\n> then we'll learn more about how people actually engage, and I think the ontology is starting to be mature enough to\n> allow this to take place now/soon\"*\n\nOn calibration against resolution (§ 9b): *\"Right, true.\"*\n\n### 10a. BC8 in practice: the estimate is of what is not yet written\n\n**A relevance reading.** The gate: *today is Tuesday* supports *the gate is open now*, under *this gate is open\nevery Tuesday*. A reader judging that link with nothing written under it might say 0.9: Tuesdays usually open\nthe gate, and the 0.1 prices, on average, the holidays, repairs and lock failures nobody has named. Now a\nperson writes *today is a public holiday* with *the gate is closed on public holidays*, attacking the link's\nrelevance claim. If the reading stays at 0.9, holidays count twice: once inside the 0.1, once as the written\nattack. So the reader is handed what is already written under the link (its relevance claim's parts and the\nattacks on it) and asked for the rest: *setting holidays aside, which are argued below, how far does a Tuesday\nbear on the gate being open?* Say 0.97; the written attack then lowers it as far as its own strength goes.\nIn code: the relevance judge's inputs carry the link's written defeaters, and its reading is stored with the\nlist it was given, so a defeater written later marks the reading outdated (the standing-work queue already\ntracks readings by version).\n\n**A prior.** *H. pylori causes most stomach ulcers*, 1982. If its kind is read as *a claim that contradicts the\ntextbook consensus in medicine*, the kind's low rate already contains the clash with *stress causes ulcers*,\nand BC7 writes that clash again as an attack link. So **a kind may not be described by a relation to another\nclaim that the graph holds as a link**; it is described by what the claim says (*a bacterial cause proposed for\na common chronic disease*), and the clash counts once, through its link. The membership pass is told which\nclaims are already linked to the claim and refuses a kind defined by them.\n\nDraft resident line, for his blessing: **A consideration counts once: while unwritten it lives in a coarse\nestimate (a prior from the claim's reference class, a link's relevance reading), and once written as a claim it counts\nonly through its own link, so each estimate is made given what is written beneath it.**\n\n### 10b. BB4, explained (already ruled)\n\nA real case from a friend's map of AI governance. A link: *researchers left their jobs saying the risk is high*\nsupports *the risk is high*. The author adds: *sincere alarm is evidence of belief, not of probability*. He does\nnot dispute that they left or what they said; he disputes how far their saying it bears on the risk. BB4 files\nhis sentence as an attack on that link's relevance claim (*their leaving saying so bears on the risk being\nhigh*), never on *researchers left saying the risk is high*, which stays true and unweakened. Filed on the claim\ninstead, it would weaken a fact the author accepts and leave the inference he doubts untouched.\n\n### 10c. One home per act, many doors\n\nHis standing aim: *\"a single source of truth/disagreement on each claim/question/rebuttal/case/exception/consideration,\nand a single correct way to channel every type of interaction from either users or agents using Deliberus\"*.\nThe sameness pass dedups claims after the fact; it does not route acts. Routing is the other half: **each kind\nof act has one home, and every door that accepts it files it there** (AL1, open, made a rule about acts as well\nas issues). A twelve-year-old who types *that doesn't matter* under a claim's reason has made a relevance attack\nthrough the reason's door; a routing step asks AL6's question (*granted the reason, does the objection still\nstand?*: yes) and files it on the link's relevance claim, saying so to the person in plain words. The door\nstays wherever people reach for it; the home is fixed.\n\n### 10d. What goes into language and what is lifted out\n\nA test that sorts every case met tonight: **lift a thing out of the sentence when code must compute or compare\nit** (a link's kind and direction, a rule's kind and stated proportion, a hedge, an attribution, a time), and\n**keep it in language when people dispute its content**. Where a lifted thing can itself be disputed, it gets a\nclaim twin anyone can argue: the relevance claim for a link, a membership claim for a kind, the hedge kept as the\nauthor's stance. This is the storage law proposed as AA1 (open: uses, judgments and readings, never a reading\nstored as fact). Prior art worth a research pass: the Argument Interchange Format and Carneades on argument\nstructure, Wikidata's statements with qualifiers and references, nanopublications, legal XML. Not run.\n\n### 10e. Negation as a mirror\n\nAL2, ruled for now, says a defeated attack does not establish its opposite, and *opposes* covers contrariety\n(*the wall is red* against *the wall is blue*) as well as contradiction (*the wall is red* against *the wall is\nnot red*). His intuition that a claim and its exact negation should be marked as mirror images is the\ncontradiction half: proposed, not filed as a row, a keeper answer or a code check that marks exact negation\napart from contrariety, so the pair can share one home with two poles (`one-challenge-one-home.md` § 4).\n\n**RULED 2026-10-25** (ledger BC9, rival explanations share one question), founder: *\"Rival explanations set ideas: ruled and blessed!\"*. Rival explanations of one fact are answers to one question, with a *something else* answer, their priors sharing one whole; the exact negation is the two-answer case of that set, marked apart from a contrary. Unbuilt.\n\n### 10f. The higher plane\n\nHis hunch, that the what-if and worldview views sit above foundations still being set, matches the corpus: the\nwhat-if view is one computation over the same strength rules (`what-if-and-lenses.md`, AU5 puts the strength\nrules first). The (+) button on the map, his first step toward people adding to maps, is in `TODO.md` under the\nlanding page and is an interface change for his blessing (AR14).\n\n**2026-10-24, later: how the membership pass is built and trusted** is its own note: [saying-what-kind-a-claim-is.md](saying-what-kind-a-claim-is.md) (ledger Arc BD).\n\n## Cross-references\n\n`kartpaus-auto-enrichment-design.md` (the ruling this builds on) · `starting-values-and-the-problem-of-the-prior.md`\n(the start) · `burden-of-proof-and-priors.md` and its walkthrough (the extraordinary start) ·\n`weighing-answers-and-the-badge-threshold.md` (the lowest tier) · `quality-control-for-deeper-steps.md` (the\ncontrols) · `machine-work-as-a-standing-property.md` § 7 (the standing plan) · `drowning-in-claims.md` (volume) ·\nledger AG6, AG12, R8, T1, Q3, Q4, Q35, Q36, Q46, AX3, AZ3.\n"}