{"path":"research/where-a-shared-doubt-sits.md","content":"# Where a shared doubt sits: the claim between reasons, against the object model and the two folds\n\n*2026-10-23, late. Analysis of ledger Q42 (a shared doubt becomes a claim between) against what is\nruled about the graph's objects, and against the same evening's work on folding and zoom (ledger AX8,\nruled; AX9 to AX12, open). Q42 was ruled the same hour with the clauses of § 5 and one exception of his: a machine-written middle carries its cases' strength from the start.*\n\nHis questions, verbatim:\n\n> And what's been said throughout all those different sessions talking about this, shared roots of\n> evidence, multiple witnesses having lunch, shared reasons deeper in the graph etc? What have I said in all\n> those sessions about this, or what's been mentioned in the walkthroughs about it?\n\n> I want to move ahead on this, I'm just worried if it'll fit with all other thinking about the\n> structure/object-model of the graph. How would it fit into everything else that's been decided? Look at the\n> other session from tonight where I'm discussing folding & zooming. Does it relate to this? Think deeply\n> about it.\n\n## 0. In short\n\n1. **Q42 fits without a new kind of node or link.** It is a pattern made of a claim, support links and a\n   relevance claim, all of which exist (§ 3a).\n2. **Its middle claim already has a ruled home: the case channel.** The lawn, the roof and the chairs are\n   cases of *everything left outdoors got wet last night*, filed under it and counted once (ledger AY2); that\n   claim then supports *it rained* by one link, whose rival question (*was it dew?*) sits on that link's\n   relevance claim. A shared doubt splits one inference into what was observed in common and what explains\n   it (§ 3b).\n3. **It completes a typology the corpus already had.** The walkthrough of 15 October named three ways two\n   arguments are secretly one (rewording, content, shared origin); a shared rival is a fourth, and each of\n   the four has one home in the object model (§ 3c).\n4. **Moving the reasons under the middle is a rehoming**, the same operation as the cut whole's (ledger\n   AL15, AX12), under the same rules for machine and human links (§ 3d).\n5. **The folding work had already assumed it.** AX3 is Q42 applied to reported beliefs, AX2 reads it by code,\n   and AX9's table of what a card may join needs one more row for it (§ 4).\n6. **Two tensions, both resolvable**: the middle must be one proposition at the floor (AX8), and a middle a\n   machine writes weighs nothing until examined (AG12), so the old links must keep counting until then (§ 3e,\n   § 5).\n\n## 1. What he has said, in order\n\nFrom his typed prompts (`mannaminne --said`, `~/.claude/history.jsonl`) and the ledger:\n\n- **2026-07-06**: *\"Merging/dedup HAS been considered deeply and reflected on in earlier docs, I believe?\"*\n- **2026-08-25**, on sameness groups: *\"we will need to be super careful not to count evidence twice but at\n  the same time ensure the grouping collects/holds/consolidates everything associated with the\n  other-identical-claims of the identical-group, yeah? But come to think of it, if the sameness/identity\n  deduping/merging works well, we will only ever have to worry about nodes/claims directly connected to (only\n  a single step away from) the first claim/identical-claim-group, since everything upstream/deeper will be\n  the same, yeah?\"* (the one-step invariant)\n- **2026-09-13**: *\"How can we ensure that sameness-merge is always considered first? … I feel like it's\n  important to consider sameness first so it doesn't confuse us/the ontology?\"* (sameness before relation,\n  ruled the same day)\n- **2026-09-22**, on *\"parts combine by weakest link … evidence accumulates (three witnesses beat one)\"*:\n  *\"Is this automatically sensed and tagged these days or not yet?\"*\n- **2026-10-14**, to Codex (recorded in ledger Q4): *\"Overlapping roots of evidence obviously shall not be\n  added up.\"* The same day: *\"reasons pass on their strength in whole to the claim they argue for … Whether\n  there's 1 or 14 reasons\"*, and *\"The transitive inherited strength should not let it make any difference\n  how many vertical steps the chain of a link splits into, if all that is relevant and true gets passed on up\n  the chain.\"*\n- **2026-10-21**, to Robin, on the alarm contacts: *\"De måste vara premisser där en till premiss\n  (regel/warrant) är 'både kontakt A och B måste vara slutna för att larmet ska gå'\"*; and on reading a line\n  of reasoning by its weakest piece: *\"Yes, of course!!!\"*\n- **2026-10-22**, the question Q42 answers: *\"How does one know how many independent arguments/reasons can\n  conceivably belong directly under a claim? A claim with four independent reasons, what would that look like\n  compared to a claim with just one? Could we be sure nothing is missing for either? How does this affect the\n  adding/multiplication?\"* Then the ruling of the independence question (Q40): *\"Rule the third wording for\n  the independence check now: suppose the claim is false, what is each reason's most likely explanation, is\n  it the same one.\"*\n- **2026-10-23**, reading the markets-and-experts case: *\"The experts are part of the \"market\" generally\"*\n  and *\"The inferential distance between the claim and the reasons here feels like a bit too much of a leap.\n  We need to be responsibly aware at least of all the inferences between them.\"* On the machine's answer\n  listing no premises: *\"This seems like a core weakness of LLM usage for Deliberus' microtasks, no? But\n  humans ofc make this mistake often as well, they almost don't notice their leaps through multiple inference\n  hoops... Especially when victims of motivated reasoning...\"* (`how-independent-reasons-add/his-readings.md`)\n- **2026-10-23**, on the cases rows: *\"Both blessed! But I'm not sure I FULLY grasp all this yet.\"* (BA1, BA2)\n\n## 2. What the walkthroughs said\n\n- **26 September, what counts as evidence**: agreement multiplies only across independent errors\n  (Condorcet); two reports with one origin can be worth barely more than one; but *diverse beats repeated*\n  is false in general, since two agreeing reports from one instrument can show the instrument is reliable\n  (Bovens and Hartmann), which is why the August rule made a known shared source a node, never a blind\n  discount.\n- **8 October**: a machine that drew a link, wrote its rule, imagined the case against it and checked it is\n  *\"one witness, asked four times.\"* **13 October**: *\"An echo is not a second witness.\"*\n- **14 October, parts, reasons and what the number means**: a correction, *\"Fourteen copies of one\n  observation are one. Fourteen different readings from one thermometer are fourteen, and they share a\n  weakness. Dependence is not identity.\"*; and drawing an extra box between two claims multiplies nothing,\n  since multiplying is a decision about the world, not the drawing.\n- **15 October, the ropes and the coin, part five, *Two witnesses who had lunch***: two witnesses who had\n  lunch before the trial are one witness; two newspapers copying one wire report are one source. Three ways\n  two arguments are one: rewording (the keeper's *same*), dependence in content (its *contains* and\n  *overlaps*), and shared origin, which no text shows and which belongs to provenance, sources as nodes\n  (Q4), quoting his *\"overlapping roots of evidence obviously shall not be added up.\"*\n- **20 October, one argument counted twice**: the alarm contacts, one joint argument counted as two by the\n  reasons splitter; the double count waits on how a joint argument is drawn.\n- **22 October, what an argument is made of**: the ruling (weakest premise within an argument, best across\n  until independence is read, then add); the objection *\"winning by volume\"* answered by rewordings and one\n  source counting once.\n- **23 October, four pieces**: witnesses at different windows multiply, a press release in two papers counts\n  once, and the decision Q42 put with the magnesium trials (*Why the parts of the strength calculus fit*); the\n  independence reading and the proposal that a reported belief be written down to its basis, AX3 (*Where\n  dependence hides*); a bias accusation counting heads where it should count origins (*Who must prove it*);\n  cases counted once under their rule (*What a case counts for*). The elder twice: *\"A crowd of witnesses is\n  worth what their separate eyes were worth. Ask where each one stood before you count them.\"* and\n  *\"Agreement is cheap when it comes from a single source. Ask of every chorus how many throats it has.\"*\n\n## 3. How Q42 fits the object model\n\nThe example, as it would sit in the graph. Today:\n\n- *It rained last night* ← *The lawn is wet* (support link) and ← *The car roof is wet* (support link).\n\nDrawn as Q42 proposes:\n\n- *The lawn is wet*, *The car roof is wet*, *The garden chairs are wet*: cases, filed under\n- *Everything left outdoors got wet last night*, which supports, by one link,\n- *It rained last night*; under that link's relevance claim, *It was not dew*.\n\n### 3a. Nothing new in kind\n\nA claim, support links, a relevance claim with something under it. No node type, no link type, no stored\nnumber. The rival's denial answers the abductive scheme's own critical question on that one link (`alternatives_eliminated`:\n*alternative explanations for the observation have been adequately considered and found wanting*), latent until\nsomeone answers it (latent scaffolding, built 2026-09-21).\n\n### 3b. The middle is a generalization, so two ruled channels meet\n\n*Everything left outdoors got wet last night* is a claim about many things, each of which is a case (ledger\nAT11: a claim's scope is read once; this one is many cases, normic, below). The case channel, ruled (AY2: a case\ncounts once, under its rule; AY3: a rule's number from its count), carries the lawn, the roof and the chairs\ninto it. The reasons channel (Q38, Q46) carries it to *it rained* as one reason. So the shared doubt divides\none inference into two of different kinds: **what was observed in common** (an induction, counted) and\n**what explains it** (an abduction, one link). The rival question (Q40's wording) then has exactly one home,\nthe relevance of the second step, instead of hiding in two parallel links.\n\nThe numbers come out right: with dew on three nights in ten, the lawn alone gives 76 %, the two summed 91 %,\nthe exact answer 79 %, which is what the middle passes on once its link's relevance is weighed against dew\n(`reasons-and-cases.md` § 5.9). Built today: the reasons' log-odds (Q46). Not built: the case channel's count,\nand a relevance computed from a stated rival's rate (the relevance judge proposes a magnitude, Q33).\n\n**A rule, in other words?** His question the same night: *\"A rule, in other words?\"*, and then: *\"with possible exceptions\nalso, read up on this in recent sessions\"*. Yes, in the ruled sense: a claim about many cases is a rule (AT11, AT12). Its\nkind is read by the scope pass, never set by hand (AT11), and the likely reading is **normic**, not strict as this section\nfirst said: a dry chair left outdoors would call for an explanation, and *it stood under the eaves* would leave the rule\nstanding (AT12's two questions). That explanation is an exception claim of its own, naming a kind (sheltered things) that\napplies to other cases and can fail on its own; it moves the dry chair out of the rule's count into the exception's (AT3,\nAY4).\n\nTwo rules are at work, each with its own job. The middle is the rule the cases are filed under: the lawn, the roof and the\nchairs are its conforming cases, counted once. Under its link to *it rained* sits a second normic rule (AT1): *when\neverything left outdoors gets wet overnight, it has rained*. Its exception is a claim too: *on dew nights, everything left\noutdoors gets wet without rain*, with its own count of dewy nights. For this night, *it was not dew* denies that the exception\napplies, and anyone can attack it (*the forecast reported dew*).\n\n**Exceptions are claims.** His question: *\"Why? Shouldn't everything that can be a challengeable claim be a claim in\nDeliberus? How else would the exceptions be up for debate?\"* They are claims, and debatable like any other. What AT13\nrules out is narrower than its spoken name suggests: no rule claim stores a list of its exceptions as a property, hidden\nfrom debate. A counterexample arrives as an objection (a claim), an explanation arrives as an exception claim that attacks\nthe objection's relevance, and each can be attacked, supported and counted (AT3, AY4, AT13). Example:\n\n- Rule: *this gate is open every Tuesday*.\n- Objection: *the gate was closed on Tuesday 24 December*.\n- Exception, a claim of its own: *the gate is closed on public holidays*; it attacks the objection's relevance, so the\n  closed Tuesday stops counting against the gate rule and counts for the holiday rule instead.\n- Its own objection: *the gate was open on Midsummer Eve, a holiday*.\n\nThe headline *no claim carries exceptions* misled him, which is the stranger test failing on a ruled row's name; he\nruled the plainer name the same evening (*\"Ruled!\"*): *an exception is a claim of its own*.\n\nHis 14 October condition holds and becomes a test: drawing the middle may change the conclusion only by\nremoving the double count. Drawn over cases that share no rival, the conclusion must stay what adding gave\n(the chain note's § 4.6: writing a middle must not lower a conclusion). Owed as a test with the build.\n\n### 3c. Four ways two pieces of evidence are one, each with a home\n\n| Dependence | Example | What finds it | Its home in the graph |\n|---|---|---|---|\n| Rewording | two sentences, one proposition | the keeper's *same* | a merge (sameness first, ruled 2026-09-13) |\n| Content | one reason entails or partly establishes the other | the keeper's *contains*, *overlaps* | a part, or a separability proposal |\n| Shared origin | two papers copy one wire report; two witnesses who had lunch | provenance, not text | a source as a node, its inputs one channel (Q4, open) |\n| Shared rival | a wet lawn and a wet roof, both explained by dew | the independence reading's *one story* (Q40) | a claim between, with the rival's denial on its link (Q42, open) |\n\nAnd between case bundles, a shared route (a publication filter, one lab) is filed as an explanation naming a\nkind (BA1, ruled tonight). A shared origin is a shared rival whose cause is a source; where the common cause is\na source it goes to provenance, where it is something in the world it goes to a claim. The independence\nreading's verdict then routes: *same* to a merge, *contains/overlaps* to a split, *one story* to a source node\nor a claim between, depending on what the story is.\n\n### 3d. The move is a rehoming\n\nPutting the lawn's link under the middle is the same operation as moving a cut whole's links to its parts\n(AL15, ruled; scoped in AX12): the old link kept visible and marked, the new ones naming their readings. A\nmachine's link moves on a confident later reading (R12); a person's link moves when the move is obvious, and\nthe person is told (AS1). The hand map's table of moves serves both.\n\n### 3e. Two tensions\n\n- **The floor (AX8).** The middle must be one proposition: a generalization, or the observation the cases\n  share. It is never the stored conjunction *the lawn, the roof and the chairs are wet*: that sentence is a\n  sideways join, a reading (AX8, AX9), and stored it would be cut back into its items.\n- **Who writes the middle (AG12).** Authors rarely write it, so a machine usually will, and a sentence a\n  machine writes is asserted by nobody and weighs nothing until examined. Moving the reasons under it at once\n  would drop the conclusion's support until someone looked. So the old links must keep counting, as the ruled\n  default reads one-story reasons (the best one, Q46), until the middle is examined (§ 5).\n\nTwo smaller ones: the middle may already exist in another source's words, so a machine matches before it\nmints (AL16) and asks sameness first; and if the middle fails (it was dew), the conclusion loses that one\nreason and the cases still stand (Q34).\n\n## 4. The two folds\n\nThe folding session of the same evening (`folded-routes-and-shared-premises.md`,\n`folding-joining-and-zoom.md`) already leaned on Q42:\n\n- **AX3 is Q42 for reported beliefs.** Two groups' beliefs written down to their bases meet at *the polls show\n  the incumbent ahead*, which sits under both \"as Q42 draws it\" (`folded-routes-and-shared-premises.md`\n  § 4a). The market and the panel are cases of *informed opinion followed the polls*; the polling miss is the\n  rival on its one link.\n- **AX2 reads it by code.** Once a middle is drawn, a claim upstream of two siblings marks them *partly*\n  without a model.\n- **AX9's table needs a fourth row.** Today it has parts of one cut and premises of one argument (joined into\n  one card, filled at the weakest), separate reasons (stacked), and support against attack (never). Reasons\n  that share a rival are none of these: they sit as cases under their middle. At the arguments zoom (AX10)\n  they fold into the middle's card, *3 observations*, its one offered move the open rival (*was it dew?*); at\n  the reasons zoom the middle shows with its cases beneath.\n- **The join and the middle are near-twins, and must not be confused.** *The lawn, the car roof and the garden\n  chairs are wet* says the cases together and is a reading; *everything left outdoors got wet* says what they\n  have in common, goes beyond them, and carries a link. A pass that writes a middle can take the join as its\n  draft and generalize it, marked as going beyond its cases.\n- **The folded route shows the doubt.** *Lawn → everything wet → it rained* is a two-link route; folded at the\n  grounds zoom, its weakest link is the step to *it rained* while dew is open, which is what AX4 asks a band\n  to show. Before Q42 that doubt was invisible, spread over two parallel links.\n\n## 5. Q42 with four clauses, ruled with his exception\n\n*Ruled 2026-10-23: asked \"Rule the shared-doubt claim with the four clauses?\", he chose \"Rule Q42 with the clauses, except\nthat a machine-written middle carries its cases' strength from the start instead of waiting to be examined.\" Clause 2 below\nis replaced by that, and § 3e's second tension is settled his way: the middle is not a sentence under a link with nothing\nbeneath it (AG12), it stands on asserted cases, so it carries what they establish at once, and the rehoming completes at once.*\n\n1. The middle is one proposition at the floor, a generalization or the shared observation, never the stored\n   conjunction of the reasons; its cases are filed under it and counted (AY2).\n2. ~~Until the middle is examined, the old links keep counting as the ruled default reads one-story reasons\n   (the best one); examined, the rehoming completes.~~ Replaced by his exception: the middle carries its cases' strength\n   from the start.\n3. The move is a rehoming under the rules of R12 (a machine's links) and AS1 (a person's, told).\n4. On the map, reasons under a middle fold into its card at the arguments zoom, with the rival as the card's\n   one move (a row for AX9).\n\n## 6. The polls case: markets and experts that both read the polls\n\nThe case from *Where dependence hides* and `folded-routes-and-shared-premises.md` § 3–4. As a source might put it:\n\n- *The incumbent will win the election* ← *Prediction markets give the incumbent a 70 % chance of winning* (support link)\n  and ← *A panel of political scientists expects the incumbent to win* (support link).\n\nWhat the record shows: market prices follow the polls (r = .86 to .91, US elections 1936–2008); in ordinary years the\nmarket's and the experts' errors were nearly unrelated (.01 in 2008, .36 in 2004); in UK 2015, Brexit and US 2016 both\nfailed together, and the shared cause each time was that the polls were off. So the two reasons are nearly independent\nwhen the claim is true and share one story when it is false, which is exactly where the ruled independence question\nlooks (Q40), and his reading of the case, *partly*, was the right one.\n\nDrawn by the ruled rows:\n\n- **Cases**: the market's price and the panel's expectation, filed under\n- **the middle**, a rule over forecasters: *most informed forecasters favour the incumbent*. Its kind, by AT12's first\n  question: a forecaster who favours the challenger calls for no explanation, since forecasters disagree all the time, so\n  it is statistical and states its proportion. It supports, by one link,\n- **the claim**: *the incumbent will win the election*. Under that link, the rule: *the informed forecasters' favourite\n  wins about as often as their stated chance*. Statistical too: a favourite at 70 % that loses calls for no explanation,\n  since that is what 70 % means. So the shared rival is no exception (a statistical rule has none, AT3) but a narrower\n  rule with its own count (AY4): *in years when the polls everyone reads are systematically off, the forecasters'\n  favourite wins less often than its stated chance*. For this election, the open move: *the polls are systematically off\n  this year*, or its denial, with whatever bears on it (changed weighting since the last miss, turnout models, the size of\n  the lead).\n- **The track record**, as cases: elections where the forecasters agreed, most conforming; UK 2015, Brexit and US 2016\n  failing. They stay in the broad rule's count, each moving it a little, and as polling-miss years they are also the\n  narrower rule's cases, which is what gives it a number. The narrower rule is kept only if it carries spread the broad\n  rule leaves unexplained and was named before the count was read, or holds on elections not used to find it (AY4's test).\n- **The first drawing of this section, the same evening, read the link rule as normic** with the polling miss as its\n  exception, and the chat said the middle was *most likely normic*. Run through AT12's question, both are statistical. The\n  difference reaches the arithmetic: read as normic, the three misses would leave the main rule's count; read as\n  statistical, they stay in it and move it. The dew case (§ 3b) is the normic one: a dry chair left outdoors does call for\n  an explanation.\n- **Beneath each case, what it rests on** (AX3, ruled 2026-10-23, widened to every stance in § 6b): *traders read the polls*, *the panel's models use the polls*, with\n  *the final polls show the incumbent ahead* one claim under both; and what each adds beyond the polls (the market's other\n  information; approval and the economy for the panel). Code sees the shared claim and marks the two cases at least\n  *partly* dependent with no model (AX2, open). If someone gives the polls as a reason of their own, they join the same\n  middle as a third case, since their rival is the same.\n- **The other channels his own reading named** (case 1, `how-independent-reasons-add/his-readings.md`): besides\n  the shared input (the polls), shared people (*some political scientists trade on prediction markets*) and one\n  reading the other (*traders read the experts' forecasts*; *the panel reads the market*). Each is written as a claim\n  of its own and can be attacked, and how much of the two cases' support runs through a shared route follows those\n  claims' strengths. His premise about who trades, which he caught himself holding, is one of them.\n- **The size of the overlap in ordinary years**: combining markets and experts cut forecasting error almost as much\n  as two independent methods would (Graefe et al. 2014, three US elections with no polling miss), while polls and\n  models together were worth about one and a half, because the models feed on the polls (the walkthrough *Why\n  independent reasons add*, 23 October). So the middle's link takes its relevance from the record of forecasters who\n  agreed, never from one method's calibration, or counting once undercounts; in polling-miss years the shared route\n  dominates, which is what the narrower rule says.\n- **The rule's kind is itself contestable**: read by calibration, a favourite's loss calls for no explanation; after\n  2016 many treated such a loss as calling for one (the polling post-mortems), which is the normic reading. Yesterday's\n  walkthrough found the same split in the deterrence debate, the ACLU's rule statistical and van den Haag's normic\n  (*What an argument is made of*, 22 October). A scope reading is read, never set, and a person disputes it with a\n  rival wording (AT11).\n\n**Why this shape and not another.** Considered: leaving both reasons directly under the claim, which the ruled default\nthen reads as the best one alone (safe, but it drops the second forecast's extra and hides the shared doubt in a number);\nand writing out only the bases, polls once, fundamentals once, the market's other information once (it throws away the\nforecasters' own synthesis, which weighs things nobody wrote down, and the consensus's track record is the best-measured\nnumber here). Chosen: the middle, which keeps the synthesis, counts it once, and puts its weak point where a reader can\nargue it.\n\n**One trap the build must avoid.** The consensus's track record already contains what the forecasters used. Adding the\nwritten-out bases as separate reasons beside the middle would count the fundamentals twice: the reference-class reading (how\noften the forecasters' favourite wins) and the inside reading (what each forecast rests on) are two readings of one body of\nevidence, and only one of them may add. The bases stay beneath the cases, shown at the grounds zoom, and counted through\nthem.\n\n### 6a. Partial dependence, once the grounds are written out\n\nHis question, 2026-10-23, of a sentence in this session's reasoning that said the graph cannot cleanly separate partial\ndependence: *\"Can it not when all evidence as subgraph is explicated and scaffolded using the graph object model and all\ninferences with each smallest step are mapped out, such that you can find the shared root?\"*\n\nIt can, and that is the answer to *partly*. Written down to what each case rests on, the shared root is one claim with two\nlinks up to it: *the final polls show the incumbent ahead* stands under both the market's case and the panel's. Code finds\na claim two cases both stand on with no model call (AX2), counts it once, and what each case rests on beyond it (the\nmarket's other information; approval and the economy for the panel) stays two routes that may add. \"Partly\" stops being a\nnumber someone guesses and becomes a shape anyone can read: how much of each case's support runs through the shared claim.\nThree limits remain, each with its home: a shared root nobody has written down (coherent absence, which only invitation and\nthe grounds-writing pass of § 6b reach); a shared method or prior nobody named (two forecasters using one turnout model);\nand the trap of § 6, the consensus's track record and the written-out grounds being two readings of one body of evidence,\nof which only one may add.\n\nThe \"flywheels, microtasks\" session reached the same shape the same evening, drawn on the same case, with the polls'\nreliability as the claim on the link from what forecasters expect to who wins (`SESSION-HISTORY.md`, 23 October; his\nstanding word there on what is logical: *\"want what can be settled as logical in the graph to be settled automagically.\nBeyond that, nondeterministic magic takes over, still with aggressively minimized error margins, since errors are liable\nto compound, as we've established.\"*).\n\n### 6b. Every stance written down to what it rests on (AX3, ruled widened)\n\nOf the next sentence in the same reasoning, *\"unarticulated judgment (an expert's intuition, a trader's private\ninformation) remains an unwritten remainder inside the case\"*, he wrote: *\"All of that should also go into the Deliberus\ngraph, as soon as possible, using whatever machine passes we have at our disposal or can create and scaffold to\nrecursively decompose all stances\"*. And minutes later: *\"Whoever asserts a case implicitly asserts its constituents\"*.\n\nWhat follows, as recorded in the ledger:\n\n- **Every stance, not only a group's belief.** A stance a source reports or a person states is written down to what it\n  rests on, recursively to the separability floor (B2), by machine passes.\n- **The unarticulated part too.** An expert's intuition is written as the cues an expert in that field reads (from the\n  field's own literature), marked machine-written. Where a ground's content is the holder's alone, the machine writes that\n  such a ground exists, *the price moved on information not in the public record*, which is attackable on its own (did\n  the price move without public news?), and the holder may fill the content.\n- **Constituents and grounds are attributed differently.** A case's constituents are what it says: for *this red-sky\n  morning brought rain*, that the sky was red this morning and that it rained. Whoever asserts the case asserts them, which\n  is a deduction code can settle (AL15: provenance passes down to every part); a constituent the machine writes is marked\n  as reconstructed, and the asserter may disavow it. A stance's grounds are what its holder relied on: the traders' reading\n  of the polls is the traders', never the reporter's who said *the market gives 70 %*.\n- **Strength.** A constituent of an asserted case is at least as strong as the case, since the case entails it: a lower\n  bound code computes, the kind of logic his word above wants settled automatically. A ground the machine writes, with no\n  case beneath it, weighs nothing until examined (AG12), unlike Q42's middle, which stands on asserted cases.\n- **Causes stay out.** Grounds are reasons a holder relies on. Why a person believes (stakes, motive, group) is a cause,\n  never shown publicly (AN5), and a hidden guess at it waits for its own build (AN13).\n- **Depth is capped by measurement.** Machine work building on machine work compounds error (T1), which is his own reason,\n  stated the same evening, for the reviews: *\"Without sufficient guidance and grounding, I don't trust the compounding rate\n  of errors in judgment that would follow if I were to just hand over the reigns for building Deliberus to Claude or any\n  other frontier LLM.\"* So each level the recursion adds waits on the pass below having a measured error rate, and each\n  new pass is tried at three or four model calls before any sweep (R8).\n\nWhether this needs a pass of its own was audited the same night, on his question: § 6c.\n\n### 6c. No new pass: the grounds are answers to questions the schemes already ask\n\nHis question, after ruling the canon line: *\"Ruled! But are we sure a new pass is needed for this? Perhaps the passes\nshould be composable. In that case, I think several existing passes could trigger this new pass, or they overlap in\ntheir purpose, since I half-remember this kind of decomposition at least being my intention for several of the others,\nbut let's go through and audit them.\"*\n\nHis half-memory is right twice. The enrichment ruling of 25 August already ordered it: *\"All unstated premises need to\nbe clarified and scaffolded. Ideally whenever the weighing questions are applicable they should be auto opened AND\ntentatively responded to in the graph as faithfully as possible according to our reading of the source/sources\"*\n(`kartpaus-auto-enrichment-design.md`; its orchestrator loop is unbuilt). And the schemes a reported belief falls under\nalready ask for its grounds, as critical questions latent until somebody answers them (`deliberus/extraction/schemes.py`):\nexpert opinion's *\"{expert}'s assertion about {claim_topic} is based on evidence\"*, popular opinion's *\"the popular\nbelief that {specific_claim} is based on adequate information rather than bias or incomplete evidence\"*, position to\nknow's *\"{source} is in a position to know\"*, and the witness's *\"was in a position to observe\"*.\n\nThe 28 registered passes (`standing_work.PASSES`), read for what each writes toward a stance's grounds:\n\n| Pass | What it writes | Its share of the grounds | State |\n|---|---|---|---|\n| reported speech | the stance as a claim, held by its holder, joined by a link that carries no strength | the stance itself | built; standing in phase 2 |\n| scheme classifier | the link's pattern; a reported belief used as a reason gets a source-based scheme | which questions apply | live; standing in phase 3 |\n| critical questions and latent questions | the scheme's questions, rendered on read, minted when answered | the slot each ground goes in | live |\n| *(missing)* an answering step | a model reads the stance and its source and answers such a question with the grounds, as claims | the grounds | unbuilt: the August ruling's *tentatively responded to* |\n| reasons cut off a sentence | the grounds a holder stated with *because* | stated grounds | built |\n| the splitter and its coverage check | a case's constituents | constituents | built, writing marked parts |\n| implicit premises | a premise one source's argument relied on unsaid | unsaid premises in one source | live on the Temporal path |\n| cross-source premises | a frame premise two opposed sources disagree over | frame premises | built, run by the operator |\n| machine sentences, and the chain writer | the smallest steps under a link | the leap from a stance to what it supports | the store built; the writer a trial |\n| the keeper | a written ground that says what an existing claim says joins that claim | where two cases' grounds meet | standing |\n| independence reading, and code under AX2 | whether two reasons share a route, once the routes are written | dependence | standing, not swept |\n| standing work | each pass owes work on a condition in the graph, and order follows the data (principle 5 of its plan) | the recursion: a written ground is a new claim the claim passes owe a reading, and its link one the link passes owe | phase 0 built |\n\nSo grounds-writing composes from what exists. What is left: **one answering step**, registered inside the\ncritical-questions family (owed by a source-based question with no answer and no reading), the August loop's\npre-answer made standing; **a store variant**, since a ground a machine writes is attributed to the holder and marked\nreconstructed, where a machine's sentence under a link is asserted by nobody (AG12); **phases 2 and 3 of the\nstanding-work plan**, which make each written ground trigger the passes above; and the recursion's brakes, the\nAugust loop's stop rule (spend the next action where a reader would see a change) beside T1's depth cap. Two cautions:\na reported belief whose link is classified as a sign or a cause gets questions that do not ask for grounds, so the\nanswering step reads the scheme first; and an answer is a proposal of what the holder relies on, which the holder may\ndisown (the consent channel).\n\nThe polls case through it: the market's link is classified popular opinion → its question *the popular belief is based\non adequate information* is answered with *traders' prices follow the final polls*, *the final polls show the incumbent\nahead* and *the price also moved on information not in the public record* → the keeper finds the polls claim already\nwritten under the panel's case → the independence reading reads *partly* → the claim between of Q42.\n\n## 8. Open loops of this session, and what the two days' walkthroughs add\n\nWritten 2026-10-23 on his request: *\"Document all open loops and loose ends from this session, and double check if\nanything conflicts with or is augmented/deepened by a reading of what the walkthroughs said about any of this today or\nyesterday.\"* The six walkthroughs of 22 and 23 October were read for this session's threads.\n\n**Ruled tonight, unbuilt** (each in `TODO.md`): the claim between reasons that share a rival (Q42); the rival question\nasked between bundles and a unit's cases under the unit's rate (BA1, BA2), both built with the case channel; every\nstance written down to what it rests on (AX3, widened, its canon line resident): the answering step, the store variant\nand the lower bound code computes for a constituent; the spoken name *an exception is a claim of its own* (AT13), shown\non the site's ledger wheel after the next deploy.\n\n**First instance on a live map, 2026-10-24, by hand**: on the institute map the four Nordic neighbours behind the\ncall (*Danmark/Island har ställt sig bakom*, *lanserades av Finlands president / Norges statsminister*) go under\n*Alla Sveriges nordiska grannländer har ställt sig bakom uppropet*, which supports *Sverige bör ställa sig bakom … uppropet*\nonce, with the shared doubt written as its co-premise, *Sverige bör göra som sina nordiska grannländer i frågan om\nuppropet*. Kept as a case for the unbuilt pass: `tests/fixtures/microtask_cases/institute-map.json`, `claim_between`.\n\n**The answering step's first answer key, 2026-10-24, by hand**: on the institute map five consultation answers (MPF,\nFOI, RISE, PTS, KTH) written down to the grounds each gives in its own text, quoted and attributed to the holder, every\nlink marked as not independent of the stance it serves; FOI's *we have what is needed* rests on four of FOI's\ndescriptions of itself, read as one argument since one rival (FOI speaking for itself) explains all four. 14 cases in\n`tests/fixtures/microtask_cases/institute-map.json`, `grounds`; a mapper's work, not his labels (Y13). None needed a\nground only the holder knows: every one was stated in the source, so the reconstructed-ground half of AX3 has no case yet.\n\n**Waiting on him**: who builds the answering step (this session, or the session that owns machine passes and their\nquality control); code marking a pair that shares an upstream claim as at least partly (AX2); a reading names its\nchannels (AX1); a step's rule is the grounds of its link's relevance claim (AD22); the folds and zoom levels (AX9 to\nAX12, the folding session); which takes of the elder's sounds stay; his reflections on the case rows, promised when he\nblessed them; his thoughts on each chapter of *Where dependence hides* (filed by the microtasks session).\n\n**What the walkthroughs add or contradict:**\n\n1. *Where dependence hides*, part two, and his case reading: he named three channels where § 6 first drew one. § 6 now\n   carries all three, each as an attackable claim.\n2. His reading also called the leap from each reason to the claim too large: *\"We need to be responsibly aware at least\n   of all the inferences between them.\"* The claim between and the rule under its link are that leap written out; the\n   steps left inside it are the chain writer's (AG14).\n3. *Why independent reasons add*: markets and experts were worth almost two methods in ordinary years. Counting once\n   then undercounts unless the link's relevance comes from the agreeing pair's record (§ 6). The microtasks session's\n   direction *counting once isn't always safe* says the same from the other side.\n4. *What an argument is made of*: two sides can disagree about a rule's kind, not only its cases. § 6 now says the\n   polls rule is read statistical and that the reading is contestable.\n5. The same walkthrough's open row on relevance (a rule's strength times how often its outcome follows; for a\n   statistical rule, the rate its words state) supplies the polls link's number.\n6. *Where dependence hides*, part four, predicted his answer as *\"Even a strict rule has exceptions, argued under it\"*,\n   from his 21 October words *\"the rule isn't certain for a deductive argument either, there are exceptions\"*. The\n   substance agrees with what was ruled tonight (a rule used in a deduction keeps one number, how sure we are it\n   holds), the word does not: an exception now names an explained counterexample to a normic rule (AT13), and a\n   counterexample to a strict rule refutes it, repaired by a narrower wording (AT12). A note says so in that\n   walkthrough's header; the audio stays as rendered.\n7. Yesterday's walkthrough mentioned a stored deterrence link that points the wrong way; it is already in `TODO.md`\n   (noticed 17 October), not a new loop.\n\n## 7. Proposed (unruled)\n\n- ~~AT13's spoken name~~ ruled 2026-10-23 (*\"Ruled!\"*): *an exception is a claim of its own*, the rule's text unchanged.\n- The resident line proposed with AX3's ruling (§ 6b): *every stance is written down to what it rests on, unarticulated\n  judgment included; whoever asserts a case implicitly asserts its constituents*.\n\n## Cross-references\n\n[reasons-and-cases.md](reasons-and-cases.md) § 4, § 5.9 · [retrofitting-the-reasons-channel.md](retrofitting-the-reasons-channel.md)\n§ 9k, point 3 (Q42's home) · [what-counts-as-evidence.md](what-counts-as-evidence.md) § 2c (Q4) ·\n[folded-routes-and-shared-premises.md](folded-routes-and-shared-premises.md) § 4 · [folding-joining-and-zoom.md](folding-joining-and-zoom.md)\n§ 3, § 5 · [what-a-case-counts-for.md](what-a-case-counts-for.md) · [the-chain-inside-a-link.md](the-chain-inside-a-link.md) § 4.6 ·\nwalkthroughs of 26 September, 14, 15, 20 and 22 October and four of 23 October in `../listening/`.\n"}