{"path":"research/folded-routes-and-shared-premises.md","content":"# Folded routes and shared premises: where two reasons' dependence hides, and how a map can show a long leap\n\n**Date:** 2026-10-23. **Standing:** research, a reading of his first independence case, two\nmeasurements on the graph, and proposals entered as ledger Arc AX; nothing ruled, nothing built.\n\n## 0. In short\n\n1. **His reading was right where the note's label was not.** The case set expected *different* for\n   prediction markets and a panel of political scientists; he read a possible overlap he could not\n   settle. The evidence says *partly*: prices largely follow the polls (r = .86 to .91 over US\n   elections 1936–2008), many forecasting models use the polls, and in the three big misses on\n   record (UK 2015, Brexit, US 2016) markets and experts erred the same way. In ordinary years their\n   errors were nearly unrelated (.01 and .36), which is all the label rested on. § 3.\n2. **Adding needs independence in the cases where the claim is false**, and that is where these two\n   share a cause: a polling miss. The ruled question (*suppose the claim is false; what most likely\n   explains each reason?*) asks exactly that, so the machine's *one story* and *partly* were closer\n   than the fixture. § 3c.\n3. **The overlap he looked for first (shared people) is unmeasured; the one that is measured (shared\n   information) needs no shared people.** § 3a.\n4. **The dependence hides in the distance he named.** Written down to what each group's belief rests\n   on, both chains pass through *the polls show the incumbent ahead*; over the bare reasons it cannot\n   be seen. § 4.\n5. **Code can already see part of it.** Of 13,315 pairs of reasons supporting one claim in the\n   graph, 1,084 have one reason supporting the other within four links and 80 share a claim upstream\n   of both. § 4c.\n6. **A long leap can be folded on the map, not drawn as a link.** 2,003 pairs of claims are six or\n   more support links apart; support does not carry along a chain (non-transitive except under\n   screening off), so a folded route carries no strength of its own and shows instead where it is\n   weakest, what is unwritten in it, and what it shares with other routes. § 5.\n\n---\n\n## 1. His words, 2026-10-23\n\nOn the answer forms, after a report said one was too big for Anthropic's structured output:\n*\"'Forms'? What is a 'form'? The form in the prompt vs... what? Can you explain?\"* (answered in § 6).\n\nOn the first independence case, in full in\n[how-independent-reasons-add/his-readings.md](how-independent-reasons-add/his-readings.md) § Case 1.\nIts turns, quoted:\n\n- *\"The experts are part of the \"market\" generally.\"*\n- *\"But that's ideological and I don't think more than a handful are actually interested, aware of\n  the existence of prediction markets and well-read enough\"*\n- *\"generally I guess it's well-read people who we could expect to have a good grasp on the matter\n  that participate on prediction markets, but again that's just an idea I hold that becomes a\n  premise in how I make sense of this.\"*\n- *\"How would we know that for sure? Who would we trust on that matter?\"*\n- *\"The inferential distance between the claim and the reasons here feels like a bit too much of a\n  leap. We need to be responsibly aware at least of all the inferences between them.\"*\n\nOn the map: *\"But I have this idea of... collapsible inference leaps. If there's 10 hops between\ndifferent claims you could see both in one view and compress/visualize the hops in the map view\nsomehow. I wonder how to realize that. The zoom-out in the current map view is quite limited in\nthis regard.\"* And: *\"Read up on the smallest-step / one inference at a time / link in the chain docs\nand anything vaguely related in the corpus and think deeply about this using the corpus graph etc.\"*\nThen: *\"I want you to reflect deeply on my reading as well.\"*\n\n## 2. What the corpus already holds\n\n| Angle | Where | What it settles |\n|---|---|---|\n| Every link is a chain folded up; the map unfolds none yet | [the-chain-inside-a-link.md](the-chain-inside-a-link.md) § 0, § 2; [../listening/2026-10-10-the-chain-folded-inside-an-arrow.md](../listening/2026-10-10-the-chain-folded-inside-an-arrow.md) | 1 of 1,569 links had anything under it (10 Oct) |\n| A leap may be left, never unmarked; a support link whose middle is on the map shows that middle | the chain note § 4.1–4.2, ledger AG1 | 104 links with a two-link path beside them |\n| Levels opened one at a time, four to ten steps a level (Lamport) | the chain note § 4.5 | how a reader opens a chain |\n| The smallest step: one stated rule, once, no gap, nothing removable | the chain note § 11.8, ledger AG14 | what gets written under a link |\n| Writing the middle must not lower the conclusion | the chain note § 4.6 | 0.519 bare against 0.500 with the middle written |\n| Reasons sharing a likely rival go under a claim of their own | ledger Q42 (open) | the wet lawn, car and chairs under *everything outdoors was wet* |\n| The independence question, ruled: suppose the claim is false, what explains each reason | ledger Q40; [how-independent-reasons-add.md](how-independent-reasons-add.md) § 10 | built, 25 cases |\n| Agreement among believers counts only where they reached it by independent routes | Goldman 2001 in [what-counts-as-evidence/philosophy-of-evidence.md](what-counts-as-evidence/philosophy-of-evidence.md) | the social-epistemology form of the same question |\n| Zooming out risks flattening (his words, 4 Oct) | [how-a-map-is-cut-and-what-it-shows.md](how-a-map-is-cut-and-what-it-shows.md), *A first outside reaction* | *\"då är risken förstås att det blir tillplattat och obalanserat\"* |\n| Semantic zoom, focus+context | [graph-visualization.md](graph-visualization.md) § 3–4 | topic clusters at the far level, no path folding |\n\nNot in the corpus before today: evidence on markets against experts, the non-transitivity of\nsupport, and prior art for folding a path. Two research helpers read them; their reports are kept\nverbatim in [folded-routes-and-shared-premises/](folded-routes-and-shared-premises/).\n\n## 3. His reading, reflected on\n\n### 3a. Two channels, and he checked the weaker\n\nHe looked first for shared **people**: political scientists among the traders. He doubted it on good\ngrounds (few people trade, under selection), and he was right to: the record holds no count of\npolitical scientists who trade. Trader surveys find them young, male, well educated and politically\nactive (Forsythe et al. 1992 on the first Iowa market; Berg and Rietz; PredictIt 2016 92 % men),\nwhich says nothing either way about academics.\n\nThe channel that is measured needs no shared people: shared **information**. Erikson and Wlezien\n(2012), over US presidential markets 1936–2008, find prices correlate with polls at +.86 to +.91 and\n*\"add nothing to election prediction beyond polls\"* when both exist; in 1948, when the polls were\nwrong, *\"the market followed the polls.\"* Many political-science models put trial-heat polls in\nbeside approval and the economy (Campbell, Holbrook, Erikson and Wlezien in 2016). So the two\nreasons can share an error through what both read, whoever the readers are. His sentence *\"it's\nwell-read people who we could expect to have a good grasp on the matter\"* points at this channel\nwithout naming it: well-read people read the same polls.\n\nMarkets are not only polls. Rothschild (2009) finds debiased Intrade prices beat polls in 2008's\nuncertain races, and in 2024 markets priced Trump five to ten points above the poll-based models.\nSo the channel is partial.\n\n### 3b. *\"The experts are part of the market\"*: the warrant is itself a claim of dependence\n\nHis first sentence holds a sharper point than overlapping people. What is supposed to make a\nmarket price a good reason is that it aggregates all available information. A published expert\nforecast is available information. So the more one trusts reason one for the reason markets are\ntrusted, the more reason one already contains reason two. A reason whose warrant is *this\naggregates what is known* cannot be independent of the known things it aggregates; doubt the\nwarrant and reason one weakens instead. Either way the pair does not count as two full reasons.\n\nHe then set that warrant aside as ideological, and the measurements side with him: markets mostly\nfollow polls rather than aggregating everything, and the expert panel's errors and the market's\nwere close to unrelated in 2008 (.01) and only moderately related in 2004 (.36) (Graefe et al.\n2009, daily errors over the last 93 days). In theory the market contains the experts. In fact it\ncontains part of what they read.\n\n### 3c. Where dependence counts: in the failures\n\nThe note labelled the case *different* from Graefe et al. 2014: over 2004–2012, market and experts\ntogether cut the error by 29 %, what two independent forecasters would. Three ordinary elections.\n\nAdding two reasons in log-odds needs them independent given the claim, and above all given that it\nis false: the overclaims of § 5c in the independence note all came from a shared cause that showed\nonly when the claim failed. On record, when it failed, these two failed together:\n\n| Case | Experts | Markets | Outcome |\n|---|---|---|---|\n| UK general election 2015 | 6 % of surveyed experts expected a majority; all 12 academic teams a hung parliament | about 6 % for a Conservative majority | Conservative majority |\n| Brexit 2016 | Leave 38 % (PSA survey of 596, late May) | Remain 73–77 % on 22 June, 84–93 % on the day | Leave |\n| US 2016 | 75 of 77 panel forecasts overestimated Clinton | Iowa market +4.8 points for Clinton | Trump |\n\nThe common rival in all three is the one the ruled question finds: *the polls were wrong*. The PSA's\nown report warned that agreement across its expert groups *\"may just be a sign of groupthink\"* and\nthat experts and bettors *\"have access to the same opinion polls\"*. So the reading that fits is\n**partly one story**: little shared in ordinary years, much shared in the years that decide whether\nadding was safe. The machine's five answers (one story four times, partly once) erred towards\ncounting once, as the swap rule intends; the fixture's *different* would have let the two add.\nCorrection owed: the expected answer for this case becomes *partly*, and the scores in\n`how-independent-reasons-add.md` § 10 are recounted against it (pointer added there).\n\n### 3d. A premise he caught becoming a premise\n\n*\"that's just an idea I hold that becomes a premise in how I make sense of this.\"* This is the move\na reader rarely makes and the machine never does: its answer form asks for two rival explanations\nand a verdict, and never for the premise the verdict hangs on. His reading hangs on at least three:\nwho trades, whether political scientists trade, whether markets are calibrated. A second reader with\nother beliefs about traders would read the same pair differently and could not see why. That is the\nlens case of Arc AU (where two minds part is often in a premise neither stated). So a reading should\ncarry its premises, each a claim another reader can contest.\n\n### 3e. *\"Who would we trust on that matter?\"*\n\nThe regress he names ends in two places. Where a measurement exists, it ends there: that prices\nfollow polls is measured; that experts and markets missed together in 2015 and 2016 is a matter of\nrecord. Where none exists (how many political scientists trade), it ends in a stated open question,\nand the reading holds at *partly* rather than pretending to settle it. That is AG1 applied to a\nreading: an unexamined channel may be left, never unmarked.\n\n### 3f. *\"There's so many thoughts to be had about this\"*\n\nOne pair of reasons opened about a dozen claims in his reading. That is the descent he asked for in\nAugust (*\"ask 'Why?' recursively\"*), and it is also why his reading cannot be the instrument: the\nmachine should enumerate the channels and premises, and his attention should go to the premises\nwhere the evidence runs out (the constraint of `flywheels-in-building-deliberus.md` § 3).\n\n## 4. The distance he named is where the dependence hides\n\n### 4a. Both chains written out\n\nAt the floor (AG14), each reason reaches the claim through a reported belief of a group:\n\n**Reason one** · *Prediction markets give the incumbent a 70 % chance of winning.*\n1. A contract's price is the traders' collective probability *(rule: a binary contract's price is\n   its market's probability, less fees and the long-shot bias)*.\n2. Those traders' probability tracks the outcome *(rule: this market is calibrated)*; their belief\n   rests on what they read: **the polls show the incumbent ahead**, and other news.\n3. So the incumbent wins, with about 0.7 *(what has a 70 % chance usually happens)*.\n\n**Reason two** · *A panel of political scientists expects the incumbent to win.*\n1. The panel's expectation is the members' shared belief.\n2. Their belief tracks outcomes *(the expert-opinion scheme; its question* is it based on\n   evidence?*)*; it rests on approval, the economy and **the polls show the incumbent ahead**.\n3. So the incumbent wins, with the strength of that track record.\n\nThe shared sentence appears one level down, in step 2 of each: what each group's belief is based on.\nThat is Goldman's test (agreement counts only where the believers reached it by independent routes)\nand the expert-opinion scheme's backup-evidence question, written as a node. Drawn as Q42 proposes,\nthe two reasons go under *the polls show the incumbent ahead* as far as they rest on it, and only\nwhat each adds beyond the polls stands as an independent route.\n\n### 4b. What this says about the independence reading\n\nRead over the two bare reasons, a model has to imagine the shared middle; over written chains, the\nshared sentence is there to be found, and finding it is a comparison of two lists. His complaint\nabout the leap and the independence question are one question: the dependence lives in the middle\nthat is not written.\n\n### 4c. What code finds today\n\nOn the graph `deliberus_extraction` (read 2026-10-23, superseded links left out): 13,315 pairs of\nreasons each supporting one claim. In 1,084 one reason supports the other within four support links\n(the pair is at most *contains*); in 80 more a claim upstream of both supports each within four links\n(at least *partly*). Example: *the Singularity could bring subtle catastrophes* and *utility-driven\nAI faces pressure to use its resources maximally*, both for *once takeoff begins there will be no\ntime to deliberate*, both resting on *an AI with a harmless goal could consume the Earth's resources\nfor computing*. With chains written under links, more pairs would show their shared ground the same\nway. The count is of the whole support graph, not the 334 pairs the standing pass owes.\n\n## 5. Collapsible leaps: a folded route\n\n### 5a. Two directions of one operation\n\nThe corpus has the unfold direction: *every arrow is a chain folded up* (the chain note; walkthrough\nof 10 October). His idea is the other direction: two claims ten links apart, both in view, the links\nbetween them folded. A link with its chain written under it and a path of links between two distant\nclaims are the same object, a **route** between two claims, drawn folded or open.\n\n### 5b. There is material for it\n\nShortest support distances between pairs of claims, same graph: 1 link 3,340 pairs, 2 links 3,415,\n3 links 4,589, 4 links 3,400, 5 links 2,048, 6 links 975, 7 links 634, 8 links 260, 9 links 108,\n10 links 25, 11 links 1. Six claims sit in a two-link support loop. The map today shows two links\nout from one claim.\n\nThe eleven-link route runs from *advanced AI systems will exhibit instrumental drives regardless of\ntheir goals* through capability, agency, self-improvement and *Future AI systems may greatly exceed\nhuman performance*, ending at *Some people believe the consciousness carried forward by AI\ndescendants would persist for an extremely long time*. Read whole, its two ends barely bear on each\nother, and two of its links run from a claim into a report of what some people believe.\n\n### 5c. A folded route carries no strength\n\nProbabilistic support is not transitive: x can raise y and y raise z while x does not raise z\n(Hesse 1970; Eells and Sober 1983; Shogenji 2003). It carries along a chain only where each middle\nclaim screens off what came before (Shogenji's condition: given the middle, the start tells nothing\nmore about the end), and threshold versions fail even then (Roche 2012). For defeasible steps the\nstandard policies (weakest link, Pollock; last link, ASPIC+) rank chains differently from each\nother. So a folded route is drawn as a route, never as a derived support link, and shows no single\nnumber. The eleven-link example is the reason: as one arrow it would claim what nothing in it\nestablishes.\n\n### 5d. What a folded route shows\n\nFrom the literature (report § 1, § 6) and the corpus, each with its source:\n\n- **Sound and complete** (the provenance views of ZOOM/WOLVES): the fold implies no dependency the\n  route lacks and hides none. A side attacker or a side premise on any of its links splits the fold\n  where it attaches, rather than vanishing inside it.\n- **Where it is weakest, not a total**: the number of links, the weakest link, how many links have\n  nothing written under them (AG1's mark), how many are contested.\n- **What it shares with sibling routes**: where two reasons' routes pass through one claim, the fold\n  shows the join. This is § 4 drawn: the shared polls visible on two folded routes at once.\n- **Opened a level at a time** (Lamport, the chain note § 4.5), keeping the folded route's place on\n  screen (Raitner 2004); embedded at small scale or inserted at full scale, the trade-off Kieffer\n  names for Proofscape, untested by anyone.\n- **Any written summary of a fold points at the links beneath it** (Coscia et al. on Semantic\n  Bundling: a fluent label can claim a connection the path does not support).\n\nNo study compares folded and flat chains for readers' understanding (report § 4). Graphionale's\nusers rated progressive disclosure the most helpful of four graph features, a rating, not a measured\neffect.\n\n## 6. What an *answer form* is (his first question)\n\nEvery machine pass returns its answer in a fixed set of fields defined in code. The independence\nreading's form has five: whether the two are premises of one argument, the most likely explanation\nof reason one if the claim were false, the same for reason two, a comparison, and one answer from\nseven fixed words. The engine can hold a model to a form two ways:\n\n- **Structured output**: the form goes to Anthropic as a JSON schema, and its decoder can only\n  produce text that fits. Guaranteed.\n- **The form written in the prompt**: the fields are described in the instructions, the model is\n  asked to answer in that shape, and the code checks the reply and sends it back with the error\n  when it does not fit.\n\n`FocusedResult`, the form of the claim-reading pass (a list of claims, each with its type, its words\nin the source and more), is too large for Anthropic to compile as a schema (*\"Schema is too\ncomplex\"*), so it goes the second way; on 11 calls it was sent back 8 times. Glossary entry:\n*answer form*.\n\n## 7. Recommendations (Claude's, unruled; ledger Arc AX)\n\nIn order of dependency.\n\n1. **Relabel the first case** *partly*, and recount the independence runs against it (§ 3c). A\n   test-fixture correction, his reading and the evidence agreeing.\n2. **AX1 · a reading names its channels and its premises.** The independence reading lists the\n   channels by which the two reasons could share an error (shared inputs, shared people, one reading\n   the other, a shared method, a shared incentive or bias) and the premise its verdict hangs on; it\n   answers *different* only when every named channel was examined; a plausible channel left\n   unexamined holds it at *partly*. His reading turned into a rule (§ 3a, § 3d, § 3e).\n3. **AX2 · code reads the graph before a model reads a pair.** Where one reason supports the other\n   along written links, the pair is *contains* at most; where a claim upstream supports both, it is\n   at least *partly*. 1,164 of 13,315 pairs today (§ 4c).\n4. **AX3 · a reason that reports a group's belief is written down to what the belief rests on**\n   before its independence from a sibling of the same kind is read: a market, a panel, a poll, a\n   meta-analysis, a consensus statement. The shared basis then sits under both as Q42 draws it (§ 4).\n5. **AX4 · a long leap is shown as a folded route**, never as a link: no strength of its own; its\n   length, weakest link, unwritten and contested links, and claims it shares with sibling routes;\n   opened a level at a time; split where a side attacker or premise would otherwise vanish (§ 5).\n\n**Only he decides:** whether the first case's label changes (1); whether a reading must name its\nchannels (AX1), which costs a longer answer form and makes *partly* the common answer until channels\nare examined; and what a folded route looks like on the map (AX4), a choice about what readers see.\n\n## Sources\n\nReports kept verbatim: [folded-routes-and-shared-premises/markets-and-experts-report-2026-10-23.md](folded-routes-and-shared-premises/markets-and-experts-report-2026-10-23.md)\n(Erikson and Wlezien 2008 and 2012; Rothschild 2009; Graefe et al. 2009 and 2014; the PSA surveys of\n2015 and 2016; Fisher and Shorrocks 2017; Forsythe et al. 1992; Crane and Strumpf, preprint) and\n[folded-routes-and-shared-premises/folding-paths-report-2026-10-23.md](folded-routes-and-shared-premises/folding-paths-report-2026-10-23.md)\n(Shogenji 2003, abstract only; Roche 2012; Eells and Sober 1983; Pollock; Modgil and Prakken;\nZOOM*UserViews; Raitner 2004; Kieffer on Proofscape; Coscia et al., arXiv 2608.04002, uncorroborated).\nGraph counts: [folded-routes-and-shared-premises/graph_counts.py](folded-routes-and-shared-premises/graph_counts.py),\nread-only, on `deliberus_extraction` 2026-10-23.\n"}