{"path":"research/what-if-and-lenses.md","content":"# The what-if view: one door for a reader's suppositions, two people's lenses, and a policy's rules\n\n*22 October 2026, evening. His question and his framing; the measurements are this session's, run on\nthe live graph; the proposals are Claude's, unruled. Outside reading in two reports kept verbatim beside\nthis note: [what-if-and-lenses/outside-reading-formal.md](what-if-and-lenses/outside-reading-formal.md)\nand [what-if-and-lenses/outside-reading-psychology.md](what-if-and-lenses/outside-reading-psychology.md).*\n\n## 0. His words\n\nOn hearing part eight of the walk of 22 October, where the first of the group's unbuilt ways in was\nTomas Kronvall's of 2012 (grant a few premises and see what follows,\n[what-the-group-held-2011-2017.md](what-the-group-held-2011-2017.md) § 1.4 and § 4 item 1):\n\n> *\"The what-if view relates strongly to what I was talking about last night when I stayed up til 3-4 am.\n> The worldview-lenses, attunement etc. That's always been my sense of roughly how to effectively render\n> such a mind-meld from the mind of Deliberus in bridging the perspectives/worldviews of two people who\n> don't see eye to eye. This could also be used for hypothetical imaginings of what a policy of multiple\n> given rules could result in if they would hold up, right.\"*\n\nThe night before (22 October, 03:01), in the session that wrote\n[normic-rules-normative-reasons-and-trying-rules-out.md](normic-rules-normative-reasons-and-trying-rules-out.md):\n\n> *\"This seems to relate to normative reasons in Deliberus also. The normic rules. You could imagine a\n> normative claim or set of claims being tried out as rules hypothetically, for example. This could be a\n> way to imagine/simulate - in the mind of Deliberus - what the results of certain complex policies could\n> be. Potentially this sort of interplay and creative interweaving could also help with the attunement and\n> worldview lenses related machinery and features.\"*\n\nThe mind-meld image is his of 18 October ([three-minds-and-the-leap.md](three-minds-and-the-leap.md) § 1):\n*\"a capability of the mind of Deliberus to perform a Vulcan mind-meld between two other minds, bridging their\nperceptions and beliefs about the world, showing what a worldview looks like from the inside\"*. And the\nquestion he left open on 28 March, which this note tries to answer in object terms: *\"it's a very good\nquestion what that concretely consists of, what fundamental dials or levers are tweaked by viewing the graph\ndifferently.\"*\n\n## 1. The answer, in plain words\n\n**Yes: the three are one operation, fed from three places.** The operation is: take a set of claims, hold\neach one granted or denied in supposition, recompute what follows, and show which supposition carries the\nweight. What differs is where the set of suppositions comes from:\n\n| Use | Whose suppositions | What it shows |\n|---|---|---|\n| A reader's what-if (Kronvall, 2012) | the reader picks them, one at a time | *if this held, here is what moves* |\n| A lens, and the mind-meld between two people | each holder's own accepted claims and their own weightings | *given your premises this follows; given theirs, that; and here are the few premises where the difference lives* |\n| A policy trial | a set of rules held in supposition, with the case facts | *if these rules held, these outcomes follow; these are the rules the outcome leans on most* |\n\n**The arithmetic already exists.** The function that computes a claim's badge,\n`compute_badge_recursive` in `deliberus/graph/badge_inputs.py`, takes `overrides`: any number of claims,\neach fixed at a strength, entering every parent and target at exactly that value. The hinge score uses it\none claim at a time (`deliberus/hinge.py`); a what-if is the same call with several claims at once. Used\ntoday only one claim at a time: the route `/claims/{id}/hinge`, and the order of a claim's critical\nquestions (`deliberus/question_ordering.py`). No page lets a reader grant anything.\n\n**What it cannot yet do well is the part that matters most**, and that is measured below (§ 2): the\nshipped strength arithmetic makes every single premise move a conclusion by the same three points, and\nadds up reasons whether or not they share a source. A what-if view built on today's arithmetic would show\neither almost nothing or too much. So the view's honesty depends on rulings already made and not yet built\n(the weakest premise within an argument, independent arguments add only once their independence is read)\nand on one still open (a step's rule is the grounds of its link's relevance claim).\n\n## 2. One example, measured on the live graph\n\n**The claim.** *Once a model crosses a dangerous-capability threshold, stronger custody (secured weights,\nmediated access, monitoring, revocation) is the right response* (`claim_aig_custody_warranted`), the focus of\nthe custody map: Anthropic's framework as Daniel Lindqvist's paper reads it, in wording Claude reconstructed.\nIts four required parts are the four measures, one each. Forty-four claims stand under it; fourteen move it\nat all.\n\n**One premise at a time (the hinge, as shipped).** Every one of the fourteen moves the conclusion by the\nsame amount: granted, the conclusion reads 0.529 or 0.471; denied, 0.5. Hinge 0.029 for all fourteen. The\nranking that is meant to say *this is what the conclusion stands on* says *everything equally, and barely*.\nAn earlier dogfood run found the same on a seventeen-claim tree: flat when unengaged, and one evidence item\nshifted a hinge by about 0.001 after engagement (`dogfood-run-2-orthogonal-experiments.md`, result 3: *\"The\ninstrument never lied; it is just quiet\"*).\n\n**Two whole positions at once.** Person A grants every claim that supports the conclusion and denies every\nchallenge; person B does the reverse.\n\n| | Conclusion reads |\n|---|---|\n| nobody's suppositions | unexamined |\n| person A | 0.892 |\n| person B | 0.276 |\n| A, except one premise set as B has it | 0.785 |\n\nSo whole positions do separate, by a lot. **Which premises carry the gap?** Swapping one premise at a time\nfrom A's value to B's: each of the conclusion's four required parts (secured weights, mediated access, monitoring, revocation) drops it by 0.392, because denying one required part caps the whole by\nthe weakest-part rule; each of the ten reasons and challenges drops it by 0.026. The single swaps sum to\n1.83 against a gap of 0.62. The effects are not additive, so *which premise accounts for the difference* is\nan attribution question with a known formal answer (the outside reading, § 3), not a sum of single swaps.\n\n**Two caveats the example itself shows.** All fourteen claims come from one source: seven are the author's,\nfour are Anthropic's framework as the author reads it, three are longtermism as the author reports it. A\ncomparison of two lenses on this map compares two halves of one person's account, the same limit the claim\nreadings met (`synthesis-build-plan.md` § 10). And A's 0.892 adds ten reasons as if independent; under the\nruled reading (best argument until independence is read), it would read lower until the keeper has read\nwhether those reasons share an origin.\n\n**A second example, from the death-penalty map.** Under the ACLU's claim that the death penalty does not\ndeter murder better than other punishments (`claim_eb8a71030abb`), granting the statistic that\ndeath-penalty states had higher murder rates *lowers* the claim (0.471). The stored link from that\nstatistic is an attack whose own explanation says it refutes the opposite claim; its polarity is inverted\n(recorded in `TODO.md` on 2026-10-17, unrepaired). **A what-if view is also an error detector**: granting a\nclaim's own supporting evidence should never lower it, and a reader who sees that happen has found a\nbroken link.\n\n## 3. What the outside reading says\n\nTwo reports, kept whole beside this note. Each line below names the report section; read depth is stated\nthere per source (no source was read start to finish; most were abstracts plus long excerpts).\n\n### 3a. Psychology, deliberation and the philosophy of disagreement ([report](what-if-and-lenses/outside-reading-psychology.md))\n\n- **The lens must come from the holder, and this is now measured, not only principled.** Imagining another\n  person's view did not make people more accurate about them in 25 experiments; asking them did (Eyal, Steffel\n  and Epley 2018, § 1a). Writing from an opponent's standpoint moderated attitudes only when that opponent\n  would check it; for an abstract \"other\" it did nothing (Tuller et al. 2015, § 1d). Models voicing a group\n  misportray and flatten it (Wang, Morgenstern and Dickerson 2025, § 8c), and the strongest pro-simulation\n  paper says in its own text that it does not license simulating an individual (Argyle et al. 2023, § 8a).\n  So AN3 has outside grounds, and a what-if view must refuse to run on premises the system inferred for a\n  person, or show them marked as inferred.\n- **Do not sell the view as humbling or moderating.** Explaining how a policy would work lowers people's\n  sense that they understand it; that part replicates. That it moderates their views did not survive close\n  replication (Crawford and Ruscio 2021, § 4b), and a later study found moderation only for issues argued by\n  consequences, never for issues held as protected values (Sloman and Vives 2022, § 4c). The view's warrant\n  is legibility, which is the project's own stance.\n- **Arguments alone moved nobody in the field; arguments with the person's own story did, a little and\n  durably** (Kalla and Broockman 2020, § 3a). Opponents respected a moral belief more when it came with a\n  personal experience of harm than with facts (Kubin et al. 2021, § 3c). AN11's line again, now measured: a\n  view of four reasons shows two arguments; with the stake in the holder's words it shows two people.\n- **Cruxes are usually a crux for one side, not both, and finding them did not close the gap.** In the\n  Forecasting Research Institute's structured AI-risk debates, each side could restate the other's arguments\n  well and neither moved; the best near-term crux was expected to close about 5 percent of the gap; the two\n  sides' most informative cruxes were mostly different ones (§ 9a). So the two-person view reports *what\n  carries your conclusion* and *what carries theirs* separately, and a shared crux only when there is one.\n- **Deep disagreement is about what counts as relevant and as evidence** (Stanford Encyclopedia on Fogelin\n  and the later literature, § 6c). That is the weighting place of § 4, the relevance claims, so the view\n  must be parametrised by which relevance claims a holder accepts, not only by which premises.\n- **Same conclusion, different reasons, is a result worth showing**: Rawls's overlapping consensus (§ 6a);\n  and *agree on this case, differ on the principle* is Sunstein's incompletely theorized agreement (§ 6b),\n  which is exactly what a trial of two rule sets over cases finds (§ 6 of this note).\n- **A tool for one person alone, reasoning on their own premises, tends to entrench** (Mercier and\n  Landemore 2012, § 6d). Built for one reader, the what-if view needs the other position beside it.\n- **Feeling understood and being understood are only modestly linked** (Reis et al. 2017, § 1g), and\n  perspective-taking raised confidence faster than accuracy (§ 1a). The mind-meld image invites false\n  confidence unless the view marks what each holder stated and what the system derived.\n- **Models in deliberation so far optimise agreement**: the Habermas Machine writes group statements people\n  prefer (Tessler et al. 2024), and dialogue with a model reduced conspiracy belief (Costello et al. 2024)\n  (§ 8f). Both are tuned for convergence; the project treats convergence as a prediction that may come true\n  when reasoning is legible, never as the objective (`../convergence.md`). A model's place here is finding\n  candidate premises in a holder's own text for the holder to confirm.\n- **The only evidence on a structured map as such is a pilot**: showing an argument as a map rather than\n  prose removed the link between prior attitude and how intelligent opponents judged it, and a jumbled map\n  did not (Cullen and Sharma 2018, not peer reviewed, § 9c).\n\n### 3b. Formal argumentation, supposition and existing systems ([report](what-if-and-lenses/outside-reading-formal.md))\n\n- **The clamp is an intervention, not a supposition in the full sense** (§ 3.4, with a check of `hinge.py`\n  and `badge_inputs.py`). A clamped claim returns its set strength at once, so its own supporters and\n  attackers are never read: Pearl's do-operator, cutting the arrows into one node. Strength flows only\n  forward (Potyka's directionality, § 4.1), so the view can say *if this were settled, the conclusion would\n  stand at X*, never *if you believed this you would believe that* (the Gärdenfors impossibility is the\n  formal reason that phrasing cannot be made exact, § 3.2). Two duties follow: the view lists the supporters\n  and attackers the grant set aside, which is where the other side will object; and a partial grant (*I lean\n  towards this*) is a different operation, a starting-weight change, and is labelled as one.\n- **A lens has a formal home: value-based argumentation** (Bench-Capon 2003, full text, § 1.1). An audience\n  ranks a few named values, and a conclusion is then accepted under every ranking, under some, or under\n  none. That three-way readout is directly usable, and so is the finding of *which single priority decides\n  a contested conclusion*. Facts stay outside the ranking, and Bench-Capon advises settling factual disputes\n  before value disputes: the same split as AN11's fact place against its weighting place. Two audiences\n  have *grounds for agreement* on a claim when each accepts it under some ordering of its own (§ 1.2), the\n  formal version of a bridge.\n- **A priority is a product of reasoning, not an input fixed in advance** (Bench-Capon, Doutre and Dunne\n  2007, § 1.2), and a preference can itself be an argument that others attack (Modgil 2009, known from\n  secondary sources only, § 1.5). So a lens is not a set of global sliders; *this outweighs that, for me* is\n  the holder's own claim. That is AD22's relevance claim seen from the lens side, and it is the second\n  reason to rule AD22 first. No graded version of a preference-as-claim was found (§ 7 of the report).\n- **A number hides an open conflict** (Horty's reasons as defaults, slides and a review read whole, § 2.1).\n  When two rules conflict and nothing ranks them, the honest result is several outcomes, not a mild lean;\n  a gradual semantics always returns a number and so shows *unsettled* as *slightly favoured*. Two cases\n  stay apart: one lens leaves a conflict open; two lenses settle it opposite ways. A conclusion every lens\n  reaches by different premises (Horty's floating conclusion; Rawls's overlapping consensus in § 3a) gets\n  its own state, not a plain *agree*. Horty also argues that an exclusionary reason is not merely a very low\n  weight, which bears on how exceptions act on relevance claims (§ 9).\n- **Lenses are shown side by side, never merged** (Caminada and Pigozzi 2011, abstract and introduction,\n  § 5.11). Combining views premise by premise can produce a group result whose reasons do not support its\n  conclusion. Their acceptance test fits a common-ground display: could each holder defend what is shown?\n- **Attribution has a known answer, and the hinge is one of four imperfect ones** (Kampik, Potyka, Yin,\n  Čyras and Toni 2024, part, § 4.3). Removal, removal after cutting incoming links, Shapley and gradient each\n  fail a different principle. One failure is § 2's measurement exactly: when many independent supporters\n  carry a conclusion, each one alone scores near zero. A group clamp or a sampled Shapley score shows the\n  cluster; Shapley's contributions also add up to the whole gap, which single swaps did not. The better\n  reader question is counterfactual (CE-QArg, part, § 4.5): *what is the smallest change that flips this?*,\n  with a cost model under which a claim a holder asserted costs more to change than one they only voted on.\n- **People confuse a claim's starting weight with its final standing** (Oren and Yun 2025, excerpts,\n  § 4.6). A holder moving a slider probably reports where the claim should end up. Ask for the final standing,\n  as a range, and back out the rest.\n- **Existing tools stop short of deriving** (§ 5). Kialo's Perspectives redraws vote averages per person,\n  group or an invented persona, derives nothing, and folds truth and relevance into one impact vote. Polis\n  infers groups from votes and imputes missing votes; a lens must show *no stance* where there is none.\n  Consider.it lets a person adopt others' points, a cheap way for a holder to build a lens from chosen\n  contributions. Carneades is the nearest working what-if: an audience is a set of assumptions and weights,\n  an accepted statement holds whatever the arguments say and is shown as the reader's act, and its 2006\n  version lets an argument's weakest premise govern, as Q38 does. Community Notes publishes about one note in\n  ten and rests on one learned axis; a bridging score favours what both sides already accept, while a hinge\n  favours what a conclusion rests on, so the claims worth showing are those **high on the hinge and split\n  across lenses**.\n- **Policy what-if tools refuse to predict** (§ 6). Catala, read whole, stops with a conflict error when two\n  equally ranked exceptions apply and returns an empty result when no rule decides: the normative trial's\n  two outputs, which for rules under dispute are results to show rather than faults. OpenFisca states it is\n  static, with no behavioural response. RAND's robust decision making sweeps assumptions and reports the\n  regions where a policy fails; assumption-based planning lists the load-bearing assumptions and the signs\n  that one is breaking. Legal hypotheticals (HYPO) are invented cases built to test a rule's edge, marked as\n  invented.\n- **The mind-meld framing overclaims** (summary point 10). The view shows what follows from stated premises\n  and weights, not the other person's reasons or background. Granting a premise can make a holder reject the\n  link instead of accepting the conclusion (Jackson's case, § 3.1), which a clamp cannot show; the reader\n  needs a way to mark a link as no longer accepted under the grant. And *holds under every lens* covers only\n  the lenses entered (Perelman's universal audience is projected by whoever builds it, § 1.3).\n\n## 4. What a lens is, in the objects the graph already has\n\nHis question of 28 March asked for the dials. The corpus already proposes an anatomy: **six places two\nminds part** (ledger AN11, unruled; [causes-of-belief-and-the-three-minds.md](causes-of-belief-and-the-three-minds.md)\n§ 10): a fact, a prior, a sense, a weighting, a standard, a stake. The first four are reasons; the last two\nare causes, entered by holders. Read against the what-if computation, the six split cleanly by what can be\n*supposed*:\n\n| Where two minds part (AN11) | What a holder's lens supposes | Can the what-if view compute it today? |\n|---|---|---|\n| a fact | the claims the holder accepts or rejects (their agree and disagree, the claims they wrote) | **yes**: `overrides` on those claims |\n| a prior | the base-rate claims the holder accepts (Q35, the starts dial) | when starts are claims (Q35, open) |\n| a sense | which sense of a contested word the holder uses | when senses are stored per use (AM7, ruled, unbuilt) |\n| a weighting | which considerations bear for the holder, and how much: their acceptance of relevance claims | **when AD22 is ruled and built**: the application sentence is the link's relevance claim, so a weighting is a supposition on a claim like any other |\n| a standard | the same evidence judged by which side it serves | **no, and it should not**: a standard is read across a population by the symmetry instruments (AN6), never supposed |\n| a stake | what the position protects for its holder | **no, and it should not**: a stake is testimony, from the holder only (AN3) |\n\nAnd one place AN11 does not list, added by the rule-kind work: **what the holder finds in need of\nexplaining** (AT7), carried by which rival wording of a rule they accept (AT2's contested clause), so it is a\nsupposition on claims too.\n\nTwo things follow. **The weighting dial, the one his lens page names as where people diverge** (*\"the premises\nare shared; the weighting is where you diverge\"*), **is the one that waits on AD22.** Until a link's relevance\nis a claim, a holder's weighting has nothing to be held in supposition; this is the strongest practical reason\nyet to rule AD22. And **the view computes reasons and shows causes**: the four supposable places are computed,\nthe two causal places are shown beside them in the holders' own words, which is AN11's own line, *\"A mind-meld\nthat shows only the first four shows two arguments. With the last two it shows two people.\"*\n\nVotes stay telemetry for the shared graph (ruled 2026-08-31). A lens uses them as suppositions, held apart,\nmoving no strength anyone else sees: the same standing as AM5's `supposed` claims. None of it asks a model\nto play anyone (AN3).\n\n## 5. The mind-meld as a computation\n\nTwo people, A and B, on one claim. The view computes three things:\n\n- **Where each lands**: the conclusion under A's suppositions and under B's.\n- **Where they agree**: the claims both accept, and how far the conclusion would go on those alone.\n- **Where the difference lives**: the smallest set of premises whose change carries most of the gap,\n  attributed fairly when the effects interact (§ 2's single swaps overcount).\n\nThat third line is the position fingerprint designed in [lowering-the-cost.md](lowering-the-cost.md) § 3\n(*you already agree with 14 of these 17 premises; here are the 3 that matter*), computed between two\npeople rather than between a reader and a source, and it is double crux done by the graph. It is also what\nthe lens page's 2026-08-27 section says the lens is for: not transcendence of one's position but *holding\ntwo positions in one addressable structure, so the gap between them becomes an object rather than an\natmosphere*.\n\n## 6. The policy trial is the same view over rules\n\nA policy is a set of rules. Held in supposition (asserted by nobody, AM5), with the case facts, the same\ncomputation answers: which outcomes follow, which rule the outcome leans on most (the hinge, now over\nrules), and where two rules conflict with nothing to settle them. That is the normative trial of\n[normic-rules-normative-reasons-and-trying-rules-out.md](normic-rules-normative-reasons-and-trying-rules-out.md)\n§ 4a (ledger AT5). Its consequence half (AT6, AT10) adds the behavioural rules scoped to the policy (AT9,\nruled and built as a critical question on three schemes), and the hinge over those scoped rules is RAND's\nload-bearing assumptions. So the policy trial needs no separate engine: it is the what-if view whose\nsuppositions are rules. Two things from § 3b shape its output: it reports *rules conflict with nothing to\nrank them* and *no rule decides* as results, the two outputs Catala treats as errors; and for the\nconsequence half it names the behavioural rules it holds fixed and, following RAND, reports where the verdict\nflips across them rather than one predicted outcome.\n\n## 7. Hazards\n\n- **The ultimate tool for rationalization.** A member of the 2011–2017 group warned in 2013 that such a tool\n  would let people pick superficially reasonable premises to reach the conclusions they wanted\n  ([what-the-group-held-2011-2017.md](what-the-group-held-2011-2017.md) § 2). A what-if view is exactly that\n  tool if it shows only the conclusion. The guard follows from the computation: show every consequence of\n  the granted set, not only the one wanted; show which granted premises the graph currently reads as weak or\n  refuted; and in a two-person view, show both sets side by side, so nobody's suppositions are hidden.\n- **Supposing is not learning.** Fixing a claim's strength changes what it supports. It does not change the\n  claims beside it that would also be likelier if it were true, and it silences the claim's own attackers\n  (an intervention in Pearl's sense, § 3b). The view answers *if this were settled and nothing else\n  changed*, says so, and lists the attackers and supporters the grant set aside.\n- **Today's arithmetic is inert one premise at a time and inflated many at a time** (§ 2).\n- **Tourism.** The lens page's Žižek section: if switching lenses feels like browsing a museum, it has failed.\n- **Selling it as moderation.** The claim that explaining a mechanism humbles people did not replicate\n  (§ 3a). The view's warrant is that the gap becomes visible, not that anyone moves.\n- **The image of a mind-meld invites false confidence.** Feeling understood and being understood are only\n  modestly linked, and taking a perspective raises confidence faster than accuracy (§ 3a). Every premise in\n  a lens is marked as stated by its holder or derived by the system, and each holder sees and confirms their\n  own side before the other sees it.\n- **A model filling in a holder.** Measured to misportray and flatten (§ 3a). Premises inferred from a\n  person's votes or their other writing are put to them as questions, never computed silently (AN3).\n- **Two cruxes shown as one.** In the forecasting debates the cruxes were mostly a crux for one side (§ 3a).\n  The view names each side's own crux, and a shared one only when both carry it.\n- **Protected values.** For a position held as a protected value, supposing consequences moves nothing\n  (§ 3a), and argument produces outrage where symbolic recognition produces flexibility\n  (`coupling-the-poles.md`, the sacred-values finding). The four computed places are not shown as if they\n  were the whole: the stake stands beside them in the holder's words.\n- **Ranking premises by how far they would move someone.** That is an agreement objective under another\n  name (the Habermas Machine's, § 3a). The view ranks premises by how much of the gap they carry, which is a\n  fact about the graph, never by how persuadable a person is.\n- **One reader alone entrenches** (Mercier and Landemore, § 3a). The single-reader what-if shows the\n  position that disagrees beside the reader's own whenever one is on record.\n- **Merging lenses manufactures a position nobody holds** (the discursive dilemma, § 3b). Lenses stand side\n  by side; a common-ground display passes only what each holder could defend.\n- **A number hides an open conflict** (Horty, § 3b). Where a lens ranks neither of two conflicting\n  considerations, the view says *unsettled*, not *slightly favoured*.\n- **Imputing a stance** (Polis's missing-vote fill, § 3b). A claim a holder never marked shows *no stance*.\n- **A slider reports a final standing, not a starting weight** (Oren and Yun, § 3b).\n- **\"Holds under every lens\" covers only the lenses entered.** Said as *holds under all N lenses entered*.\n\n## 8. Proposals, unruled\n\nIn the ledger as Arc AU, rows AU1 to AU13 in this order; none ruled. Spoken names first.\n\n1. **One door, three sources.** A reader's suppositions, a holder's lens and a policy's rules are three\n   sources for one computation; it is built once. Extends AM5 (`supposed`) and AT5.\n2. **The first form is the return edge on one branch** (L9, open). After a descent, the reader is offered\n   *this branch from the position that disagrees*: that holder's stated premises, in their words with the\n   quote and the person attached, and the branch's conclusion recomputed under them. It needs holders on\n   record for the branch's claims and the `overrides` arithmetic, nothing else; the whole-graph lens follows\n   when the corpus is dense enough. The research adds a reason L9 did not have: the measured route to\n   accuracy about another person is their own testimony (§ 3a), and the return edge is built from it.\n3. **A lens is computed only from what its holder stated**: their agree and disagree marks, the claims they\n   wrote, the relevance claims they accepted. Narrows AN3 from *no model plays a holder* to *no premise\n   enters a lens unconfirmed*.\n4. **Rule AD22 before building the weighting dial.** Where people most often part is in what bears and how\n   much (§ 3a, the deep-disagreement literature; the lens page's own sentence). Until a link's relevance is\n   a claim, that place cannot be supposed. The formal side agrees from the other end: a priority is the\n   holder's own attackable claim, not a slider set in advance (Bench-Capon, Doutre and Dunne; Modgil; § 3b).\n5. **The arithmetic before the view.** Build Q38 (the weakest premise within an argument; the best argument\n   across arguments until their independence is read), and rule § 9j of\n   `retrofitting-the-reasons-channel.md` (multiply within an argument unless the keeper reads dependence),\n   before the view shows a number. On today's arithmetic it shows either the flat 0.029 of § 2 or a sum of\n   reasons that share one source.\n6. **The gap is attributed fairly, and the reader is asked the flip question.** *Which premises carry the\n   difference* uses a sampled Shapley score or a group clamp (§ 3b), not single swaps, which summed to three\n   times the gap in § 2 and score near zero where many independent supporters carry a conclusion. Beside it,\n   *the smallest change that flips this*, costed so that a claim a holder asserted costs more to change than\n   one they only voted on.\n7. **The view says *supposing*, shows every consequence of the granted set, and shows how the graph\n   currently reads each granted premise.** The guard against rationalization (§ 7).\n8. **Success is measured by the holders, not by movement.** Each holder accepts the rendering of their own\n   position, and the other holder judges it as one that side would recognise (the ideological Turing test,\n   § 3a). Whether anyone changed their mind is not the measure.\n9. **The causes stand beside the reasons.** The four supposable places of § 4 are computed; the standard and\n   the stake are shown in the holders' own words (AN11).\n10. **The policy trial reuses the view** (AT5): the suppositions are rules, the hinge runs over rules, and\n    the conflicts and gaps are offered as moves (Catala's two error values, read as results, § 3b).\n11. **Lenses are shown side by side, with a three-way readout.** For each claim: holds under all N lenses\n    entered, under some, under none; and for those reached by every lens through different premises, a state\n    of its own. Never one merged position (§ 3b).\n12. **What the view puts in front of a reader first: claims high on the hinge and split across lenses.**\n    A bridging score finds what both sides already accept; the hinge finds what the conclusion rests on; the\n    claims where the two meet are where the disagreement is worth reading (§ 3b).\n13. **A grant may reject the link instead.** Granting a premise can make a holder drop the relevance claim\n    rather than accept the conclusion; the view lets them mark that, and records it as their act (§ 3b).\n\n## 9. What is not settled\n\n- **The numbers in § 2 change with § 9j.** If premises within an argument multiply, a required part denied\n  still caps the whole, but the reasons that today move the conclusion by 0.026 each move it differently.\n  The measurement is rerun when the arithmetic changes.\n- **How independent arguments add** (log-odds, or the summed energy as shipped) is open in Q38, and it sets\n  how far a whole position like person A's reads.\n- **A holder's mark as a supposition: at full strength or at a degree they state.** An *agree* entered as\n  1.0 overstates most people. Whether a lens takes a degree from the holder is unasked.\n- **What a partial lens does with the claims its holder never marked**: leave them at the graph's reading,\n  or at *unexamined*. The first mixes the holder's view with everyone else's; the second makes most lenses\n  empty. Unmeasured.\n- **Consent.** A lens is a person's stated position. Showing it to someone else, by name, needs that\n  person's agreement; the rule for it is not written.\n- **Whether a valid exception should cut a link completely.** Horty argues an exclusionary reason is not just\n  a very low weight; today an exception lowers a relevance claim numerically. This bears on the exception\n  rulings, not only on the view.\n- **A lens over a few named values, or over a holder's acceptance of each claim and link.** The first is\n  legible and cheap to sweep (Bench-Capon); the second is closer to what holders actually say. The reading\n  does not decide between them, and no graded semantics for *A outweighs B for this holder* was found.\n- **Whether showing where a gap lives helps anyone** is untested here. The only study of a structured map as\n  such is a pilot (§ 3a); the first workshop is the place to measure it (proposal 8).\n\n## Cross-references\n\n`../worldview-lenses.md` · `../analysis-and-attunement.md` · `../bridging.md` · `lowering-the-cost.md` § 3 ·\n`three-minds-and-the-leap.md` § 1 · `causes-of-belief-and-the-three-minds.md` § 5 (AN3) ·\n`normic-rules-normative-reasons-and-trying-rules-out.md` § 4–5 · `what-relevance-is.md` § 3b–3d ·\n`starting-values-and-the-problem-of-the-prior.md` § 6.6 · `what-the-group-held-2011-2017.md` § 1.4, § 2, § 4 ·\n`synthesis-build-plan.md` § 10 · `deliberus/hinge.py` · `deliberus/graph/badge_inputs.py`\n(`compute_badge_recursive`, `overrides`).\n"}