{"path":"research/shared-sensemaking-what-deliberus-holds-and-owes.md","content":"# Shared sensemaking: what Deliberus already holds, and what it still owes (2026-10-03)\n\n**Status**: research, nothing ruled. The proposals are ledger Arc V (`session26-open-rulings.md`).\n**Added 2026-10-05**: § 9, a live instance from Swedish government formation, read against the five\nproperties.\n**Origin**: the founder, after the AI-risk intake of 2026-10-01: *\"Think deeply about how Deliberus could\nembody the shared sensemaking discussed and go through the details of what's already implemented and what\nmore we need to realize, guided by the existing docs corpus. Ground extremely deeply.\"* This came ahead of\nthe four open reviews because he asked for it; its proposals go to the ledger, so it feeds them.\n\n**How it is grounded.** Everything called *built* below was checked in the code on 2026-10-03, not taken\nfrom a doc. Everything called *designed* cites the doc and section that designed it, and founder rulings are\nquoted. The outside sources are the ones read in full this week:\n- [ai-risk-for-mathematicians-and-the-coordination-game.md](ai-risk-for-mathematicians-and-the-coordination-game.md);\n- [four-cuts-of-the-ai-debate.md](four-cuts-of-the-ai-debate.md);\n- [schmachtenberger-and-the-ai-race.md](schmachtenberger-and-the-ai-race.md);\n- Michael Chwe's *Rational Ritual* (2001), chapter 1, which the corpus had cited once without reading;\n- three contributions to the *Philosophical Transactions A* issue on world models, and the discourse map\n  built over it (§ 6).\n\n---\n\n## 0. In short\n\n- **What the material means by shared sensemaking** is narrower than \"people understanding each other\".\n  A group holds one picture of what it disagrees about. The picture is:\n  - faithful to each holder;\n  - located, and typed by kind;\n  - current, and says what would change it;\n  - known by each member to be held by the others.\n\n  Choices are then made from that picture. Deliberus is built for most of this. It is not built for the\n  last item on the list, nor for the step from the picture to a choice.\n- **What it already holds is substantial.** One addressable object, maps over it, reports kept as reports,\n  disagreements located and typed, an aggregate incoherence check, and compartments a community can\n  deliberate in.\n- **Five things are designed in the corpus and unbuilt, and each is load-bearing here:**\n  1. the act of consent to a rendering of one's own position;\n  2. positions read from what people asserted rather than set by whoever draws the chart;\n  3. the position fingerprint, a diff rather than a tree;\n  4. a slot for the conditions under which a holder would revise a claim or restrain an action;\n  5. the shared event that turns a public record into common knowledge.\n- **Where it stops.** Deliberus can make reasons common knowledge. It cannot verify behaviour, and the\n  game theory, read in full, makes a shared picture necessary for escaping a race but not sufficient. The\n  audience a public map can serve is the public's political will, not the racers' bilateral assurance. The\n  verification literature narrows ledger U9: frontier *creation* can be verified without watching people,\n  while capability that has already *spread*, and secret work on algorithms, cannot.\n- **Where shared sensemaking hurts**, at full strength:\n  - capability information can speed a race up;\n  - publicity can break agreements that held because they were vague;\n  - a group's defences of its beliefs trade truth for the cohesion it acts on;\n  - more reasoning capacity hardens the incurious.\n\n  The corpus has designs against some of these and nothing against others (§ 4).\n- **A sibling instrument exists, and it shows what to build against.** A discourse map of a Royal Society\n  special issue on world models sorts fifteen debates into genuine disagreement, complementary and\n  talking past. Read against the papers, it does three things a map drifts into unless built against\n  them. Some of its quotes follow preprints, though every quote is flagged verified. It turns hedges into\n  poles. And it shows a relationship between two papers that neither states as a finding (§ 6).\n\n## 1. What the material means by it\n\nThe week's sources state the requirement in their own terms, and the terms differ in useful ways.\n\n| Source | The requirement, in its own words | What it adds |\n|---|---|---|\n| Schmachtenberger, *Joe Rogan Experience* #1736 | *\"if you can't make shared sense of the world, you can't act effectively on issues\"*; and *\"can we make sense of anything collectively, adequately, to be able to make choices effectively in the environment we're in?\"* | Two stages: making sense together, then choosing |\n| RAND RR-A4245-1 | Cooperation becomes possible when *\"a common shared assessment emerges\"*, and is implementable only with mechanisms *\"to establish common knowledge and verify mutual commitments\"* | Common knowledge and verification. The report leaves *common knowledge* undefined, and its model assumes the assessment is already shared |\n| Papenhausen, *The AI Race is Not a Prisoner's Dilemma* | In a stag hunt, *\"once both sides are sure the other will comply, a treaty doesn't need enforcement … The hard part is verification and assurance\"* | Assurance: each side must know the other would hold back |\n| Katzke & Futerman, *The Manhattan Trap* | *\"The challenge now is to establish mutual understanding of the dangers of an ASI race\"* | Mutual understanding named as the task now |\n| Joe Carlsmith, *On restraining AI development for the sake of safety* | *\"widespread political will is the first and most central bottleneck to serious capability restraint\"* | The public's will, ranked first |\n| Erik Hoel, *Humanity's Flight 93 Moment* | Time for the public *\"to digest the situation and vote on how much they want their world to be transformed, and how quickly, and why\"* | Time, and the reasons along with the decision |\n| Xiaoyu He's essay, S19 | *\"Coordination is confusing, so humans will probably fail to do it\"* | Confusion named as the obstacle |\n| A commenter on the LessWrong crosspost of Papenhausen | *\"Politicians and the public haven't started thinking in these terms, but they will, and we should be ready to make that debate as sensible as possible before it happens.\"* | The debate prepared before it is needed |\n\n**Five properties and a bridge.** Read together, the requirement decomposes into five properties a\npicture of a disagreement can have, plus a bridge to choice. This decomposition is a proposal.\n1. **Shared**: one object everyone can point at.\n2. **Faithful**: each position as its holder would accept it.\n3. **Located and typed**: where the disagreement actually sits, and what kind it is.\n4. **Common**: each knows the others hold it too.\n5. **Current and conditional**: it moves when the world or the people move, and it says what would move\n   them.\n\nThe bridge is the step from the picture to a choice.\n\nThe deliberative-democracy literature has a name for the first three together: *meta-consensus*, Dryzek and\nNiemeyer's term for agreement on what a disagreement is about. Their papers were not read in this pass, so\nthe term is a pointer here, not an authority. The corpus's own version is older. Bridging says *\"Where\nreasoning stays opaque, you can count agreement but never locate it\"* (`../bridging.md`). The worldview-lens\ndoc says a lens holds *\"two positions in one addressable structure, so the gap between them becomes an\nobject rather than an atmosphere\"* (`../worldview-lenses.md`, § What the lens is actually for).\n\n**Why it matters, and why it is not enough.** The game theory read this week (He doc § 3) gives the\nproperty its stakes. Where the AI race is a trust problem rather than a dilemma, the shared assessment is\nmost of the incentive: each side already prefers mutual restraint and needs to be sure of the other. Where\nit is a true dilemma, no picture is enough, and payoffs must change by other means. Read in full, the\nsources add a third caution: even inside the trust problem a shared picture is necessary, not sufficient.\nPlayers who understand the risk still race when they cannot see each other's actions. Assurance between a\nfew governments and labs runs on classified information no public map holds. What a public picture can\nserve is the bottleneck Carlsmith ranks first, the public's political will, which is Hoel's time for voters\nto digest the situation (He doc § 3d, limits 4 and 5). Both of the first two halves were in the\ncorpus before this week: the founder's 2011 *\"we need to build a collaborative sensemaking platform,\nDeliberus or something like it\"*, and the March doc's warning that sensemaking without a coupling to action\n*\"changes no outcomes\"* (`zizek-schmachtenberger-connections.md` § 3.4, refined 2026-10-01).\n\n## 2. Property by property: built, designed, missing\n\n### 2.1 Shared: one object everyone can point at\n\n**Built.**\n- Every claim is a page and a link.\n- Claim and source pages open on a computed map of about ten boxes, supporters on one side and challengers\n  on the other. Every box is a link, and one place is kept for the region's rarest claim\n  (`specs/map-view/design.md`; live since 2026-09-20/21).\n- Machine readers get the same object as markdown (`deliberus/agent_readable.py`, `llms.txt`).\n- A community can hold its own compartment: a separate graph whose readers are always told which one they\n  are in (`deliberus/blob.py`, `deliberus/scope.py`; UX principle P25, ruled 2026-09-22, *\"agreed\"*).\n\n**Designed.** P23: *\"If a claim is readable it is contestable, and the interface must make that true\neverywhere.\"* The rival is named honestly: prose with permalinks would also be a shared object\n(`structure-versus-scale.md`).\n\n**Missing.** Little in the object itself. What is unobserved is whether people *point* at it. P20 makes\nthat the signal that would redirect the build, and no two people have yet used the object together.\n\n### 2.2 Faithful: each position as its holder would accept it\n\n**Built.**\n- **Reports are kept as reports.** A reported position (*\"opponents argue that X\"*) is minted as its own\n  claim and linked by a `REPORTS` edge that carries no strength. This runs as a post-store pass on both\n  extraction paths (`deliberus/reported_speech.py`, `deliberus/extraction/post_store.py`), and the badge\n  excludes `REPORTS` by name (since 2026-09-17).\n- **Attribution**: `attributed_to` sits on most substantive claims (1,082 of 1,642 when last measured). Its\n  normalization is designed and not yet written to the graph (`deliberus/attribution.py`).\n- **Disagreement preservation** measures how far a paraphrase moved from its source\n  (`GET /extractions/{id}/disagreement-preservation`). It measures one hop\n  (`grounded-critics-and-quarantined-imagination.md`).\n\n**Designed and blessed, not built: consent.** The ratification end-state, blessed 2026-08-31: there is no\nratify act; some contributions have a sole possible author, and that is all consent ever meant\n(`what-human-judgment-is-for.md` § 4). The act is *\"I do not endorse that rendering as a statement of my\nposition\"*: a speech act, not a report, so it cannot be factually wrong. Its value is that *\"the person\ngets to refuse a rendering and supply their own, and the graph then holds both\"* (same doc, act three). The\nconsent row of that doc's table reads *\"Ship it.\"* The corpus has designed it from three sides:\n- `lowering-the-cost.md` § 3: *\"A steelman is trustworthy when the person steelmanned signs it\"*;\n- `incentives-analysis.md` § 5: *\"let opponents propose your premises, and let you only ratify or contest\n  them\"*, the defence against strategic silence;\n- `keeper.py`: its prompt calls sameness *\"a consent test\"*.\n\nIn the code there is no such act: no endpoint and no button, and nothing tells a person that their position\nhas been rendered.\n\n**What this week showed.** In two rounds of paraphrase, Aaronson's *accept* moved from believing that it is\nhappening to resigning oneself to it (four-cuts § 1). A chart set people's positions instead of reading them\n(four-cuts § 3). Neither is malice, and both are what an unconsented rendering does. The Aaronson chain is\nthree texts long and the instrument reads one hop of it.\n\n### 2.3 Located and typed: where the disagreement sits, and what kind it is\n\n**Built.** This is the most developed property.\n- **The hinge**: which premise, if granted or denied, most moves a claim. It is computed for one root at a\n  time (`GET /claims/{id}/hinge`, `deliberus/hinge.py`).\n- **`worth_asking`**: critical questions ordered by how much their answer would move, with a stopping rule\n  (shipped 2026-08-23).\n- **Horizon readings**: where shared mapping currently ends, read from the arrangement of claims and never\n  set (`GET /claims/{id}/horizon`; ruled 2026-09-15).\n- **The keeper of the seven keys**: before any relation is written, it asks whether two claims are the\n  same, contained, overlapping, opposed, unrelated or undecidable (`deliberus/keeper.py`).\n- **The weighing scheme**, with its fourteen questions (`deliberus/weighing.py`).\n- **The concept layer**: contested terms, their senses, and the interpretation checkpoint.\n- **Stance conflicts** (`GET /graph/stance-conflicts`).\n\n**Designed, not built.**\n- **The position fingerprint**, which `lowering-the-cost.md` § 3 calls its *\"Highest-leverage new item in\n  this document\"*:\n  *\"you already agree with 14 of these 17 premises; here are the 3 that matter\"*. It is double crux at\n  corpus scale. `incentives-analysis.md` § 4.3 adds why it is cheap: *\"The feature that needs reader data\n  and the only contribution readers will give are the same thing.\"* Per-person votes already exist as\n  `AGREES_WITH` and `DISAGREES_WITH` edges (`deliberus/api.py`), and nothing compares two people's.\n- **The cross-root fan-out**: one answer closing many open gaps (`curiosity-as-growth-fuel.md` § 9.1).\n- **Pairwise term mapping**: *\"when they say freedom they mean roughly what you would call\n  non-interference\"* (`lowering-the-cost.md` § 3).\n- **Sense attribution**: who uses which sense. It is *\"modelled but not stored\"*\n  (`core-periphery-and-the-concept-layer.md`).\n\n**A worked case from this week (§ 3 below).** Papenhausen's threshold turns the race's type on two\nquantities of different kinds: an estimate of the chance of doom, and a weighing of how bad doom is against\nhow good winning is. The graph types both: one is an empirical claim that takes evidence, the other a\nweighing that takes the weighing questions. That is the property working as designed.\n\n**A class the corpus names in prose and not as a reading.** The discourse map of the *Philosophical\nTransactions* world-models issue (§ 6) sorts every debate into three:\n- *genuine disagreement*;\n- *complementary*;\n- *talking past*.\n\nThe corpus has the third as an idea: the founding conviction says lowering the cost *dissolves*\nconfusion-conflict, and He's S19 reason 1 is one. It does not have it as a reading on a pair of claims.\n\n### 2.4 Common: each knows the others hold it too\n\nThis is the property the corpus has thought about least, and the one this week's sources lean on most.\n\n**What common knowledge requires.** Chwe's chapter 1, read for this doc, puts it exactly: *\"knowledge of\nthe message is not enough; what is also required is knowledge of others' knowledge, knowledge of others'\nknowledge of others' knowledge, and so on—that is, 'common knowledge.'\"* What generates it is publicity of\na particular kind. *\"A public ritual is not just about the transmission of meaning from a central source to\neach member of an audience; it is also about letting audience members know what other audience members\nknow.\"* Two people manage it *\"simply through eye contact\"*. An inward-facing circle works because *\"each\nperson can visually verify that others are paying attention\"*.\n\n**What that means for a public graph.** A public page gives mutual knowledge to whoever reads it: each\nreader knows the content. It does not tell any reader who else has read it, so on its own it does not give\ncommon knowledge. RAND's *\"common knowledge\"* and Papenhausen's *assurance* need the second kind.\n\n**Built.**\n- Public pages give mutual knowledge to their readers.\n- Inside a compartment the membership is known, which makes it a known audience.\n- The read log records who read what. By ruling it is the keeper's alone (2026-09-27, *\"Read log should\n  only be visible to me\"*), so it cannot serve as a common-knowledge device. That is deliberate, because a\n  shared read log is surveillance.\n\n**Designed, and newly warranted.** P21, co-present use, decided 2026-08-17: *\"Two people in the same room,\ntwo laptops, one shared graph, discussing through Deliberus while the disagreement is still live.\"* Its\nstated reason was attunement: the affective channel does not survive delay. Chwe gives it a second,\nindependent reason. Two people looking at the same map with each other in view are an inward-facing circle,\nso the session generates common knowledge of the map. **That is an argument for P21 that the corpus did\nnot have**, and it holds even if the attunement reason were dropped.\n\n**Missing.**\n- **The co-present session has never run.** The rehearsal's obstacles are listed in\n  `islands-of-coherence.md` § 5c:\n  - an account wall at minute one;\n  - rate limits measured at 3–10 calls a minute;\n  - two to three minutes per extraction;\n  - a stream path with no forensics row and no cleanup;\n  - *\"Concurrent contribution is untested\"*.\n- **There is no artifact that carries common knowledge beyond the room.** For that, two or more parties\n  would each mark a dated state of the map as *this is what we disagree about*, and each would see the\n  other's mark. That is the textual form of eye contact, and consent at the level of the pair rather than\n  the person (V5).\n\n### 2.5 Current and conditional: what would move each holder\n\n**Built.**\n- **Staleness**: dates on every claim, a flag-tier staleness daemon, and validity lapses ratified with\n  their grounds (`POST /claims/{id}/validity`).\n- **Edge supersession**.\n- **Lifecycle states** that keep superseded material off default surfaces.\n\n**Missing, in the ontology rather than the code.**\n- **No epistemic status means *open*.** All six assert something (`EpistemicStatusLit` in\n  `deliberus/extraction/models.py`). The governance map's twenty-four open questions had no faithful home.\n- **No slot for an update condition.** The same map reports, as its first finding about the ontology, that\n  *\"Update conditions have no slot\"*: a holder's *I would revise this if X* is *\"a pre-registered defeater\n  written by the claim's own holder\"*, and *\"the single most useful thing in the file for disentangling a\n  disagreement\"*. The map kept them in a field no instrument reads\n  (`mapping-a-structured-position-on-ai-governance.md`).\n- **An outside protocol makes it the entry condition.** Schmachtenberger describes a facilitated debate\n  (JRE #1736, 2021; the speaker is probably him, as the transcript has no labels) whose participants first\n  agree on *\"what would I need to change my mind about this? If the answer is nothing, then you don't even\n  engage in the debate.\"* It then *\"formalize[s] our disagreement\"* and asks whether the gap needs *\"a new\n  piece of science\"*, or is *\"an intuitive hunch or a different value\"*. The protocol makes explicit\n  disagreement, a typed basis and shared ground first, all of which Deliberus has. Its entry condition is\n  the slot Deliberus lacks\n  (`schmachtenberger-and-the-ai-race.md` § 2).\n- **No slot for a conditional commitment** (*A will hold back if B does*). The corpus located this gap\n  twice. The non-zero-sum doc's reading of Dafoe's four capabilities places Deliberus in understanding,\n  communication and institutions, *not* commitment (`non-zero-sum-economics-and-civilizational-cooperation.md`\n  § 6). The parity doc rules that commitment is not platform machinery (`parity-and-the-zone-interior.md`).\n\n**Why this matters for the race in particular.** In a stag hunt each side would restrain if it could count\non the other. A public, attributed statement of the conditions under which each would restrain is the part\nof assurance that talk can carry, and it is what RAND's common knowledge would be common knowledge *of*.\nThe graph could hold such statements as claims and let them be examined: are the conditions coherent, are\nthey credible, what would count as meeting them? It could not enforce them or check compliance (§ 2.6 and\nU9). There is a hazard the corpus already names. A commitment structured so that later parties cannot revise\nit is *deliberate lock-in*, a strategic move (`fractal-scales-and-temporal-frame.md`, the temporal rung).\nAnything the graph holds here must stay dated and supersedable. Proposal V4, which is an ontology decision\nand the founder's.\n\n### 2.6 The bridge: from the picture to a choice\n\n**Built.**\n- **Compartments with an advice process**, piloted with a real community: decisions made on shared reasons\n  with no counting step (`blobs-community-sandboxes-and-the-advice-process.md`).\n- **Read, never set** everywhere a kind or a position is shown, so the map never chooses for anyone.\n\n**What the corpus warns.**\n- **The coupling critique**, which the 2026-10-01 refinement sharpened (`zizek-schmachtenberger-connections.md`\n  § 3.4). Where the game is a trust problem, the shared assessment is most of the incentive. Where it is a\n  true dilemma, payoffs must change by other means: law, liability, mechanism design. A friend's second\n  document, held privately, is one such mechanism (He doc § 5).\n- **Institutional orphaning**: tools *\"built without governance connection; outputs go nowhere\"*\n  (`ai-augmented-cooperation-infrastructure.md` § 5).\n- **The epistemic constitution**, demanded by three docs and written nowhere\n  (`fractal-scales-and-temporal-frame.md`).\n\n**The founder's own statement of the bridge**, from 2012, is a market analogy: *\"some algorithm for humanity\nto reason and make decisions together in a way that combines the democratic elements and general\nefficiency of a decentralized system like markets with the discerning force of logic\"*\n(`non-zero-sum-economics-and-civilizational-cooperation.md` § 2). The doc's gloss: prices tell you *that*\nresources should flow somewhere, not *why*.\n\n**Missing.** A surface that hands a decision what it needs from the map: the options, the residual\nobjections, the minority positions, and what would reopen the question. The corpus cites this shape from a\nstudy of deliberative tools (*\"decision packets\"* with *\"reopening conditions\"*,\n`ai-augmented-cooperation-infrastructure.md` § 4) and has not designed its own.\n\n### 2.7 Trust in the map itself\n\nA shared picture is only shared if each side trusts it was not drawn against them.\n\n**Built.**\n- No verdicts anywhere.\n- Every count labels its provenance.\n- Absence is a state: unexamined and unsettled are shown as such, never as confidence (P22).\n- Synthesis cites the claims it rests on and reports what it left out.\n- The agent surface is read-only.\n- Confession channels on the machine passes.\n\n**Ruled and unbuilt.** A third state beside unexamined and unsettled, *nothing stands (N checked)*,\nruled for now on 2026-09-30. It answers an objection of the founder's that P22 records in its own words:\na middle reading would suggest some truth where no good reason stands.\n\n**What this week added.**\n- **The fake middle.** A public post that reached the no-verdict stance on its own was read by its\n  sympathetic commenters as *\"Sanningen finns någonstans i mitten av allting\"*, the truth is somewhere in the\n  middle (`the-post-that-states-the-problem-from-inside.md` § 6). The founder's stance on steering\n  (2026-08-21) is the answer the map must make visible: *\"it SHOULD be obvious from the graph/ontology if the\n  middle is actually the wise/rational/balanced stance or if the AVAILABLE reasoning and evidence ends up\n  leaning one way or the other GIVEN what people… have contributed up until that point.\"*\n- **Corrections travel less far than claims.** Hoel reports that Anthropic's introspection paper *\"got over\n  a million views on social media\"* and the paper questioning it *\"13,000 views\"*. On Deliberus's own\n  surfaces the correction is attached to the claim as an attack, so a reader of the claim meets it. Off\n  them, nothing carries it. P9, *\"Notify when someone challenged your reasoning\"*, is designed; the code\n  sends no notification to anyone but the operator.\n- **Checking facts is not checking arguments.** Schmachtenberger, in the same 2023 interview: *\"fact-checking\n  is not enough is because you can cherry-pick your facts. You can frame them in a particular way and create\n  a conclusion that the totality of knowledge wouldn't support\"*. That is the case for checking structure\n  rather than claims one at a time. It is also the corpus's coherent-absence class, stated from outside: a\n  frame can mislead with every fact in it true.\n- **Machine reasoning confabulates its reasons.** Hoel's case is that a model's chain of thought is not its\n  reasoning and that models *\"must constantly confabulate the reasons for its own actions\"*. His *\"graduate\n  student from hell\"* drops a raised objection and returns to it later as if it had been answered. That is\n  the forgotten-premise failure the harness conviction names. A persistent record where an unanswered\n  attack stays attached is the corpus's answer, and Hoel's essay is an outside statement of the problem.\n- **The world growing uninterpretable.** In *Don't dethrone consciousness!* Hoel calls the AI revolution\n  *\"a mechanism for transferring the processes of our civilization from under the supervision of\n  consciousness to unconsciousness\"*, which *\"renders the world increasingly uninterpretable\"*. A\n  shared picture answers part of that: machine-made structure that stays legible because every part is\n  marked by its provenance, attributed, and open to attack. It does not make a model's own processes\n  interpretable. It keeps inspectable the reasoning people act on (`four-cuts-of-the-ai-debate.md` § 5).\n\n**Missing.**\n- No instrument here has met a motivated stakeholder (`fractal-scales-and-temporal-frame.md`).\n- The per-cluster ingestion balance is proposed and unbuilt (`the-residual-error-taxonomy.md`).\n- Every instrument is published, and *\"publishing how your safeguard works is the same act as handing over\n  the workaround\"*.\n\n### 2.8 Scale\n\nThe AI race sits on the civilization rung of the scale ladder, which the corpus marks as lineage, not\nroadmap: *\"A rung is earned by an instrument that can fail there, not by an analogy that fits\"*\n(`fractal-scales-and-temporal-frame.md`). Three applications are honest now:\n- a map of the debate cluster, for readers (U3);\n- the two-person rehearsal;\n- a community's advice process.\n\nNothing here claims Deliberus can coordinate labs or states. It claims that the variable the race turns on\nis the one Deliberus works on, which was already the He doc's position (§ 3d).\n\n## 3. A worked example: the race's type, as the graph would hold it today\n\nThis sketch uses only material read this week, and marks for each part whether today's system can hold it.\n\n**The contested claim.** *C0: \"The AI race is a prisoner's dilemma.\"*\n- **Asserted** by He's S17, by Leopold Aschenbrenner (2024) and by Hunter Ash (2026), the last two as\n  quoted in Papenhausen's post.\n- **Attacked** by Papenhausen, Katzke and Futerman, RAND and Carlsmith.\n- *Holdable today*: claims, attributions, attack links, and reports minted as reports.\n\n**Papenhausen's attack, decomposed.** Assumptions:\n- A1, two symmetric players;\n- A2, winning as good as losing is bad;\n- A3, an even race;\n- A4, whoever reaches superintelligence first dominates.\n\nTwo premises of different kinds:\n- **E, empirical**: the chance of doom, as decision-makers estimate it. The evidence offered is the\n  CEOs' stated probabilities.\n- **W, a weighing**: doom is at least nine times as bad as winning is good.\n\nConclusion: above the threshold *q\\* = W / (W + |D|)* the race is a stag hunt. Papenhausen's own caveat\nbelongs to the claim: the feasibility is *\"game-theoretically\"* only, with *\"messy politics,\npersonalities and relationship between the specific parties involved\"* set aside.\n\n*Holdable today*: the decomposition, an empirical premise with evidence links, a weighing with its\nquestions, a horizon reading on the weighing, and the hinge, which would say whether E or W moves C0\nmore.\n\n**The attacks on W**, from a commenter on the crosspost:\n- winning might be worth as much as doom costs, because *\"Win means using the whole resources of the\n  lightcone for your values\"*;\n- for an individual decision-maker, racing *\"can easily be the difference between suffering and dying from\n  aging and living approximately forever\"*;\n- losing to a rival who shares many of your values is not so bad.\n\nEach is an attack on the weighing, and the second is an implicit premise about the decider's personal\nstakes: the corpus's finding that *the crux is left unsaid by at least one side*,\nreappearing in a comment thread.\n\n*Holdable today*: all three, as attributed claims attacking W.\n\n**Two conditions and one observation.**\n- RAND's condition: a common shared assessment plus verified commitments.\n- Carlsmith's caveat: one actor who does not believe the risk.\n- Hoel's observation that Chinese discussion of the risk *\"look[s] a lot like our own\"*.\n\n*Holdable today*: as claims bearing on E and on assurance.\n\n**What cannot be held today, and is exactly what coordination needs:**\n- **Which decision-makers hold E and W.** There is no per-holder reading (U4, V3).\n- **\"Lab X would hold back if Lab Y did.\"** There is no conditional form (V4).\n- **That both sides have seen the same map.** There is no common-knowledge mark (V5).\n- **What would move each holder.** There is no update-condition slot (V4).\n\n## 4. Where shared sensemaking hurts\n\nEach of these is a reason to build some of the above carefully, and some not at all.\n1. **Information about capability can speed a race.** Armstrong, Bostrom and Shulman model teams racing\n   and find that more knowledge of each other's capabilities raises the danger. Every team in their model\n   knows the risk and still cuts safety; the lever they name is enmity, reduced by trust and common aims. A\n   map that turned into a capability scoreboard would feed the race (He doc § 3d).\n2. **Publicity can break an agreement that held because it was vague.** This is Chwe's mechanism in its\n   other face, and the red team's *constructive ambiguity* (`legibility-under-power-red-team.md`). Hence\n   compartments first, and opt-in for the joint mark (V5).\n3. **A group's defences of its beliefs buy the cohesion it acts on.** The corpus's memetic-immune read,\n   drawing on Turchin, concludes that ideological immune systems *\"trade truth for coordination\"*\n   (`memetic-immune-system.md` § 7). Making a group's reasons legible may weaken the cohesion that lets it\n   act. The corpus has no answer to this.\n4. **Reasoning capacity hardens the incurious**, and one unpersuaded player breaks a stag hunt. This is\n   Kahan's result with Carlsmith's caveat (`curiosity-as-growth-fuel.md`).\n5. **Machine summaries flatten toward the middle and drop the tail** (`saturation-and-the-long-tail.md`\n   § 3b). Hence: show structure, ask questions, never restate.\n6. **Open reasoning is exposed.** Hoel's dark forest: work in progress is absorbed and scooped. The full\n   reading narrows this to work whose value is priority (four-cuts § 6).\n7. **Selective legibility.** An actor is legible where it pays and vague elsewhere\n   (`incentives-analysis.md` § 5). The defence is consent's other side: opponents may propose your premises,\n   and you answer.\n8. **Verification of spread capability leans toward surveillance** (§ 5, U9 narrowed).\n9. **Information may not be what is scarce.** Tristan Harris, in the same 2021 conversation: *\"people think\n   about it as a, as an information problem, misinformation, disinformation, it's actually about meaning and\n   identity\"*. A record repairs what a reader can inspect, not what a community needs in order to belong.\n   The corpus's answer is the attunement pole and co-presence, both unbuilt as mechanisms\n   (`../analysis-and-attunement.md`).\n10. **Arguments are funded, not selected.** Schmachtenberger: *\"Good ideas don't get naturally selected, but\n    they get upregulated by whoever it is that has marketing budgets to get that idea out a lot\"*. A public\n    record is a surface to flood. In the graph this is asymmetric ingestion becoming asymmetric strength\n    (`the-residual-error-taxonomy.md`). The per-cluster balance count that would show it is proposed and\n    unbuilt.\n11. **Awareness can raise the pressure to strike.** In Jervis's most dangerous world, where offense has\n    the advantage and postures look alike, *\"the 1950’s would have been more hazardous if the decision\n    makers had been aware of the dangers of their posture and had therefore felt greater pressure to strike\n    first\"*. In Fudenberg and Koh's model, faster detection of a rival's stop is *\"double-edged\"*: at middling\n    trust it can first remove the equilibrium in which firms stop early.\n\n## 5. Verification, narrowed\n\n**What U9 said.** The trust problem's hard part, verification, is where the dystopia attractor lives\n(`schmachtenberger-and-the-ai-race.md` § 3). The refutation condition it carried: it weakens if\nverification exists that does not require watching everyone and holds at the scale of the race. The\nverification literature cached during this intake meets that condition for one half of the problem.\n\n**Verifying creation does not require watching people.** Shavit's compute-monitoring design (*What does it\ntake to catch a Chinchilla?*, 2023) aims to give governments *\"high confidence that no actor uses large\nquantities of specialized ML chips to execute a training run in violation of agreed rules\"*, while it\n*\"does not curtail the use of consumer computing devices, and maintains the privacy and confidentiality of\nML practitioners' models, data, and hyperparameters\"*. The literature's verification methods are aimed at\n*\"unauthorized AI training\"* and *\"unauthorized data centers\"* (Wasil et al., *Verification methods for\ninternational AI agreements*, 2024). A treaty draft from MIRI's technical governance team pairs inspection\nwith satellite and aerial surveillance of large data centres (Scher, Abecassis, Barnett and Abeyta,\narXiv:2511.10783). These watch facilities, chips and supply chains: a few hundred sites, not citizens. The Manhattan Trap\nstates the ideal a privacy-preserving design aims at: verification that *\"would reveal only whether a state\nis compliant with an arms control treaty, and nothing else\"*.\n\n**The gap inside creation is algorithms.** Carlsmith, read in full: compute can be governed for *\"at least\na few years\"*. Verifying that nobody is doing secret algorithmic work, though, needs a surveillance\napparatus that *\"(even modulo AI-powered privacy preservation) quickly becomes extremely scary in itself\"*.\nWithout that verification, *\"the game-theoretical logic at stake in a more general race dynamic gets\nrecapitulated at the level of algorithmic progress alone\"*, even among nations convinced of the risk.\n\n**Verifying spread capability does.** Shavit again: once a violating run has produced a model, its weights\n*\"may from then on be copied undetectably, and it can be deployed using consumer GPUs\"*, and enforcing rules\nat that level *\"would require surveilling and policing individual citizens' use of their personal\ncomputers, which would be highly unacceptable on ethical grounds\"*. Schmachtenberger drew a related\nasymmetry in 2023, between developing a model and running it: monitoring all large GPU clusters *\"to some\ndegree can be done\"*, but *\"it takes a very large GPU cluster to develop an LLM. It takes a very small one\nto run that LLM afterwards\"*. That is from his *Theories of Everything* interview, read in full from its\nautomatic transcript on 2026-10-03. In the same interview he voices the control side of the trap,\n*\"we need ubiquitous surveillance to know what everybody's doing in their basement\"*, and rejects it:\n*\"most versions of ubiquitous surveillance are pretty terrible\"*.\n\n**The window narrows.** Shavit: as chips grow stronger and algorithms more efficient, more sampling is\nneeded to verify a fixed-capability run. Open-weight models trail the frontier by about three months\n(Epoch AI, *Open-weight models lag state-of-the-art by around 3 months on average*).\n\n**U9 should be narrowed, not dropped.** The dystopia attractor lives in verifying capability *after it has\nspread*, and in verifying the *absence of secret algorithmic work*. Creation can be verified at chokepoints\nthat watch facilities rather than people, for as long as frontier training needs monitorable compute. Openness decides which side of that line a capability is on.\nThis also separates two things the governance map's openness argument runs together:\n- **verifying compliance** (who trained what);\n- **verifying an artifact's properties** (what the model does).\n\nOpen weights help the second and hurt the first. (Ledger U9, revised.)\n\n## 6. A sibling instrument: the discourse map of the world-models issue\n\n**What it is.** Adam Safron and Michael Levin co-edited a special issue of *Philosophical Transactions of\nthe Royal Society A*, *World models in natural and artificial intelligence* (vol. 384, issue 2320, 2026),\nwith an editorial and seventeen contributions. Michael Garfield's podcast *Humans On The Loop* opened a\nseries on it with Safron (H-38, 30 July 2026). The episode notes point to *\"this illuminating interactive\ndiscourse map by Van Bettauer of Ideoscopic\"*. The map counts 18 papers, 217 concepts, 28 themes,\n15 tensions and 84 quoted positions. It has an agent that *\"reads the graph, the debates and the full\ntexts, then answers with citations you can touch\"*, and four views; one of their presets is named\n*talking past*. Safron is not new to this corpus. In January 2011 the founder wrote that *\"Adam Safron,\nfamous neuroscientist, will also be joining in\"* a brainstorming session about building what became\nDeliberus (`simplenote-archive-analysis.md`).\n\n**What its data holds**, read from the map's own data file. Each of the 15 tensions is a question with:\n- a *kind*: five framework clashes, four definitional axes, three necessity claims, two empirical claims\n  and one methodological;\n- two poles;\n- a *relationship*: nine genuine disagreements, four complementary, two talking past, each with a note on\n  why;\n- a position for each paper that takes part, with a stance (one pole, the other, or nuanced), a summary,\n  a verbatim quote marked verified (76 matched exactly, 8 after repairing hyphenation or layout), its\n  section, and a confidence.\n\nThat is close to Deliberus's own shape: an attributed position with its source span, typed by the kind\nof disagreement. Two differences matter. The map's unit is a question with two poles, so it shows\n*where* papers disagree and not which premise carries the disagreement, which is what decomposition and\nthe hinge are for. And it is a reading of a closed set of papers that nobody can answer. An author it\nrenders has no way to say *that is not my position*: the act Deliberus has designed and not built\n(§ 2.2).\n\n**What a full reading found in it.** Three contributions were read in full for this doc: the editorial,\nLevin and Lyons, and Collins, Bhatt and Sucholutsky. They let the map's renderings be checked against the\ntexts.\n- **The quotes are real, but some come from a different version.** Every quote for the editorial is\n  exact. Of the others that could be tested against published text, four follow the preprints instead:\n  two for Levin and Lyons (a quote's wording and a section number) and two for Collins et al. All 84\n  quotes are flagged verified. The flag does not say\n  *against what*. Deliberus half-answers this. It keeps the text it read, write-once, with a hash and the\n  time it was fetched (`deliberus/database.py`: *\"keeping the first observation is what makes drift\n  detectable at all\"*). But nothing checks that a stored span occurs in that text. The span is the\n  extraction model's *\"Exact quote from source text\"* (`deliberus/extraction/models.py`), and no code\n  compares the two.\n- **Hedges become poles.** The map puts the editorial on one pole in six of the fifteen tensions (T01 and\n  T03–T07). The editorial mostly hedges, and says of its own synthesis that its interpretations\n  *\"should not be taken to be definitive statements and may diverge from the views of contributing\n  authors\"*. One placement rests on a phrase inside a question, *\"seemingly essential elements for\n  selfhood\"*. The map's summary has the editorial concluding that LLMs *\"likely model the world through\n  statistical surface regularities\"*. The editorial poses *\"statistical surface regularities\"* as one\n  side of *\"a recurring theme\"* it leaves open. Aaronson's *accept* and the two-axis chart met the same\n  fate this week (§ 2.2): a position set by whoever drew the map, with the hedge lost on the way.\n- **A relationship neither paper states is shown as found.** T11 sets Levin and Lyons (one generic template\n  for every cognitive glue) against Sacco, Sakthivadivel and Levin (binding gated by the topology of the\n  interaction graph) as a genuine disagreement. The two papers share an author. Levin and Lyons never\n  mention the other paper or topology, and the editorial describes both without comparing them. Yet the\n  map's own entry has the editorial presenting them *\"as complementary lenses\"*. The Sacco paper was not\n  read here, so the label may be right. What can be seen is that it is the mapper's inference, and\n  nothing marks it as one. Deliberus's rule for machine judgment, *propose, never assert*, would show it\n  as a proposal attributed to whoever made it.\n\nThese are not quirks of one map. A map of a disagreement drifts into all three unless it is built\nagainst them, and they are the failures this doc's proposals target: consent (V2), rendering chains (V7), and\npositions read rather than set (U4).\n\n**What the issue says about shared models.** Two contributions bear directly on § 1's question, and a\nthird on who can take part in inquiry.\n- **Levin and Lyons, *cognitive glue*.** The published version cites it as Levin M, Lyons B. 2026, while\n  the editorial and the map say Lyons and Levin. The thesis is *\"All intelligence is collective\n  intelligence\"*, and the price system is *\"the cognitive glue of the economy\"*: *\"a shared model of\n  parameters that are constructed to track variables, called relative scarcities, that reflect the\n  degree of incompatibility between plans\"*. Members use it to choose compatible plans *\"not out of\n  altruism but simply to maximize the probability of their own plan being fulfilled\"*.\n  - *Against this doc's framing*: the best-understood glue needs no shared understanding. *\"coherence\n    within a group can be achieved without any shared goals among the individual members of the\n    group\"*, and the price system *\"in no way requires that people actually understand the concept of\n    relative scarcity\"*. A message telling members what to do fails, since it is *\"not obvious why the\n    members of the collective would be motivated by such a message, nor why they should believe its\n    veracity\"*. Of their five conditions on a shared model, the second says it must not be *\"'mere\n    information'\"* but must connect with members so that they are motivated by it. The fifth requires\n    its updates to be *\"direct causal consequences of changes in plans\"*.\n  - *For it*: *\"the price system depends on people being able to disagree with it—to reject the shared\n    model when it contradicts other data they have. The price system cannot be corrected … unless the\n    individual members of the collective have the freedom and the competency to defy it\"*. Disagreement\n    is the glue's correction channel, *\"just as the progress of science depends on the ability of\n    individual scientists to disagree with widely shared models\"*. The first of those lines is in the\n    published version and not in the preprint.\n  - *What follows (our reading, unruled).* Measured against the five conditions, a public record of\n    reasoning meets the third, *\"The model should not direct the members of the collective as to how to\n    change their plans\"*, which is *read, never set*. It also builds in the correction channel, since\n    any claim can be attacked and nothing is overwritten. By itself it fails the second and the fifth:\n    nothing ties a reader's plans to the record, and the record changes when someone contributes, not\n    when anyone acts. So a record is a different kind of glue from a price. It carries the *why* a price\n    leaves out, which is the founder's 2012 market analogy and its gloss (§ 2.6), and it charges a reading\n    cost a price never does. Condition (2) is the incentive analysis's question in other words: *when is\n    legibility individually advantageous?* The nearest thing the design has to a motivating connection\n    is consent, because being represented accurately is in the reader's own interest. Their warning\n    belongs here too: *\"Virtual governors are dictators that do not control by force but by adjusting\n    parameters to achieve a distributed consensus among the parts\"*. That is the steering the threat\n    model watches for.\n- **Collins, Bhatt and Sucholutsky, *the Rogers paradox*.** Cheap AI trained on all human data *\"may\n  not, on its own, in the long term, improve our collective world model\"*. What helps is *\"critical\n  social learning\"*, where a learner can switch to learning for themselves after seeing the AI's output.\n  It *\"yields a higher equilibrial state of collective world understanding\"*. Interventions that only\n  made use selective, a transparency variant among them, *\"do not affect population equilibrium\"* in\n  their simulations. That matches the cognitive-commons finding from another field: question-framed help\n  builds skill and finished outputs wear it down (`cognitive-commons-and-deliberus.md`). It also warns\n  against assuming that showing the reasoning is enough. They add that a collective world model *\"will\n  disproportionately represent some experiences and cultures over others\"*, and that an aggregate\n  metric *\"washes out much of this potentially heterogeneous dispersion of understanding\"*, which is the\n  tail argument of `saturation-and-the-long-tail.md` in their terms. **Weigh this before citing it as\n  support.** Their proposed future work is that *\"an AI system could summarize many different humans'\n  responses\"*, citing Tessler et al. 2024, the Habermas Machine. The authors of the later DeepMind\n  facilitation study implicate summarizing as the mechanism that steers\n  (`live-election-test-design.md`). Collins et al. themselves *\"urge caution in interpreting these kinds\n  of simple network models\"*.\n- **Who can take part in inquiry.** The H-38 notes ask *\"what might it take for machines to actively\n  participate in that inquiry?\"*; the episode's answer was not read. The editorial says current systems\n  *\"remain dependent on human guidance for conceptual innovation and problem formulation\"*. Battleday\n  and Gershman (read as arXiv v2 only) picture the first AI scientists as *\"research assistants or\n  first-year grad students\"*, whose growth *\"requires the examination and emulation of the communal\n  aspects of science and related cultural institutions\"*. Their reason is that *\"Lab meetings,\n  conferences, and presentations and discussions are ultimately the place where judgements on the\n  quality of a scientific problem are made.\"* A public record of reasoning is one such communal surface\n  that a machine can read, which is the case `agents-as-a-consumer-class.md` makes from the other side.\n\n**Where the editors land, and what their reassurance does not reach.** The editors display disagreement\nrather than settle it. They write that *\"the term 'world model' has multiple meanings\"* and *\"runs the\nrisk of becoming semantically vacuous\"*. They trace part of the LLM dispute to which meaning is used: on\na weak sense, *\"LLMs clearly reach this standard\"*, while on causal or agentic senses *\"the divergence may\nbecome even greater\"*. Their conclusion: *\"Rather than offering definitive answers, we must refine the\nquestions we ask.\"* That is a talking-past diagnosis made by hand, the reading V8 proposes to compute. It\nis also the concept layer's job: a contested term whose senses split the verdict\n(`core-periphery-and-the-concept-layer.md`). Their reassurance about risk concerns generality and\nself-improvement: *\"recursively self-improving superintelligences may not be one of the problems that we\nhave to face from such systems\"*. Gradual disempowerment needs neither. Its loss is driven by *\"more\ncompetitive machine alternatives to humans\"* (He doc § 4a), and neither it nor Critch and Tsimerman nor\nSchmachtenberger is cited in the editorial. Our inference: the reassurance does not reach that route.\n\n**What Deliberus can take from it.**\n- **The three-way relationship class**, as V8 proposes, read from the concept layer and the keeper\n  rather than assigned.\n- **The tension kinds, as a comparison to run.** The horizon kinds and claim types cover *definitional*\n  and *empirical*. *Framework clash* has no counterpart: the corpus writes about frames (the reframe, the\n  frame gap, frame lock-in), but no reading types a disagreement as a clash of frameworks.\n- **Two rules the map breaks, and how Deliberus stands on each.** A verification flag should name the\n  text it was checked against: Deliberus keeps that text but does not check spans against it (above). A\n  relation the mapper inferred should be marked as inferred: Deliberus marks machine-written links by\n  their provenance (`auto_connect`, `extraction`, `split`).\n- **An agent that answers *\"with citations you can touch\"*** applies the inspectable-synthesis discipline\n  (`synthesis-build-plan.md`). It shares that discipline's risk: a door that hands readers summaries is\n  where the tail gets lost.\n\n**Read depth.** The editorial, Levin and Lyons and Collins et al. were read in full in their published\nversions. Levin and Lyons was checked against the OSF preprint and Collins et al. against arXiv v1.\nBattleday and Gershman were read as arXiv v2 only, and fifteen contributions were not read. The map was\nread from its landing page, its explore page and its data file. H-38's transcript is unpublished, so\nonly its notes and chapter list were read.\n\n## 7. Proposals (ledger Arc V)\n\nEach is unruled, and each is designed against a named failure from this week. Two existing rows carry the\nrules these builds would serve: U4 (*a position is read, never set*) and U5 (*attribute before charging\nincoherence*).\n\n- **V1. The definition.** Shared sensemaking, as this project uses the term, is a group's picture of what\n  it disagrees about that is shared, faithful to each holder, located and typed, common, and current and\n  conditional, with a bridge to choice (§ 1).\n- **V2. Build the consent act before any other sensemaking feature.** On any rendering of a person's\n  position (a minted report, a mapped source, a chart's placement), that person can mark *this states my\n  position* or *it does not*, and supply their own. The graph holds both, attributed. Blessed in principle\n  2026-08-31 and unbuilt; this row proposes it as first in line (§ 2.2).\n- **V3. A per-holder coherence check beside the community-level one.** The discursive-dilemma endpoint\n  already reports when a community's premise votes and conclusion votes point different ways\n  (`GET /claims/{id}/discursive-dilemma`). Before anyone calls a side's case incoherent, the record should\n  also show whether any single holder holds the conflicting claims. This is U5 made buildable, from the\n  vote edges and `attributed_to`, and the same per-holder reading is what U4's placements need.\n- **V4. Update conditions and conditional commitments as attributed, attackable claims of a conditional\n  form.** *H would revise C if E*; *A will hold back if B does*. They are dated, supersedable, and never\n  enforced or verified by the graph. This is an ontology decision. It answers § 2.5 and the assurance half\n  of the stag hunt.\n- **V5. A joint mark of disagreement.** Two or more parties each mark a dated map state as *this is what we\n  disagree about*, and each sees the others' marks. Opt-in, compartment first. It is the textual form of\n  Chwe's eye contact, and the rehearsal's natural output (§ 2.4).\n- **V6. The position fingerprint, built on the existing vote edges and first used in the rehearsal.** *You\n  agree on 14 of 17; here are the 3 that matter.* Designed in August (§ 2.3).\n- **V7. Rendering chains.** Extend disagreement preservation from one hop to the chain of renderings: who\n  reported whom, through how many hands. The first fixture is Aaronson to Romero to O'Connor (§ 2.2).\n- **V8. *Talking past* as a reading on a pair of claims.** It would be computed from the concept layer's\n  sense mismatch and the keeper's verdicts, read and never set, beside genuine disagreement and\n  complementarity. The class is borrowed from the world-models discourse map (§ 6).\n- **V9. Co-present use has a second warrant: it generates common knowledge.** Two people at one map, with\n  each other in view, know that each other knows what is on it, as Chwe's eye contact does. This holds even\n  if the attunement reason for P21 were dropped (§ 2.4).\n- **V10. By itself, a record of reasoning is *\"mere information\"*.** Of Levin and Lyons's five conditions\n  on a coordinating shared model, it meets the third and builds in their correction channel, and fails\n  the second and fifth unless something ties it to readers' own plans. Consent is the nearest such tie in\n  the design (§ 6).\n\n**Sequencing.** Under the ruled order of work, V2 to V8 are builds, so they come after the reviews. Three\nof them, V2, V5 and V6, are what the two-person rehearsal itself needs. They belong to that milestone and\nare not a reason to start building early.\n\n## 8. Cautions\n\n- **The definition in § 1 is a proposal.** It was assembled from this week's sources and the corpus, and\n  the founder has not ruled on it.\n- **Nothing here is verification of behaviour**, and nothing should be described as if it were.\n- **The co-present argument from Chwe is a game-theoretic warrant, not evidence.** Whether two people at\n  one map actually come to know that each other knows is the rehearsal's to observe.\n- **The verification half rests on readings of five documents in one pass.** U9's revision should be read\n  as narrowed, not settled.\n- **§ 6 checked the discourse map only where a paper was read in full**, three of eighteen. Its other\n  renderings were not checked, and nothing here says they are wrong.\n\n**Sources read in full for this doc**: Chwe, *Rational Ritual* (2001), ch. 1 · Papenhausen, *The AI Race\nis Not a Prisoner's Dilemma* (LessWrong crosspost and its comments, August 2026) · Hoel's three essays ·\nSchmachtenberger's 2023 *Theories of Everything* interview (from its automatic transcript) · from the\nworld-models issue, the editorial, Levin and Lyons, and Collins, Bhatt and Sucholutsky (§ 6). **Read in\npart**: Shavit (2023), abstract, introduction and limitations · Wasil et al. (2024), abstract · the MIRI\ntreaty draft, verification passages · Epoch AI's open-weight comparison, headline finding · Battleday\nand Gershman, arXiv v2 · the discourse map, its pages and data file · H-38, its notes and chapters. The\ngame-theory and Schmachtenberger readings are in the He doc and the Schmachtenberger doc.\n\n---\n\n## 9. A live instance: a party leader asks for the event (added 2026-10-05)\n\n**What arrived.** Nearly three weeks after the Swedish election of 13 September 2026, on the day the\nSpeaker handed the Social Democrats' leader the task of exploring a government for a second time, the\nCentre Party's leader, Elisabeth Thand Ringqvist, invited the Social Democrats, the Left Party and the\nGreens to a joint review of the country's situation (SvD Debatt, 2 October 2026, *\"Vi bjuder in till\nsamtal om läget för Sverige\"*). Representatives of business, the unions, public agencies and other\norganisations would describe the labour market, the security situation and what climate policy requires,\nso that all four parties get the same material. According to TT all three accepted the same day.\n\nThe founder, on a friend's public post that shared the article: *\"Btw this feels quite relevant to\nDeliberus. It could enable this on a CONTINUOUS basis. Right?\"* The friend's post asked why parties do not\nhold such reviews before they start on forming governments, and why formation starts from given blocs and\nnot from the challenges. The post is held with its author's name in `.private/sources/`.\n\n**How it was read.** The article itself could not be opened, because the paper's site refuses this reader.\nWhat follows rests on a news write-up read from its own page (Bulletin, which quotes three of her sentences\nand paraphrases the rest), on two more read through a summarising reader, and on the business federations'\nreply, read in full on its publisher's page. Her sentences in quotation marks are hers as relayed. The\nparaphrases are a reporter's.\n\n**What the proposal says.**\n- **The order.** *\"Innan partier börjar förhandla om politik och fördela poster bör de ha en gemensam bild\n  av var landet står\"*: before parties negotiate policy and hand out posts, they should have a shared\n  picture of where the country stands. And *\"Det är i nuläget och utmaningarna för Sverige som samtalen\n  måste börja. Inte i vem som ska sitta var.\"*\n- **No agreement is expected.** In the write-up's words she does not count on the parties agreeing, since\n  her party and the Left Party differ on enterprise, taxes and the market economy, and that is still a\n  result, because the differences become clear to everyone.\n- **Two precedents.** Finland, where public institutions supply material before and during government\n  negotiations, and the Netherlands, where an independent agency costs the parties' election promises.\n  The Dutch body is the CPB, which by the search results has done this at every national election since\n  1986, on proposals the parties hand over voluntarily. Neither was opened for this note.\n- **What it leaves alone.** *\"Vår linje att V inte ska sitta i regeringen ligger fast.\"* The lock is about\n  who sits in a government, and the review does not touch it.\n\n**Read against § 1's five properties and the bridge.** This is Claude's reading and unruled. It applies\nV2, V5 and V9 to a new case and adds no ledger row.\n\n| Property | The convened review | A standing public map |\n|---|---|---|\n| Shared | A meeting, with each actor's picture arriving as its own briefing or article | One object, every claim of which can be pointed at (built) |\n| Faithful | High by construction: each actor speaks for itself | A map made from published material owes each holder a say on how it renders them (V2, unbuilt) |\n| Located and typed | Nothing in the write-ups says the format separates what is the case from what is wanted | Built: each claim typed by kind, each disagreement located |\n| Common | **What the room gives.** Four leaders at one briefing each know the others heard it (§ 2.4) | A page gives each reader the content and tells no reader who else read it. The joint mark (V5) is unbuilt |\n| Current and conditional | A snapshot taken at a negotiation | Dates and staleness flags are built. A slot for what would change a holder's mind is not (V4) |\n| Bridge to choice | Placed, on purpose, before the bargaining | No surface hands a negotiation what it needs from the map (§ 2.6) |\n\n**What \"continuous\" would then mean.**\n1. **The record can stand between elections, and the room cannot.** A meeting makes a picture common among\n   the people in it and then ends. A standing map is what such a meeting could start from, and what would\n   remain when it closes. They are complements. § 2.4 already says a public page does not by itself\n   produce common knowledge.\n2. **Said once, pointed at from then on.** The federations' reply says they have described the economy's\n   situation *\"i dialoger med samtliga riksdagspartier\"*, and it closes *\"Den upprepar vi gärna för var och\n   en som vill lyssna.\"* One picture, delivered party by party. That repetition is the retrieval term of\n   the cost in `lowering-the-cost.md`, named by an actor who pays it.\n3. **A public map has no guest list.** The review has one: four parties in the room and four outside it.\n   *Contributions are added, never overwritten — nothing needs permission to exist* removes the convener,\n   and with the convener the choice of who briefs. A community's own compartment does have a membership\n   list, which is why this is said of the public map.\n4. **It can be made from what is already printed.** Parties, agencies and federations publish their\n   pictures, so the exposure that makes people guard their reasoning is already spent\n   (`incentives-analysis.md` § 6b). What the map then owes each of them is consent (V2).\n\n**Two specimens this supplies.**\n- **A situation picture that is half demand.** The reply is headed *företagens lägesbild*, the companies'\n  picture of the situation. Of its six numbered points, three describe (a fragile recovery, better\n  conditions for growth, harder competition), one states a growth target, one calls a sixty-point reform\n  agenda the minimum, and one warns against named proposals (higher taxes on work and investment, an end\n  to the sick-pay waiting deduction, shorter working hours) and insists on new nuclear power. Description\n  and demand travel as one object under a descriptive name, which is the case § 2.3's typing exists for.\n- **The legibility floor, stated from outside.** A party leader calling clear differences a result is the\n  project's own floor (`../convergence.md`), reached without its vocabulary.\n\n**Where this instance hurts, on § 4's list and beyond it.**\n- Formation talks are the standard case of an agreement that may need vagueness (§ 4.2).\n- Parties would be legible where it pays (§ 4.7), and every source in the corpus is machine- or\n  operator-extracted, so no instrument here has met anyone with a stake.\n- The review's parties are one side of the house. A picture shared among prospective partners is not a\n  picture shared across the divide, which is what the friend's question asks for.\n- The convener is the party in the middle. Sympathetic readers already hear *map the reasoning* as *the\n  truth is in the middle* (`the-post-that-states-the-problem-from-inside.md` § 6), and an instrument taken\n  up by a centre party would inherit that reading.\n- The lock is an interest conflict. The founding conviction says lowering the cost dissolves conflicts of\n  confusion and only clarifies conflicts of interest.\n\n**A candidate debate cluster, not started.** The invitation, the replies to it and the other side's answer\nare a live, published, Swedish cluster. The August thread on bloc arithmetic is its *before*: its poster\nargued that a Social Democratic government would need both a Left Party saying no government without us\nand a Centre Party saying no government with them in it (`election-2026-bloc-debate-cluster.md` § 1).\nUnder the ruled order of work it waits for the reviews.\n\n**Sources**: Bulletin, *C bjuder in V, S, MP till samtal*, 2 October 2026 · TT through Hela Gotland, TV4\nand Börsvärlden, all 2 October, each through a summarising reader · Svenskt Näringsliv and five member\nfederations, *Till C-ledaren – detta är företagens lägesbild*, SvD Debatt, dated 6 October 2026 on Svensk\nHandel's page and read there on 5 October.\n"}