{"path":"research/how-a-map-is-cut-and-what-it-shows.md","content":"# How a map is cut, and what it shows — sections and the quantities that fill a map\n\n**Date**: 2026-09-21 · **Status**: research and measurement; **every proposal here is unruled** ·\n**Home of the build it concerns**: `specs/map-view/design.md` (§ 4 sections, § 9 three quantities)\n\n**Origin.** The founder, the day the map view went live, on a short reflection about sections being a\nlabel and about three quantities doing three jobs: *\"This stuff is important. I want to reflect more\ndeeply on this. Read and ground more deeply on anything and everything in the corpus that you could\nimagine (and stuff you couldn't imagine) to be relevant.\"* And, while that reading was under way:\n*\"The sections help but maybe they should be auto-clustered/created, I'm not sure if maybe it's ok to\nlet the initial extraction pass LLM/you do it, maybe that's fine for now, or it's not ideal, depends what\nalternatives there are.\"* His two earlier questions are the frame: *\"How are the sections created? How\nare they stored? Do they auto-cluster somehow? Are they represented in the DB?\"* and, on link strength\nagainst what a claim hangs on, *\"aren't these two both important, in their own way? Think deeply about\nthe user engagement gradient etc?\"*\n\n**Read for this, in full or by section, this session**: `engagement-gradient-priors.md` ·\n`curiosity-as-growth-fuel.md` §§ 3, 6, 8, 9 · `game-feel-and-the-first-screen.md` §§ 3–5 ·\n`saturation-and-the-long-tail.md` §§ 3, 3b, 7 · `the-ratification-record-and-the-triage-feed.md` § 3 ·\n`the-residual-error-taxonomy.md` class 7 · `what-human-judgment-is-for.md` act two and the 2026-09-10\nruling · `docs/sketches.md` sketch 2 · `docs/object-model.md` § Argument Bundle · `docs/ux-principles.md`\nP15 and the synthesis-path passage · `docs/depth.md` on Simon · the first pass's instructions\n(`deliberus/extraction/scout.py`) and the relationship pass's (`cross_structure.py`).\n\n---\n\n## Part one — sections\n\n### 1. What a section is today\n\nA string. Every claim carries one `argument_label`, stored on its graph node and in the source's stored\nrecord. It comes from one of three places: the pipeline's first pass, which reads the whole text and\nnames each **argument** it finds; the `group` field of a hand-made file; or, when neither named anything,\na fallback in the browser that reads sections off the shape of the links and stores nothing. There is no\nsection object. Nobody can link to a section, rename one, say a claim is in the wrong one, or say one is\nmissing.\n\n**The first pass knows more than the graph keeps.** What it returns for each argument is a short label,\n**the conclusion quoted word for word from the source**, the type of that conclusion, where the argument\nstarts and ends in the text, its key premises quoted, the contested words in it, whether the conclusion\nis only implied, and who makes the argument (`ArgumentStructure` in `deliberus/extraction/models.py`).\nThe stored record keeps all of it. The graph keeps the label and drops the rest, including which claim\nis the conclusion. The map then has to guess a head for each section from link counts, when the first\npass had already named it.\n\n### 2. Measured on the live public graph, 2026-09-21\n\nCounting claims that are not machine-made question scaffolding. Script: a one-off, not kept; the queries\nare `(:Claim)-[:EXTRACTED_FROM]->(:Source)` grouped by `argument_label`, and the six argument link types\nbetween claims of the same source.\n\n| | |\n|---|---|\n| Sources / sections | 32 / 260 |\n| Sections per source | median 8, most 41; three sources have one |\n| Claims per section | median 6, most 44; four sections hold one claim |\n| Links between two claims of the same source | 672 |\n| … that stay inside one section | **506 (75 %)** |\n| … that cross a section line | **166 (25 %)** |\n| Links between claims of **different** sources | **688** |\n| Claims with no link to anything in their own source | **half, in the median source**; 147 of 166 in the largest |\n\nThree readings.\n\n**Sections behave like the modules Simon described.** `depth.md` holds near-decomposability as the\nreason layered reasoning can be revised a part at a time: links inside a part far outnumber links between\nparts. Three to one, measured, is that signature, on a cut no link-maker was shown: the relationship pass\nsees each claim's text, type, status and attribution, and **not** its section (`cross_structure.py`,\nread 2026-09-21). One caution: claims reach that pass listed argument by argument, so nearness in the list\nmay pull links toward neighbours.\n\n**The quarter that crosses is where the interest is.** `depth.md` already names the boundary of Simon's\nproperty: it fails where a sub-claim belongs to two wholes at once. A link across a section line is that\ncase in the small. A section drawn as a closed frame hides exactly those links, and on the three sources\nwhere crossing links are the majority (a game-theory transcript 73 %, a Swedish essay 66 %, a\ndeath-penalty defence 65 %) a partition misdescribes the text.\n\n**A source is not a closed world.** More links leave a source than stay in it. The source page draws only\nthe ones that stay. On a claim page the far end of a link may sit in another source, and nothing on the box\nsays so (checked: neither `claim_map.py` nor `claimMap.ts` carries a source for a box). The brief asked for\ntravel *\"down/sideways endlessly\"*; the sideways doors are these links, and they are unmarked.\n\n### 3. The alternative he asked about, tested: cluster by the links alone\n\nFor each source with at least three sections and ten linked claims (24 sources), group the claims by link\nstructure only (greedy modularity, `networkx`), and compare with the sections the first pass named, on a\nscale where 0 is unrelated and 1 is identical (normalised mutual information, over linked claims).\n\n| | agreement | links crossing sections |\n|---|---|---|\n| Median of 24 sources | **0.81** | |\n| Best (a philanthropy essay) | 0.97 | 0 % |\n| The hand-made AI-governance map | **0.57** | 35 % |\n| Worst (the game-theory transcript) | 0.22 | 73 % |\n\n1. **Where the text is a run of separate arguments, clustering gives back nearly the cut we already have.**\n   It would add a computation and change little.\n2. **It cannot place half the claims.** A claim with no link in its source belongs to no cluster. The label\n   is the only structure those claims have.\n3. **The hand-made map scores lowest among well-linked sources because its sections are a different kind\n   of thing.** They follow the document's own outline (*The goal*, *Anthropic's approach*, *Longtermism and\n   what matters*), which is a cut by **topic**. The first pass cuts by **argument**: a conclusion and its\n   reasons. Links follow arguments, so they agree with the second and cut across the first.\n\nGrouping by meaning (embeddings) was not run. The corpus already holds why it is the wrong tool for this\njob: it measures shared vocabulary, so it finds topics and not arguments (dogfood run 6); it cannot tell a\nclaim from its denial; and its groups arrive unnamed, so a model would name them anyway.\n\n### 4. Two things are wearing one name\n\n**The bundle** is the founder's object, from his 2013 sketch (`sketches.md` sketch 2, `object-model.md`,\n`ux-principles.md` P15): *a line of reasoning toward one conclusion*. Several bundles hang under one\nconclusion, each with its own check mark, and a row of names steps across them. Argumentation theory has\nthe same object: premises that are needed **together** make one argument, and separate arguments\n**converge** on a conclusion. The graph half-knows it already (the scheme on each link; whether a part is\nrequired or merely adds weight). `game-feel-and-the-first-screen.md` § 4 reached the same place from games:\n*an argument is a vignette*, and slicing by argument is the answer to the hairball.\n\n**The region** is what the map view built: a partition of one source for its overview.\n\nOn a pipeline-made source the two nearly coincide, since the first pass cuts by argument. On a hand-made\nsource they came apart, because the mapper cut by outline. And on a **claim page there are no bundles at\nall**: supporters are one column ranked by strength, so three reasons that only work together and three\nthat each stand alone look the same. That is the 2013 sketch's main idea, and it is the part not built.\n\n### 5. A cut is the frame made visible\n\nThe residual-error taxonomy's seventh class says the first extraction of a domain fixes its vocabulary and\nits axes of decomposition, and that *nothing in the ontology can currently say \"this region's\ndecomposition is one frame among possible frames.\"* A section list is that frame written out in a dozen\nshort names. Two consequences.\n\n**It is the cheapest place for the two acts only a person can do.** *Nothing here is about X* and *this\nwhole picture is cut wrong* need no fluency with the ontology; they need a position and a list short enough\nto take in. The founder ruled on 2026-09-10 that the product invites the reframe, with *\"a standing,\nlow-key invitation on every map\"* as the proposed shape (`what-human-judgment-is-for.md`). The sections\noverview is where that invitation has something to point at. The map carries none yet.\n\n**A cut stored as a string fails the sorting test.** `what-belongs-in-the-ontology.md`: a commitment\nbelongs in the graph if two people can disagree about it and the system keeps working. Today a second\nopinion about a cut has nowhere to go, and re-storing a source overwrites the first. By that test the cut\nis baked in, by whoever ran the first pass.\n\n### 6. Who should cut? Options, with what each costs\n\n| Who cuts | For | Against |\n|---|---|---|\n| **The first pass names arguments** (today) | covers every claim; names carry meaning; comes with a quoted conclusion; agrees with the links (0.81) | one model, one reading, decided once; undisputable; the conclusion is thrown away on the way to the graph |\n| **The source's own outline** (hand-made files) | the author's own cut, so it is attributable to a person | topic, not argument; 35 % of links cross it |\n| **Cluster by links** | needs no judgement; follows the argument | gives back nearly the same cut; places only the linked half; unnamed |\n| **Cluster by meaning** | places every claim | topic, not argument; blind to denial; unnamed; the averaging the corpus distrusts |\n| **A section is a conclusion and what bears on it** (proposal) | the section is a real claim, so it can be linked to, disputed and shared across sources; membership is read off links anyone can add; nothing new in the ontology | needs a fallback for the unlinked half (the label); overlapping sections must be drawn honestly |\n| **Several cuts side by side** (later) | a cut becomes a lens among lenses (`worldview-lenses.md`); *contributions are added, never overwritten* covers it | needs a cut to be an addressable thing, and nobody has asked for it yet |\n\n**Proposal, unruled: keep the first pass as the cutter for now, and change what a section is.** His own\nleaning (*\"maybe that's fine for now\"*) matches the measurement. The cheap steps, in order:\n\n1. **Keep the conclusion.** Store, per section, which claim is its conclusion (the first pass already quotes\n   it). The map stops guessing heads, and a section's name becomes a claim's short title, which is a link.\n2. **Read membership off the links where links exist**, with the label as the fallback for the unlinked\n   half. A claim that bears on two conclusions is drawn once and reachable from both.\n3. **Let a hand-made file say which kind of cut it made** (by argument, or by the document's outline), so\n   an outline cut is shown as the author's table of contents and not mistaken for the argument structure.\n4. **Put the reframe invitation on the sections overview.**\n5. **Store nothing about a disputed cut until someone disputes one.** This is the pattern of\n   `latent-scaffolding-and-the-type-token-split.md`: derivable until touched.\n\nAnd separately, the larger piece: **bundles on the claim page.** Group a claim's supporters into lines of\nreasoning, using what the graph already records, so that *needed together* and *each stands alone* look\ndifferent. Not designed.\n\n---\n\n## Part two — what fills a map\n\n`specs/map-view/design.md` § 9 gave three quantities three jobs: link strength picks the boxes, what a\nclaim hangs on picks the emphasis, where an answer would count places the one offered move. The reading\nconfirms that split and adds five things to it.\n\n### 7. The three jobs are three neighbouring rungs of the engagement gradient\n\n`engagement-gradient-priors.md` lists the tiers and two rules: every tier's act must be reachable **one\nrung** from the tier below, and every tier must get real value **at that tier**.\n\n| Rung | Its act, in that doc's words | What it needs from the map | Quantity |\n|---|---|---|---|\n| Reader (nine in ten) | browse claims, follow a descent | a fair picture of the argument | link strength |\n| Asker | open a critical question, **follow a hinge** | to see where it turns | what the claim hangs on |\n| Answerer (about one in a hundred) | answer a question, add one piece of evidence | one place where input counts | where an answer would count |\n\nSo *both are important, in their own way* has a precise form: they serve neighbouring rungs, and the\none-rung rule says each must be **visible from the rung below without being required of it**. A reader\nsees where the claim turns; following it is one click, which makes them an asker. An asker sees one place\nto push. Nobody is asked to jump two rungs, and the reader who never clicks still leaves with the argument.\n\n### 8. Emphasis makes a gap, so it must arrive with a move — and the reader's move is travel\n\n`curiosity-as-growth-fuel.md` § 3: how visible a gap is made is a design variable, and a noticed gap turns\ninto curiosity only when the person feels able to close it; otherwise into avoidance. A box that glows\n*this is where it turns* is a gap made visible on purpose. The ratified rule applies: no gap displayed\nwithout a move offered. What the rule does **not** require is that the move be a contribution. For a\nreader the move is **go there**. That is what converts *both sides seem to have a point* into *I know which\npremise would settle this*, which `does-descending-change-the-descender.md` § 8 calls located uncertainty,\nand which is orientation and not doubt.\n\n### 9. On a source map the emphasis quantity is a different one\n\nWhat a claim hangs on is defined for one conclusion (`hinge.py` moves one root). A source map has many\nconclusions. The quantity with a meaning there is the one `curiosity-as-growth-fuel.md` § 9.1 proposed and\nnobody built: **fan-out**, how many conclusions would move if this one claim were granted or denied. That\ndoc's reason for it is the strongest known driver of curiosity: one answer that closes many open questions.\nOn the AI-governance map, those would be the few claims most theses rest on.\n\n### 10. A fourth job: keep a place for the tail\n\nThe founder, 2026-08-30, on the finding that machine summaries drop minority and unusual arguments first:\n*\"this feels like a super important finding, also considering that most people … will likely mainly\ninteract with Deliberus through a truth-graph-summary or whatever.\"* The corpus's conclusion was that\nevery summary-shaped surface Deliberus serves kills the tail by the same mechanism\n(`saturation-and-the-long-tail.md` § 3b).\n\n**A map that shows the ten best-connected boxes is a summary-shaped surface, and since 2026-09-21 it is the\nmain door.** Measured on the saved real responses:\n\n| First view of a source map | shown | of which thinly linked (at or under the source's median) | unspoken premises shown |\n|---|---|---|---|\n| AI-governance (116 claims) | 12 | **1** | 1 of 3 |\n| Minimum wages (57 claims) | 12 | 8 (the whole source is thin: median one link) | none exist |\n\nOn the densely linked source the first view is eleven-twelfths head. The unspoken premises are the items\nrun 3F found to be the crux of a real debate, and two of the three are absent until the second or third\nzoom step.\n\nWhat follows, as proposals:\n\n1. **Reserve one place in the first view for the tail**: the unspoken premise, the objection from an\n   unusual direction, the claim few link to. `game-feel` § 3.2 has the same instinct from detective games:\n   show what the solution needs *plus the interesting misdirection*.\n2. **Mark it as what it is.** The DeepMind facilitation study found restating can inflate an outlier as\n   well as bury one. A reserved place that looks like the main objection would do that. So the tail box\n   says *less common* or *nobody said this out loud*, as the unspoken premise already does.\n3. **Give the map its own omissions line.** *N more* says how many were left out and nothing about which\n   kind. The synthesis path already reports what it left out; `ux-principles.md` records why the count must\n   not be shown as a ratio (*a score wearing a disclosure's clothes*). Same rule here.\n4. **What a surface shows shapes what arrives** (`saturation` § 7): showing the common breeds restatement,\n   showing the rare-and-strong breeds new contributions. A head-only map would train its own contributors.\n\n### 11. The brief chose a kind of curiosity, and that is a second reason for picture-first\n\n`curiosity-as-growth-fuel.md` § 8.2 names an unmade decision: every gap-showing device in the product\nrecruits the curiosity that wants an itch scratched, while what sits at the bottom of a descent rewards the\ncuriosity that enjoys an interesting structure, and the first kind tolerates an unresolved ending least.\nThe founder's brief (*\"a visitor who's just surfing around\"*, *\"click through and drill down/sideways\nendlessly\"*) describes the second kind. **The map view is the product's first surface built for wandering\nand not for a to-do list.** Filling the default view with whatever is least examined would turn it back\ninto the first kind. § 9 of the spec argued picture-first from fairness; this is an independent reason.\n\n### 12. The offered move has a bigger supply than one quantity\n\nSpec § 9 sourced the move from `worth_asking` alone. The founder's triage-feed synthesis\n(`the-ratification-record-and-the-triage-feed.md` § 3) is wider: people *\"surfing around Deliberus should be\nsubtly encouraged (but not conscripted/required)\"*, *\"with incredibly clearly articulated yet short\n1-sentence prompts\"*, toward what the machine is unsure of and toward the two errors only people catch, and\nnever toward *\"epistemic glue … we're already quite confident in.\"* **The one move on a map is the triage\nitem for that neighbourhood**: an open question whose answer would count, a link the machine was unsure of,\na classification that confessed it did not fit, a premise nobody supports, or the standing *is something\nmissing from this picture?*. One per view, one sentence, skippable.\n\n### 13. The discipline that rides along\n\nFrom `game-feel` § 5, unchanged: build for flow and measure something else. A map that is pleasant to\nwander will be liked, and liking is the signal the threat model says to distrust. What to watch is whether\npeople point at claims, and the instruments' own readings. And the entrenchment hazard stands as a watched\nhypothesis: making descent more rewarding amplifies whichever way descent moves people.\n\n### 14. What is computed today, honestly\n\nLink strength: yes, and flat on hand-made sources (every link 0.7), where connectedness stands in. What a\nclaim hangs on: only down a claim's parts. Fan-out: not computed. Where an answer would count: yes, per\nopen question. The triage inventory: its parts exist, no feed joins them. On the present corpus almost\nevery strength sits at the middle value, so emphasis and the move would whisper today whatever is built.\nThat is the corpus's state and not a reason to drop the design.\n\n---\n\n## The founder's first response (2026-09-21)\n\n*\"Great work and lots of food for thought. Undecided about 'bundles' - we'll come back to it. Likewise for\nsections.\"* And on what he means by a section: *\"I just noticed that sections started showing up after we\nbuilt the Borderland page/maps and thought it would be a nice UI element that complemented the other stuff,\nbut I'm not sure what it should correspond to ideally, so it's worth philosophizing/strategizing about.\"*\nSo both stay open, by his word, and nothing in Part one is to be built on as though chosen.\n\n## A first outside reaction to two source maps (2026-10-04)\n\nThe founder showed two source maps, cold, in a community chat of people who follow the AI-risk\ndebate, after a member had asked unprompted for a way to keep a running overview of the arguments\nthat need testing or answering. His own words on showing them: *\"De här kartorna är för röriga, det\när eg för tidigt att visa, men nu gör jag det ändå haha\"*, and on the idea: *\"Tanken är att alla\nunika argument splittas upp så finkornigt som går och har en plats man kan återkomma till för att\nfundera på varje påstående osv\"*.\n\nThe reader's response, paraphrased and unnamed. They found the idea interesting while saying they\ngrasped only its outer contour. Their worry was that fine grain makes the tool hard to use in an\neveryday conversation with someone less versed. Their image for what is needed: out birdwatching you\nlook first for shapes and movements in an area, and only after you know the species do you look at\none bird's feathers. And they asked how the material gets in.\n\nHis answers: one should be able to zoom in for detail or out for an overview, *\"men då är risken\nförstås att det blir tillplattat och obalanserat osv så det är trixigt\"*; and it only has to be shown\nworking at small scale, since people follow *\"när man fattar grejen och ser det framför sig istället\nför att behöva föreställa sig det\"*.\n\nWhat this adds here, as Claude's reading. It is one reader, so it counts as a first reaction and\nnot as a measurement. The birdwatching image asks for picture-first (§ 11 here, § 9 of the spec),\nfrom a reader who had been told nothing about it. And his own reply names the cost § 10 measured, in\ntwo words, flattened and unbalanced: a zoomed-out view is a summary by selection and leaves the tail\nbehind. Whether the source pages as they open today meet this reader's need was not tested: they\nwere asked for no second look.\n\n## Candidates this leaves, none ruled\n\nSections: keep the conclusion per section · membership from links with the label as fallback · say which\nkind of cut a hand-made file made · the reframe invitation on the sections overview · mark a box that comes\nfrom another source · bundles on the claim page (the 2013 sketch's core, unbuilt).\n\nWhat fills a map: emphasis always with a way to go there · fan-out as the source map's emphasis · one\nreserved, marked place for the tail · an omissions line for the map · the single move drawn from the\ntriage inventory.\n\n**See also**: `specs/map-view/design.md` §§ 4, 8, 9 · [game-feel-and-the-first-screen.md](game-feel-and-the-first-screen.md) ·\n[engagement-gradient-priors.md](engagement-gradient-priors.md) · [curiosity-as-growth-fuel.md](curiosity-as-growth-fuel.md) ·\n[saturation-and-the-long-tail.md](saturation-and-the-long-tail.md) · [what-human-judgment-is-for.md](what-human-judgment-is-for.md) ·\n[the-residual-error-taxonomy.md](the-residual-error-taxonomy.md) · [../sketches.md](../sketches.md) · [../depth.md](../depth.md)\n"}