{"path":"research/split-review-and-the-founders-intentions.md","content":"# Split review and the founder's intentions\n\n**Date**: 2026-09-28 · **Type**: the record gathered for a founder decision. No design is proposed.\n**Status**: nothing in this document is ruled.\n\n**Occasion.** The founder saw the written-split panel on a claim page for the first time. The panel\nhad not rendered from its build on 2026-09-21 until the fix `10f7df6` earlier that day. His question,\nVERBATIM: *\"This UI... I wanna revisit the reasons for this to even be in the UI, or whether we\ncan/should entrust \"the machine\" with splitting claims reliably, or whatever's already been\nresearched and reasoned in the corpus. Look back far and deep into the docs corpus and consider\nwhat my stated intentions/reflections have been historically.\"*\n\n**Scope.** The panel shows each part the splitter wrote, with **Open · Save wording · Drop**. Its\nsibling, the panel for held proposals, is covered where it bears on the same question.\n**VERBATIM** marks his exact words. **PARAPHRASE** marks a document's restatement of him. Unmarked\ntext is this document's analysis. Live-graph figures come from read-only queries on 2026-09-28\n(FalkorDB `GRAPH.RO_QUERY`, which the server enforces as read-only). The queries are listed at\nthe end.\n\n---\n\n## 0. Answer first\n\nSince 2011 his stated architecture has not changed: *\"leverage automation as far as possible… then\nlet crowdsourcing take over at the point where human intelligence is necessary\"*. On 2026-09-21 he\nstated the test for this panel: whether work is shown to a person depends on how far the machine can\nbe trusted to do it alone. He said he suspects the splitter can be, and switched it to writing\n(*\"Wire and switch first, sitting later\"*). The panel is what remained of his condition for that\ntrust: splits marked, reversible, reviewed in use. That trust is unmeasured. No person has touched\nany of the 37 written parts, because the buttons were invisible until today. The only accuracy\nreadings come from a second model (62–68% genuine parts) and a coverage check that holds 4 of 11\nproposals for invention. The sample is mostly machine-written premises. Several button behaviours\nconflict with later rulings.\n\n---\n\n## 1. Chronology of his intentions\n\n### 1a. 2011–2013: automate as far as possible, people at the residue\n\nSources: the Collaborative Argumentation Analysis group transcript saved 2012-01-24\n(`archive/demeter_2017_dropbox/fredrik/Discussions/`, local archive, not in the public repo), his\nSimplenote notes (`docs/research/simplenote-archive-analysis.md`), and his feature map. Years\nmissing from the transcript are 2012 for January posts.\n\n- **2011-01-12…17, VERBATIM**, in a comment thread under a collaborator's note: *\"We'll make it as\n  hard for people as possible to revert to their old ugly tactics and rhetorics — the UI should\n  help (splitting compound statements and letting other people judge each part of the argument\n  individually, for example), and in general the constraining data structure should make\n  invalid/ungrounded inference painfully obvious, especially with constant feedback/tweaking of\n  statements/voting from other users.\"* (`simplenote-archive-analysis.md` § 3.) Splitting is the\n  point. Who splits is not said. Judging each part is other people's work.\n- **2011-12-05, VERBATIM**: *\"When a normal user's validity votes overlap with the judgments of\n  validators to a certain extent (9/10 maybe) after a certain period of time, they will achieve the\n  validator title, either by validator selection or automatically...\"* This is a trust ladder for\n  people, with promotion by measured agreement. The 2026 trust ladder for machine passes has the\n  same shape.\n- **2011-12-06, VERBATIM**: *\"My general aim is to try to automate (because code has no biases);\n  fail rather early and fall back to crowdsourcing, perhaps combined with simpler automation.\"*\n- **2011-12-14, VERBATIM**, on a competing tool: *\"Rationale also seems very rough in that it\n  doesn't seem to even strive to tidy up the claims to make sure they adhere to the correct\n  structure of a logical argument. It just accepts free-form text \"reasons\". I imagine a\n  crowdsourced quality-optimization process that makes sure every premise is made explicit and\n  neatly & pedagogically spelled out, fallacies sensed and avoided, irrelevant words sorted out,\n  etc.\"* Two minutes later: *\"Yes, at every step in the user's process for input and editing, we'll\n  try to integrate duplication-sensing and error-pruning through natural language processing. After\n  arguments have been sent in, if they are corrected by other users or moderators, the original\n  author will be notified of the type of fallacy he committed or whatever\"*. And: *\"the\n  merging/forking of wordings into different claims is probably going to become a bit of a\n  nightmare to manage, but... we will find the right algorithms.\"* Three ideas appear here first:\n  the machine at every step, correction after submission, and the author told of the correction.\n- **2011-12-29, VERBATIM**, posting a feature map with the minimum viable product marked: *\"The\n  general strategy is to leverage automation as far as possible, then move on to crowdsourcing;\n  designing microtasks for users to do the work that still is beyond computers.\"* The map is his\n  diagram (`archive/demeter_2017_dropbox/fredrik/111~BranstromCAA/!!Roadmap/Featurelist.png`).\n  Some of its box labels, VERBATIM: *Natural language processing* and *Crowdsourcing* both feed\n  *Process free-form text into structured argument(s)*. A second cluster reads *Queues of arguments with\n  votes… to be imported (after input) · to be reviewed (after edit) · to be deleted · to be marked\n  as valid · to be marked as a certain type of fallacy*, beside *User levels with varying powers*.\n  In 2011, human review was a set of queues and ranked users. The 2026 end-state rejects that\n  apparatus (§ 1c).\n- **2011-12-31, VERBATIM**, after he broke a polysyllogism into its constituent syllogisms by hand:\n  *\"Can one always break up polysyllogisms like that? It should be praxis within the app to try to\n  do this whenever possible, I think, to avoid unneccessary complexity...\"* Splitting is the\n  default, \"whenever possible\".\n- **2012-01-01, VERBATIM**: *\"Yes, designing microtasks will be key, just need to design macrotasks\n  first to be able to split them up.\"* Here *split* means splitting tasks, not claims.\n- **2012-01-04, VERBATIM**, three messages. On duplicates: *\"…if the same or a very similar claim is\n  made twice, users are alerted to this either instantly or after review by other users, and are\n  given the option to merge them (or it might happen automatically in some cases)\"*. On quality:\n  *\"Yes we should of course try every method available - natural language processing if possible,\n  and queue up newly-inputted arguments to be reviewed by other users for quality control, then send\n  feedback to the author/editor if errors are found\"*. And the strategy again: *\"The general\n  strategy is to leverage automation as far as possible and maybe do some natural language\n  processing, then let crowdsourcing take over at the point where human intelligence is necessary\n  for the task at hand; we will be designing microtasks for users to do the work that still is\n  beyond computers.\"* The same day he held a position he later gave up: *\"filter everything into\n  the structured format of the syllogism… Free-form comments are probably not a good idea to allow\n  within the app itself. I'm not sure.\"* (`structuring-gradient-lineage.md` § 3.)\n- **2012-01-08, VERBATIM**, sharing a talk on crowdsourcing: *\"I want to do this for every type of\n  task that is required to iteratively refine arguments and recursively hook them up with each\n  other in a giant network of statements, merge duplicates or split them up, reformulate for\n  conciseness, prune out noise and spam, etc.\"* Splitting is a crowd microtask here because the\n  machine's share was then near zero.\n- **About 2012, VERBATIM**, from his notes: *\"Crowdsourcing + automation, language parsing /\n  Microtasks, or just surfing around editing / Queues or feeds of needy args/claims\"* and *\"Focus on\n  trusting the process instead of the people?\"* (`simplenote-archive-analysis.md` §§ 5–6.)\n- **2013-01, VERBATIM fragments quoted by a PARAPHRASE**: his one-page *Consensus software* board\n  says *\"limit content to atomistic statements\"* and adds NLP-assisted extraction of structured\n  arguments from free-form text (`docs/history.md`. The artifact is in the private archive and was\n  not read). A 2013 sheet lists *\"Meta: merges/splits\"* as crowdsourced operations on concepts\n  (`docs/sketches.md`, Sketch 7).\n- **Design precedent, co-built, not only his**: the 2013–14 Meteor prototype stored a rewording as\n  a new *substitute proposition*, voted better or worse. It never overwrote the original\n  (`meteor-prototype-analysis.md` § 5).\n\n### 1b. March to August 2026: decomposition as machine work, people for meaning\n\n- **2026-03-29 (Session 7), VERBATIM**: *\"What if we need even more atomic claims? Could the system\n  include auto-finding premises for each new claim submitted, so that all claims would be the\n  conclusion of a few logical premises? That's the way I (or my left hemisphere at least) always\n  imagined it. But perhaps I need to let go of that rigorous analytical hope. Perhaps that's a\n  pipedream.\"* Then: *\"If the system would auto-decompose incoming new claims in a way that\n  automatically prunes them based on some criteria, like only finding or creating new premises that\n  could one step removed be allowing connections to other existing claims. Then perhaps it would be\n  more computationally feasible and helpful.\"*\n  (`scheme-bounded-decomposition-and-evidence-as-subgraph.md`.) The same document's *Layer 2*\n  (*\"The LLM suggests; the human commits\"*) is Claude's design, not his. The object also differs.\n  Finding premises adds content. Splitting separates content already in the sentence.\n- **2026-04-01 (Session 9), VERBATIM**: *\"recursively decomposed (eventually, not instantly tho)\n  until we aren't muddling properly separable aspects of reality that can and should be judged\n  separately\"* and *\"We must at some point in time enable and process such decomposition without\n  insisting that a muddled, too-general claim is atomic.\"*\n  (`self-similar-decomposition-and-claim-ontology.md`.) He ruled this the separability floor on\n  2026-08-31, VERBATIM *\"Both blessed.\"* (ledger B2, B3).\n- **2026-04-01, PARAPHRASE** of a design direction marked *agreed*: *\"The confirmation step +\n  disambiguation prompts are the only points of interaction between the user's natural reasoning\n  and the graph's formal structure. Everything else — decomposition, classification, scheme\n  detection, CQ generation, embedding, auto-connection — happens invisibly.\"*\n  (`disambiguation-ux-and-voice-first-contribution.md`.) This is the earliest 2026 statement that\n  machine decomposition needs no human touchpoint. The confirmation it keeps is an author's, on the\n  author's own words.\n- **2026-04-03, VERBATIM**, on how far decomposition can go: *\"It is a great question, but... it is\n  an empirical question. It doesn't matter what they believe if we can simply discover the truth\n  within a month's experiment with Deliberus or some similar system. Not a matter of opinion. It is\n  something we will explore together. Let it carry or break.\"*\n  (`session10-empirical-collaboration-and-atomization-frontier.md` § 9.)\n- **2026-04-03, VERBATIM**, on machine merging with human correction: *\"…opt-out for splitting\n  consciously (if this kind of auto-merge ran when above high similarity score and then surfaced\n  these merged claims in a feed) would probably result in less work for humans than having to merge\n  many similar claims manually. Making the HITL labor to split claims consciously is more pragmatic\n  than merging consciously?\"* (`codex-session-20260403.md`.) Here *split* means undoing a merge. The\n  same instinct returned for concept senses on 2026-04-05/06, VERBATIM: *\"…concept senses\n  automatically merging for users later to consciously split them if deemed necessary. I reasoned\n  that this would entail less friction and less wasted/manual time/effort for humans.\"* (ledger P12,\n  open.)\n- **2026-07-06, PARAPHRASE** (Claude's insight from the first dogfood run, later canon): the\n  inversion principle. *\"the system detects the decomposable structure, asks ONE plain-language\n  question at a time …, proposes each split, and the user taps confirm / edit / redirect\"*\n  (`docs/ux-principles.md` P4). Its canon line is *the machine holds the user's hand through the\n  descent*.\n- **2026-08-17, VERBATIM**, the delegation correction: *\"Part of the purpose of Deliberus is to help\n  people by all means mentioned earlier to offload large parts of the work and cost of traversing\n  opacity and stringently following whatever virtues seem most epistemologically virtuous. They\n  SHOULD of COURSE be INVITED to dive in and review and reflect and certify or fix by hand any and\n  all small or large mistakes or inconsistencies or plot holes or missing implicit premises or logic\n  jumps or WHATEVER, BUT Deliberus should ideally over time be made capable of offloading as much of\n  the idealized processing and structuring and scaffolding as possible.\"*\n  (`conviction-and-critique.md`.) The same day, VERBATIM: *\"since far back I've imagined that\n  background agents/daemons could be run on loops … that function kind of like enzymes or T-cells or\n  roaming philosophers mending and filling up the graph… Has to be approached carefully and with\n  deep thought of course.\"* (`graph-daemons-design-space.md`, whose tier table then put\n  decomposition in the propose-only tier.)\n- **2026-08-25, VERBATIM**: *\"I want Deliberus to automatically recursively decompose all claims in\n  each mapping-break…\"* and *\"…automatic recursive decomposition and deepening exploration and\n  relevant research/googling should fire/engage … and I'm leaning towards that this should also\n  occur (to some degree/cap at least) in normal use.\"* (`kartpaus-auto-enrichment-design.md`.) The\n  direction is ruled. The design puts machine decomposition at draft, *\"invisible on the default\n  map\"*, under the rule *\"do eagerly, show on contest\"*.\n- **2026-08-26, VERBATIM fragment**, ruling out a write-time *\"did you mean…?\"* for sameness:\n  *\"merely suggesting is putting too much freedom and work/mental friction on the user.\"* The ruling\n  as recorded, PARAPHRASE: *\"anyone types anything; careful auto-merge afterward; anyone (author or\n  passerby) can split when truly warranted\"*\n  (`soft-canonical-clustering-and-reversible-merge-semantics.md` § Addendum Aug 26). The same loop\n  named four conditions that make undo real: originals never destroyed, the change visible where\n  the claim is read, splitting as an accountable act *\"with attribution and reason\"*, and the\n  author notified.\n- **2026-08-27, VERBATIM**: *\"I don't really know if it makes sense to have human users\n  vote/ratify… Let's not install meaningless buttons for users to press and judgments we hold our\n  breath for them to make that don't actually change anything… when they do engage it should\n  MATTER, it should be filled with MEANING, so they can FEEL their humanness rippling across the\n  graph.\"* (`what-human-judgment-is-for.md`.)\n- **2026-08-30, VERBATIM**, the triage feed: *\"a human surfing around Deliberus should be subtly\n  encouraged (but not conscripted/required) to evaluate different kinds of discernments made by\n  LLMs in whatever circumstances we can't be 105% confident in the reliability of the targeted\n  prompting of the LLM-step... but ideally humans would PRIMARILY be nudged (perhaps a feed could be\n  created) to sift through suspected candidates for human-reserved classes of error, … instead of\n  having them re-check judgments and epistemic glue/inferences/connections we're already quite\n  confident in.\"* (`the-ratification-record-and-the-triage-feed.md` § 3.) The same day he stated\n  the harness conviction, which begins *\"If I outsource food shopping and meal planning for a\n  conference of 1000 ppl to a master chef with a calculator and a spreadsheet…\"*\n  (`the-harness-conviction.md`).\n\n### 1c. 2026-08-31 to 2026-09-19: no ratify act, housekeeping, relevance\n\n- **2026-08-31, VERBATIM**: *\"should there even be a voting/ratification beyond the stance\n  (agree/disagree) vote on a claim. I don't think the seeded-probe thing sounds useful and elegant\n  and worthwhile, do you? ... OTOH: if we don't have discernment-ratification, maybe it'll\n  unbalance or fuck up the ontology somehow?\"* He then blessed the end-state, VERBATIM: *\"Anyhow yes\n  I agree, blessing this whole proposal you gave now, go ahead.\"* That end-state, PARAPHRASE: there\n  is no ratify act. The graph holds contributions and telemetry. Assent to a machine discernment is\n  telemetry. Dissent must arrive as a reasoned reclassification. Review during the operator phase,\n  with the founder and friends, may stay authoritative. Untouched machine structure is the permanent\n  honest default under provenance display. (`the-ratification-record-and-the-triage-feed.md`\n  §§ 3b, 4b.)\n- **2026-09-03, VERBATIM question** (the answer is unruled): *\"So what's a conservative approach\n  that still uses humans to fill up the tail faithfully but has a healthy measure of distrust\n  against both LLMs and humans?\"* The answer, *cheap to add, slow to matter*, has a second model\n  family and positioned humans check machine decompositions (`cheap-to-add-slow-to-matter.md` § 1).\n- **2026-09-13, VERBATIM**: *\"Any regular housekeeping should be done that could be done with high\n  confidence in safety/accuracy and could spare human users from having to later clean up and\n  organize the graph manually etc, to instead free up their mental cycles for the uniquely\n  human-needing tasks.\"* On a machine link that fails a confident re-check: *\"Yes, replace by\n  itself, keep the old link visible.\"* (`graph-daemons-design-space.md` § Addendum. Both are canon.)\n- **2026-09-19, VERBATIM**, on the relevance of parts: *\"ideally the RELEVANCE (which I guess all\n  this is about) is judged provisionally by the machine/LLM-passes or whatever and then humans can\n  challenge it but I'm wary of saboteurs/spammers adding bogus parts that aren't actually\n  relevant\"* (`relevance-and-the-two-factors.md` § 1).\n\n### 1d. 2026-09-21: the day the panel was built\n\nSource: the session transcript that ran 2026-09-20 to 09-27 (id `dc4eae5b`, read through\n`ccdistill`), cross-checked against `the-jhu-decomposition-line.md` § 2026-09-21,\n`earning-the-right-to-write.md` and `session34-open-loops-and-unweighed-answers.md`. In order:\n\n1. **VERBATIM**, on a six-part claim: *\"This screenshotted claim should perhaps be possible to\n   automagically decompose as a phase of the extraction process, right? … Is the whole \"Deepen this\n   claim\" box unnecessary to expose to humans? Or could it at least auto-suggest how to\n   separate/decompose the claim in question, by running a separate carefully-prompted LLM pass when\n   a human user clicks to split it into smaller claims?\"*\n2. **VERBATIM**, in reply to Claude's proposal to store parts *\"as unconfirmed. The box then shows\n   them with keep, reword or drop\"*: *\"Make this happen! Maybe we can both pre-run it at extraction\n   time like you propose. But also make it possible to run on click where no such pre-run has taken\n   place, it should auto-sense this. And you can do a run yourself on the existing\n   suspected-to-be-clumped-up claims in the DB, regardless of what extraction they're attached\n   to\"*. The keep/reword/drop panel, for proposals, was in the text he approved.\n3. **VERBATIM**: *\"Beyond 3-4 rows we probably always are in need of splitting claims anyway.\"*\n4. **VERBATIM, the criterion**: *\"To me, whether anything is shown as potential work to be done by a\n   human user in the UI depends on: how confident can we be in the machine/system instead doing\n   that work on its own, so it doesn't even need a UI? What has been said about this in the past in\n   the corpus?\"*\n5. **VERBATIM**: *\"I'm curious about maybe redrawing this line... to (in general, as often as is\n   wise) spare users from having to encounter chaos/clutter [meaning maps/claims that are unsplit\n   when they could confidently, accurately be split by machine/daemon/pre-runs, …] and instead to\n   the extent possible without making glaring mistakes tidying up regularly where it makes the most\n   difference to reach and show examples of more clarity and illustrate the power of\n   separability/decomposition/addressability etc.\"* (Answered in `the-tidying-line.md`.)\n6. **VERBATIM, the discernment**: *\"…the warnings are actually most relevant to cases where an\n   AI/machine pre-digests/summarizes/makes a judgment call that goes deeper/has more\n   inferential/analytical depth/scope than what we can be confident results in an \"objectively\"\n   good, unbiased job … I want to start by testing the assumption that such a microtask of splitting\n   claims here can be done accurately and in a risk-minimizing way, I think it can if the prompting\n   and model is good enough etc. We can always mark machine-split claims in some subtle way in order\n   to know their provenance in order to review the accuracy continually/in the near future after\n   we've let the system do its thing for a little while. The cost of something like this would be\n   rather low and the benefit would be high in my mind, IFF it works reliably, which I suspect it\n   will, otherwise we will reconsider. This general approach I also want applied to other\n   machine-vs-human tasks and ratification-dependent processes/workflows etc.\"*\n7. **VERBATIM**: *\"yes I'm pretty sure I agree that \"review it later\" is worse than simply weaving in\n   the reviewing naturally in the system for users to ratify - is this what you're trying to\n   say?\"*\n8. **VERBATIM**: *\"Agreed on everything here, except ofc you need to consider the floor of\n   separability/atomicity/decomposability and not just blindly maximize for \"more places to\n   disagree\", but maybe this is nitpicking, not sure. Important context should at least be preserved\n   in each splitting.\"* Then, VERBATIM as recorded in `session34-open-loops-and-unweighed-answers.md`:\n   *\"Approved both candidates\"*. The addressability test and *a cut through a concept is an\n   argument, not a split* entered canon.\n9. **VERBATIM**: *\"If the newly-split-out part is rightfully split-out, then if it's weak, it's weak,\n   if it's strong, it's strong... It should simply be judged on its own merits/evidence/subgraph,\n   right? Even when it's machine-split/minted I'm open to letting energy flow through it.\"*\n10. **VERBATIM, who reviews**: *\"In practice, it will most likely be me and other\n    semi-deeply-invested nerds/experts/users I recruit for this purpose that would go through the\n    claims graph DB broadly and search for unsplit/decomposable claims and try to systematically\n    ensure that the system works smoothly without constant human ratification/revision/fixing...\"*\n11. **VERBATIM**: *\"Yes it should. We should expect good splits, and I have some trouble imagining\n    wrong splits, what would that even look like? I guess in some cases they could lose some\n    words/context that is important to the claim being specific enough? What else could go wrong? I\n    have rather high trust that this is machine/LLM-doable...\"* Also, VERBATIM as quoted in\n    `earning-the-right-to-write.md` § 3a: *\"I think logic would reign here. I don't expect different\n    cuts to be a matter of taste, this is probably more a science than an art?\"* And: *\"Do you have\n    reason to suspect the looks-bundled flagging is wrong?\"*\n12. **VERBATIM**, on a cut through a concept: *\"I would call that act backing the main \"policy is\n    cruel\" claim with subclaims rather\"*.\n13. **VERBATIM**, in reply to Claude's line *\"The machine may propose parts and may not write\n    them\"*: *\"I thought we (in last session) ended up in a place where the discernment here was\n    different? That maybe it makes sense to trust the splitting and we can as you mention always mark\n    them subtly to be able to reverse if need be, although I suspect we ultimately can trust the\n    splitting-machinery?\"* Then: *\"Wire and switch first, sitting later.\"* The operator sweep then\n    wrote 7 proposals and 44 parts on his word, VERBATIM as recorded in `TODO.md` (not found in the\n    distilled transcript): *\"Oh, sure, do it\"*.\n\n### 1e. 2026-09-28\n\nHis question at the top of this document. The same day he ruled two items that bear on the panel,\nin his words: retiring the `confirmed` field (ledger Q32, *\"Yes.\"*), and relevance as a second\nfactor on a link, argued and never voted (ledger Q33, *\"About the decision: your proposal is\nblessed, make it so.\"*).\n\n### 1f. What holds across fifteen years, and what reversed\n\n**What holds.**\n\n- The machine goes first and people take the residue (2011, 2012, 2026-08-17).\n- Splitting is the default direction: *\"whenever possible\"* (2011) and *\"until we aren't muddling\n  properly separable aspects of reality\"* (2026).\n- Correction comes after the fact, not approval before it. In 2011, corrected arguments notify the\n  author. On 2026-04-03 he chose opt-out splitting of machine merges. On 2026-08-26 he ruled\n  auto-merge first and splitting when warranted.\n- His stance is empirical: *\"Let it carry or break\"* (April) and *\"IFF it works reliably … otherwise\n  we will reconsider\"* (September).\n- Whoever judges earns the right by measured agreement: validators in 2011, and machine passes in\n  the 2026 trust ladder (proposed, unruled).\n\n**What reversed.**\n\n- In 2011, review was queues, validators and a *\"marked as valid\"* state. On 2026-08-31 he blessed\n  *\"there is no ratify act\"*.\n- In 2012 he wanted structured input. In 2026 he wants the structuring gradient (P20).\n- In March 2026 the design was *\"The LLM suggests; the human commits\"* (Claude's). On 2026-09-21 it\n  became *\"the machine writes what it found, and a person corrects\"*.\n\n**What he has not said in 2026.** He has not said that each machine split needs a person's\nconfirmation before it counts. He approved a keep/reword/drop design for proposals on the morning\nof 2026-09-21 and moved past it that evening. Two corpus principles ask for confirmation, and both\nconcern an author confirming a rendering of the author's own words: anti-pattern 4 (*\"The\nconfirmation step is sacred. The user always has the last word on what they meant.\"*,\n`ux-principles.md`) and the April disambiguation direction.\n\n---\n\n## 2. Arguments for and against the panel\n\n### 2a. For: the reasons recorded when it was built\n\n1. **It is the correction record that his trust was conditioned on.** His terms were to mark the\n   splits, *\"reverse if need be\"*, and review *\"continually\"*, woven into use rather than done\n   later. Each touch writes `human_action`, the account and the time on the edge. The correction\n   rate and the time to first objection are then *\"one query away\"* (`earning-the-right-to-write.md`\n   §§ 3, 4a, and commit `976d87e`). Without the panel, the pass at rung two has no feedback path except\n   an offline sitting.\n2. **An undo must be visible to be real.** The merge loop's first condition: a change must be\n   *\"visible where the claim is read, since an invisible merge is undoable in theory and permanent\n   in practice\"* (`soft-canonical-clustering-and-reversible-merge-semantics.md` § Addendum Aug 26).\n   The panel is the only surface that shows a split as a split and can undo it.\n3. **The pass makes mistakes that a reader can see.** The recorded shapes are loss (a dropped\n   attribution), invention (4 of 11 proposals), drift into commentary, and cuts through a concept\n   (`earning-the-right-to-write.md` §§ 2b, 4a · `core-periphery-and-the-concept-layer.md` § 5a-ii).\n   The *Add it as a part* button answers loss with a move.\n4. **It replaced an essay with taps.** Before 2026-09-21, the *Deepen* box opened an empty text\n   field, *\"the essay the inversion principle says never to ask for\"*\n   (`the-jhu-decomposition-line.md` § 2026-09-21).\n5. **Held proposals need a person by the pass's own logic.** Parts that say what the parent did not\n   are authorship, not separation (`deliberus/proposed_split.py` docstring).\n6. **Human labels could calibrate a cheaper decider.** One document names keep/reword/drop labels as\n   the calibration set for moving decomposition judgements to a cheaper model\n   (`jev-and-the-judgement-layer.md`, row *Decomposition kind*).\n7. **An author's objection needs a place to land.** A written part carries the parent author's name.\n   *\"I do not accept that as a statement of my position\"* is a contribution only that author can make\n   (`what-human-judgment-is-for.md` § 3, act three · `the-ratification-record-and-the-triage-feed.md`\n   § 4b). The panel does not distinguish the author today.\n\n### 2b. Against: what the corpus says\n\n1. **His own criterion points away from a UI.** *\"…so it doesn't even need a UI?\"* (2026-09-21),\n   together with his stated trust: *\"I suspect we ultimately can trust the splitting-machinery\"* and\n   *\"rather high trust that this is machine/LLM-doable\"*. The ladder's third rung makes review\n   sampled. On the build day Claude described that rung as what *\"would eventually let a split land\n   without a panel at all\"* (Claude's words, not his).\n2. **The April direction made decomposition invisible.** Human contact was to happen only at the\n   author's confirmation and at disambiguation (`disambiguation-ux-and-voice-first-contribution.md`).\n3. **No ratify act, and bare judgments are telemetry** (2026-08-31). A keep, or *Save wording* with\n   the text unchanged, is a ratification click. *Drop* is a bare dissent with full effect.\n4. **\"Meaningless buttons\"** (2026-08-27). *Save wording* on unchanged text changes no strength,\n   because a split part already caps like a person's part. It only flips a field that Q32 retires.\n5. **Human attention belongs elsewhere.** The triage feed sends people to what the instruments are\n   unsure about, *\"instead of having them re-check judgments … we're already quite confident in\"*\n   (2026-08-30). The two human-reserved error classes are coherent absence and frame lock-in\n   (`what-human-judgment-is-for.md` § 1). Ledger R13 (open) treats automation and routing as one\n   design.\n6. **The review it produces is thin and biased by construction.** Decomposers are *\"a fraction of\n   1%\"* of visitors (`engagement-gradient-priors.md` § 2). A reviewer shown the machine's cut is *\"a\n   second sample of the machine's framing\"* (`what-human-judgment-is-for.md` § 2). Rubber-stamping\n   makes the correction rate look superb (`earning-the-right-to-write.md` § 3), and *\"a tidy graph\n   looks examined\"* (ledger R11, open). The corpus's alternative is a calibration sample, with\n   seeded wrongs kept offline (`machine-work-as-a-standing-property.md` § 3d), not review of every\n   item.\n7. **One click runs against two rulings of 2026-09-28.** Relevance is argued, never voted (Q33). A\n   person's refusal becomes a visible retirement with its reason (Q32). *Drop* sets a part's effect\n   to zero with one click and no reason.\n8. **Signed-out readers never see it.** So it does not serve the *show, don't tell* aim that he gave\n   the tidying (`the-tidying-line.md` § 4). `TODO.md` records that question as raised three times\n   and never ruled.\n\n### 2c. Not decided anywhere\n\n- The trust ladder's numbers (agreement rate, minimum touches) and the sitting that *\"later\"*\n  refers to (`earning-the-right-to-write.md` § 4 · `TODO.md`).\n- The three questions `the-tidying-line.md` § 6 left open: whether machine-tidied structure is\n  visible to a signed-out visitor, whether *\"shown as a question\"* is the right form, and whether a\n  tidy-up runs continuously or only on ingest.\n- Whether machine-cut parts need their own claim kind (`TODO.md`, marked Claude's call, unruled).\n- Whether a failed reconstruction re-files a split as support\n  (`core-periphery-and-the-concept-layer.md` § 5a-ii, proposed).\n- The standing of untouched machine structure (`conviction-and-critique.md`, question 1). The\n  2026-08-31 end-state answers it as a permanent honest default under provenance display. The\n  display rule is unbuilt.\n- The four held proposals (`TODO.md`: *\"yours to rule on\"*).\n\n---\n\n## 3. Measured reliability of machine splitting\n\n| What | Date | n | Result | Source |\n|---|---|---|---|---|\n| Looks-bundled flag, reach | 2026-09-21 | 1,824 unsplit claims | 1,530 flagged (84%). 167 flagged on a single conjunction only, 246 on length only | `earning-the-right-to-write.md` § 3a |\n| Flag, false positives | 2026-09-21 | 11 read, longest first | 0 came back *already atomic* | same |\n| First proposals | 2026-09-21 | 11 claims | 64 parts, 9 flagged as a value judgement fused into the claim. The *covers* flag true on 10 (invention is reported separately). One dropped attribution (a study's university) caught | `the-jhu-decomposition-line.md` · `earning…` § 6 (*\"n = 11, which licenses nothing\"*) |\n| Failure shapes in the first four | 2026-09-21 | 4 | Loss 1, invention 1 (coverage *clean*, two inventions), drift into commentary watched | `earning…` § 2b (read by Claude, not a person) |\n| Held for invention | 2026-09-21 | 11 proposals | 4 held (36%), 7 written | `earning…` § 4a |\n| Parts judged genuine by a second model, run 1 | 2026-09-21/22 | 44 split edges | 30 genuine parts (68%), 4 miscast as evidence, 10 fit nothing. Mean relevance 0.42. Other writers: automatic linker 0 of 84, people 9 of 39, extraction 6 of 30 | `session35-the-verification-night.md` § 4 |\n| Same judge, run 2 labels on the live edges | labels 2026-09-22, read 2026-09-28 | 37 | 23 genuine (17 necessary, 6 corroborative: 62%), 7 miscast as evidence, 7 fit nothing. Mean relevance 0.47. The model-free second reader agrees on 10, disagrees on 5, is silent on 22 | read-only query |\n| Human touches | 2026-09-28 | 37 edges | 0 kept, 0 reworded, 0 dropped, 0 superseded. The panel did not render from 2026-09-21 to 2026-09-28 | read-only query · commit `10f7df6` |\n| Downstream use of parts | 2026-09-28 | 37 parts | 0 incoming supports or attacks, 0 votes | read-only query |\n| Coverage on written edges | 2026-09-28 | 37 | 31 clean, 6 carry a *dropped* verdict (one parent) | read-only query |\n| Proposals now stored | 2026-09-28 | 10 | 6 written (37 parts), 4 held, all 4 for invention. 57 parts in total, 7 flagged as fused value judgements | read-only query |\n| Theory-guided LLM decomposition, literature | 2024 | benchmark | 93–99% subclaim support against 80–94% for a syntactic parser. *\"far more subclaims … while maintaining high coherence\"* | `reading-the-text-itself.md` § 3 (measured on biographies, not opinions) |\n| Decompose-then-verify, literature | 2025–26 | benchmarks | Degrades verification without per-part evidence alignment. Gameable by adding obvious or repetitive parts (CORE) | `decomposition-axes.md` § 2 · `the-jhu-decomposition-line.md` (both about automated verification, not addressability) |\n| Human agreement on grain | literature | — | 56% fine against 91% coarse for word senses | `earning…` § 3a (transfers to cuts through a concept only) |\n| Sensitivity of strength to filing | 2026-08-22 | 87 decomposed claims, four filings | 91% of verdicts survived every filing | `docs/depth.md` |\n\n**How to read the table.**\n\n- **Every accuracy figure is a model reading a model's output.** No split has a human-labelled\n  accuracy. The census label *miscast as evidence* reliably means only *not a part*\n  (`the-parts-review.md` § 4b).\n- **The sample is narrow.** The 11 claims were taken longest first. Seven are implicit premises,\n  which no source stated in words. Three are long evidence claims that a signed-in account\n  contributed. One is a hand-mapped claim, whose proposal and seven parts are no longer in the live\n  graph (inferred: lost when that map was stored again on 2026-09-22). No ordinary claim extracted\n  from a source has been split.\n- **Invention is the failure that decides the most.** The pass refuses to write it. It writes a loss\n  with a warning beside it.\n- **The strength effect today is nil (inferred, not computed).** No part has any support, attack or\n  vote, so no part can move its parent's badge yet. When parts gain evidence, the damage channel\n  is narrow. A wrong part caps its whole at the presumed weight. An invention verdict removes the\n  cap. Splitting cannot raise a parent (the padding ruling). Measured illustration: children at\n  0.30 / 0.80 / 0.80 give the parent 0.498 at the presumed weight and 0.300 when the part is argued\n  necessary (`earning-the-right-to-write.md` § 2a).\n\n---\n\n## 4. Tension map\n\n### 4a. What each control does (code as of 2026-09-28)\n\nFiles: `web/src/routes/claim/[id]/+page.svelte`, `web/src/lib/split.ts`, `deliberus/api.py`\n(`POST /claims/{id}/split-parts/{child}`), `deliberus/proposed_split.py` (`record_action`,\n`commit`).\n\n- **Who sees and who acts.** The panel sits inside the claim page's *Deepen* card, which is hidden\n  from signed-out readers. Every action needs a sign-in. Any signed-in account can act, 20 times a\n  minute.\n- **Open** is a link to the part's own page. It writes nothing.\n- **Save wording** sends the action `reworded` with the text in the box. The server overwrites the\n  part's `text` and `short_title` in place. On the edge it writes `human_action='reworded'`, the time,\n  the account and `confirmed=true`. Unchanged text is still recorded as a reword. The stored\n  embedding keeps the old text, because the embedding debt counts only missing embeddings. The part\n  stays `claim_kind='extracted'` and keeps the parent author's name.\n- **Drop** sends the action `dropped`. The edge gets `human_action='dropped'`, `confirmed=false`, and\n  `superseded_by` set to a timestamp. No reason is asked for or stored. The part claim stays in the\n  graph, unlinked from its parent, still in the parent author's name, at `candidate` maturity. The\n  parent's strength no longer reads the edge. The panel keeps the row, faded (*\"dropped — the link\n  stays visible\"*). The panel has no undo control. Any signed-in account can call\n  `POST /edges/supersede` with `restore=true`. That call clears the supersession but not\n  `confirmed=false`, so a restored part would still not count toward its parent (read from\n  `restore_edge` and `REFUSED_EXCLUDED`, not run).\n- **Keep** exists in the API as `kept` and writes `confirmed=true`. The page has no Keep button.\n- **Cutting differently.** The written panel has no control to cut the claim another way or to add a\n  part. The one exception is *Add it as a part*, which appears only when the coverage call reported\n  a drop (one of the six written parents).\n- **Silence** writes nothing. The code comment says *\"silence is not agreement\"*.\n- **Held proposals** (4 claims) show a checkbox and an editable text per part, with *Keep N parts*\n  and *None of these — I'll write it*. Keeping calls `POST /claims/{id}/decompose`. That endpoint\n  creates the parts as `claim_kind='human'` in the keeping account's name, on edges with provenance\n  `human`, trust `high` and `confirmed=true`. The machine's wording becomes a person's claim, although\n  these proposals are held because the coverage call found invented content.\n\n### 4b. Controls against rulings\n\n*Serves* and *conflicts* are read against the ruling's text. *Tension* marks a pull that is not a\ncontradiction. *n/a* means the ruling does not bear on that control.\n\n| Ruling (date, status, home) | Open | Save wording | Drop | Why |\n|---|---|---|---|---|\n| **Bare judgments are telemetry** (agreed 2026-08-31, `the-ratification-record-and-the-triage-feed.md` § 3b) | n/a | conflicts | conflicts at scale | Unchanged *Save wording* is an assent that sets `confirmed=true`, which today grants immunity from housekeeping. *Drop* is a dissent with full effect and no argument. The ruling's operator-phase carve-out covers the founder and friends. The code admits any account. |\n| **No ratify act** (blessed 2026-08-31, same § 4b) | n/a | partly conflicts | n/a | The page has no Keep button, which is consistent. *Save wording* doubles as one. |\n| **Confident housekeeping is the machine's job** (2026-09-13, canon) | n/a | tension | tension | If the founder trusts the pass, correcting it is housekeeping done by a person. Ledger R12 (open) reads \"high confidence\" as measured calibration, and the panel is one way to get that measurement. |\n| **A machine link failing a re-check is replaced, old link visible** (2026-09-13) | n/a | n/a | serves | *Drop* supersedes and keeps the row visible. It does not use `retire_edge`, which needs a reason. |\n| **Wire and switch first, sitting later** (2026-09-21) | n/a | serves | serves | Both write the correction record. It has zero rows. The sitting has not been held. |\n| **A split part caps like a person's** (2026-09-21, *\"We should expect good splits\"*) | n/a | no effect | serves | The cap is the same whether a part is untouched or saved. *Drop* is the only human remedy for a wrong cap. |\n| **Retire `confirmed`** (Q32, ruled 2026-09-28) | n/a | conflicts | conflicts | Both write the field. Q32 moves a refusal to *\"a visible retirement with the reason\"*. Under Q32's build plan, housekeeping immunity goes to links a person wrote (`provenance = human`). A person-reworded part keeps provenance `split`, so it would become eligible for machine retirement. |\n| **Trust ladder, rung two** (proposal unruled; splitter placed there 2026-09-21) | partly | serves | serves | Rung two says *\"every surface shows the mark beside a one-click correction\"*. The part's own page, reached by *Open*, carries no mark. The map and both relation lists show none either (§ 4c). |\n| **An act's value is not its rung** (canon, re-pointed 2026-09-03, provisional) | n/a | tension | tension | Correcting splits is a decomposer-tier act, offered to every signed-in reader. It is not among the acts the canon line names as the most valuable. |\n| **The machine holds the user's hand through the descent** (canon) | serves | serves | serves | Taps, not essays. The inversion principle ends in *confirm / edit / redirect*. The written panel has no *redirect*: it offers no way to cut the claim differently. |\n| **P20, the structuring gradient** (founder-stated 2026-08-14) | n/a | tension | n/a | The parts grow from somebody else's words, mostly a machine's. The parent's author has no separate route. |\n| **Contributions are added, never overwritten** (ratified 2026-08-27) | n/a | conflicts | serves | *Save wording* overwrites the part's text in place. The machine's first wording survives only in the parent's stored proposal. A second reword leaves no trace of the first. The 2013 prototype stored a rewording as a new item. |\n| **A cut through a concept is an argument, not a split** (approved 2026-09-21) | n/a | n/a | gap | No control re-files a part as support. *Drop* discards the relation instead. |\n| **Machine work is judged by addressability** (approved 2026-09-21) | serves | n/a | n/a | The splitter passes this gate. The gate does not require a review panel. |\n| **Relevance is argued, never voted** (Q33, ruled 2026-09-28) | n/a | n/a | conflicts | *Drop* sets a part's effect to zero with one click. The endpoint for arguing a part's requirement has no screen. |\n| **Undo is real only if visible where read, recorded with a reason, and the author is told** (merge loop, 2026-08-26) | partly | partly | partly | The record keeps who and when. It keeps no reason. No author is told. The panel has no undo control, and the API restore path leaves the refusal flag set. The mark is not visible where the part is read. |\n| **Every claim is a link** (standing UX directive) | serves | n/a | n/a | The part's text sits in an editable box, so *Open* is the only link to the part's page. |\n\n**Two rulings bear on the panel as a whole, not on a button.**\n\n- **Capture immediately, publish deliberately** (canon). The splitter writes parts at `candidate`\n  maturity, which default public surfaces show. The ruled direction of 2026-08-25 put machine\n  decomposition at draft, *\"invisible on the default map\"*, until someone contests the claim. So the\n  panel works after publication, not before it. The rulings of 2026-09-21 did not say which of the\n  two applies to splits.\n- **No gap displayed without a move offered** (canon). The panel serves it. *Add it as a part* gives\n  dropped content a move, and the panel replaced a *Deepen* box that offered only an empty field.\n\n### 4c. Discrepancies found while reading\n\nThese are facts, not proposals.\n\n1. The mark that rung two promises appears in one place: the panel's own sentence, *\"These N parts\n   are in the map, cut out of the claim above\"*, on the parent's page, which signed-out readers do\n   not see. The map does not read provenance. The *Decomposes into* list on the parent's page shows\n   no provenance, and neither does the *Part of a larger claim* section on the part's own page. Both\n   print the stored constant strength of 80%.\n2. 17 of the 37 written parts were cut from implicit premises. The map draws those parents dashed,\n   *\"nobody said this out loud\"*. The parts are stored as `extracted`, so the map draws them solid.\n3. One machine output gets two attribution rules. A written part carries the parent's author (*\"the\n   content was theirs and only the cutting is ours\"*, `commit()`). A part kept from a held proposal\n   carries the account that kept it, as a `human` claim.\n4. *Drop* stores a timestamp marker and no reason. `retire_edge`, which Q32 names as the home of a\n   person's refusal, stores `retired:<who>` and requires a reason.\n5. The page cannot record a keep. *Save wording* on unchanged text records a reword.\n6. After a *Drop*, the *Decomposes into* list and the dropped part's *Part of a larger claim*\n   section still show the link as live, at 80%. Only the panel fades it and only the map leaves it\n   out (read from the code, not run).\n7. A *Drop* is not fully reversible by the existing path. `restore_edge` clears the supersession\n   and leaves `confirmed=false`, which keeps the part out of its parent's strength and out of the\n   cap (read from the code, not run). Reversibility was his condition for trusting the splitter.\n8. A reworded part keeps the embedding of its old text.\n9. Once the splitter has written a claim's parts, a reader cannot cut that claim differently from the\n   page. Writing parts by hand opens only through *Add it as a part*, which needs a reported drop.\n10. Written parts get no scheme and no critical questions. That enrichment was skipped on purpose.\n    The run over the written links was asked about on 2026-09-21 and not answered (`TODO.md`).\n11. Ledger row B7 still reads OPEN, with the pre-switch proposal *\"propose at extraction, review by\n    tapping\"*. The rulings *\"Make this happen!\"* and *\"Wire and switch first\"* have overtaken it.\n12. The correction record has no rows. The reason is that the panel did not render for a week, not\n    that nobody objected. So time to first objection has not started for any part.\n\n---\n\n## 5. Open questions for him\n\nEach question names the evidence it turns on. None is answered here.\n\n1. **Does the panel stay?** By his test of 2026-09-21 (*\"…so it doesn't even need a UI?\"*), does the\n   splitter's record justify a per-claim review surface, or is correcting it housekeeping that a\n   person should not be asked to do? *Turns on*: § 3. No human-labelled accuracy exists. A second\n   model reads 62–68% of written parts as genuine. 4 of 11 live proposals are held for invention.\n   The sample is mostly machine-written premises. His stated trust is high, and the ladder's numbers\n   are unset. *(Added 2026-09-29: read part by part, 6 of 37 written parts do harm, two of them\n   clearly, by stating as fact a clause their claim denies or concedes; the second model flagged all\n   six and the coverage check none, so its 62% is roughly the problem-free rate.\n   [microtasks-in-broad-daylight.md](microtasks-in-broad-daylight.md) § 3.)*\n   *(Added 2026-10-05: the sitting accepted on 28 September as the measurement for this question\n   was set when every part was the splitter's first version. The splitter has a second version\n   since 30 September, and the sitting as built left its parts out. Fixed that day, and the\n   results are now read per version: of 51 parts, 45 are the first version's and 6 the second's,\n   so the sitting will say what the retired version did and little yet about the one that runs.\n   [pass-quality-and-the-choice-of-model.md](pass-quality-and-the-choice-of-model.md) § 6b.)*\n2. **What is the measurement?** It could be review in use (*\"weaving in the reviewing naturally\"*,\n   2026-09-21), an offline sitting with seeded wrongs (`machine-work-as-a-standing-property.md`\n   § 3d, with seeded wrongs kept offline since 2026-08-31), or time to first objection on natural\n   mistakes. Does rung three, sampled review with no panel, need a number he sets or one read off a\n   sitting? *Turns on*: zero rows in a week, and the parts review and the sameness sitting both still\n   unsat. *Researched 2026-09-29* (the arithmetic, not the ruling): [proving-a-pass-reliable.md](proving-a-pass-reliable.md).\n3. **Whose instrument is it?** The operator and recruited reviewers (authoritative in the operator\n   phase), the parent's author (an objection only that author can make), or every signed-in reader\n   (telemetry)? *Turns on*: his own description of the reviewers (2026-09-21), the phase split of\n   2026-08-31, and the fact that any account can press *Drop* today.\n4. **What does a *Drop* do?** Retire the link at once, retire it only with a reason (Q32's *\"visible\n   retirement with the reason\"*), or record the refusal as telemetry until a second check agrees?\n   *Turns on*: *bare judgments are telemetry*, Q33 (*\"never voted\"*), and his worry of 2026-09-19\n   about *\"saboteurs/spammers\"*.\n5. **Is there a keep at all?** The page has none. *Save wording* on unchanged text acts as one and\n   sets `confirmed=true`. *Turns on*: *\"there is no ratify act\"*, *\"meaningless buttons\"*, and Q32.\n6. **What does a reword do to the record, and in whose name?** Overwrite in place, as today, or add\n   the reworded claim beside the machine's (the additivity principle, and the 2013 prototype's\n   substitute propositions)? Is a reworded part still the parent author's words? *Turns on*:\n   *contributions are added, never overwritten*, and the two attribution rules in § 4c.\n7. **Where do machine-cut parts live, and how are they marked?** Default public at `candidate`, as\n   today, or draft until contested (the direction of 2026-08-25)? Should the map mark them, as rung\n   two requires? Should they keep the kind `extracted` or get their own kind? *Turns on*: 17 of 37\n   parts cut from unspoken premises render as spoken. Ledger R11 (*\"a tidy graph looks examined\"*)\n   and the tidying line's open questions also bear on it.\n8. **What happens to a cut through a concept?** An automatic re-file as support (proposed,\n   unruled), a control in the panel, or *Drop*? *Turns on*: the canon line approved 2026-09-21, and\n   his own reading, *\"backing the main … claim with subclaims\"*.\n9. **The four held proposals** (`claim_4d6779b25e02`, `claim_14241d73113b`, `claim_c444e59749ce`,\n   `claim_eb27c825a45e`). Keep, drop or re-run them? And should keeping a held part make the\n   machine's wording a `human` claim in the keeper's name, as the code does now? *Turns on*: the\n   coverage call found invented content in all four.\n10. **Signed-out readers.** Should they see machine splits, and the panel read-only? *Turns on*: the\n    *show, don't tell* aim he gave the tidying. `TODO.md` records the question as raised three times\n    and never ruled.\n11. **Enrichment of written parts.** Should the pattern classification run over the written links,\n    with critical questions left latent? *Turns on*: no part has any support, attack or vote, and a\n    part without critical questions offers a reader no move.\n\n---\n\n## Sources and queries\n\n**Corpus read for this document** (all under `docs/research/` unless noted):\n`earning-the-right-to-write.md` · `the-tidying-line.md` · `the-parts-review.md` ·\n`the-jhu-decomposition-line.md` · `reading-the-text-itself.md` ·\n`self-similar-decomposition-and-claim-ontology.md` · `decomposition-axes.md` ·\n`the-ratification-record-and-the-triage-feed.md` · `what-human-judgment-is-for.md` ·\n`what-confirmed-means.md` · `machine-work-as-a-standing-property.md` · `auto-connect-upgrade.md`\n§ 2c · `graph-daemons-design-space.md` · `core-periphery-and-the-concept-layer.md` § 5a ·\n`a-cut-through-a-concept-is-an-argument.md` · `scheme-bounded-decomposition-and-evidence-as-subgraph.md` ·\n`session10-empirical-collaboration-and-atomization-frontier.md` · `codex-session-20260403.md` ·\n`structuring-gradient-lineage.md` · `simplenote-archive-analysis.md` ·\n`meteor-prototype-analysis.md` · `disambiguation-ux-and-voice-first-contribution.md` ·\n`kartpaus-auto-enrichment-design.md` · `soft-canonical-clustering-and-reversible-merge-semantics.md` ·\n`conviction-and-critique.md` · `the-harness-conviction.md` · `cheap-to-add-slow-to-matter.md` ·\n`relevance-and-the-two-factors.md` · `engagement-gradient-priors.md` ·\n`jev-and-the-judgement-layer.md` · `session34-open-loops-and-unweighed-answers.md` ·\n`session35-the-verification-night.md` · `session26-open-rulings.md` (rows B2, B3, B7, C8, P12, Q29,\nQ32, Q33, R8–R14) · `docs/ux-principles.md` (P4, P14, P20, anti-patterns) · `docs/depth.md` ·\n`docs/history.md` · `docs/sketches.md` · `SESSION-HISTORY.md` · `TODO.md`. Archive: the two CAA\ntranscripts of January 2012 and the 2011 feature map, under `archive/demeter_2017_dropbox/fredrik/`.\nSession transcripts, read through `ccdistill`: `dc4eae5b` (2026-09-20 → 27) and today's session.\n\n**Read-only queries** (public graph `deliberus_extraction`, run with `ro_query`):\n\n```cypher\nMATCH ()-[r:DECOMPOSES_INTO]->() WHERE r.provenance='split' RETURN r.human_action, count(r)\nMATCH ()-[r:DECOMPOSES_INTO]->() WHERE r.provenance='split' RETURN r.coverage, count(r)\nMATCH ()-[r:DECOMPOSES_INTO]->() WHERE r.provenance='split' AND r.superseded_by IS NOT NULL RETURN count(r)\nMATCH ()-[r:DECOMPOSES_INTO]->() WHERE r.provenance='split' RETURN r.support_semantics_proposal, count(r)\nMATCH ()-[r:DECOMPOSES_INTO]->() WHERE r.provenance='split' RETURN r.proposal_agreement, count(r)\nMATCH ()-[r:DECOMPOSES_INTO]->(b) WHERE r.provenance='split' RETURN b.claim_kind, b.maturity_stage, count(b)\nMATCH (a)-[r:DECOMPOSES_INTO]->(b) WHERE r.provenance='split'\n  OPTIONAL MATCH (x)-[e]->(b) WHERE type(e) IN ['SUPPORTS','ATTACKS'] RETURN count(DISTINCT b), count(e)\nMATCH (c:Claim) WHERE c.split_proposal IS NOT NULL RETURN c.claim_id, c.claim_kind, c.split_proposal\n```\n"}