{"path":"research/one-challenge-one-home.md","content":"# One challenge, one home: inspecting the ontology through examples\n\n**Standing, 16 October 2026:** the founder asks for architecture, ontology and machine-work diagrams\nin Lins, and for detailed examples mapped through alternative ontologies before choosing among them.\nThe design below is Codex's analysis, not a ruling on a new schema or calculus. The existing ruled\ndistinctions and the proposed alternatives are shown separately.\n\n## The direction, in his words\n\nOn concept tracking: *\"whether a reason is acceptable/valid/relevant depends on what concept sense\nis in use in the claim(s) in focus (both in the mother claim and in the reasons (both support/attack\nlinks)\"*. On the example with *because*: *\"It should be split by the splitter (this has been proposed\nat an earlier date!) so that the reasons are separated from the main claim and can be\nargued/attacked/dealt with separately.\"*\n\nOn ontology: *\"we need to 'deduplicate' all features of the optimal ontology that we strive towards\"*\nand *\"There needs to not be multiple points of entry for reasoning/attack/support around a claim.\"*\nHe asks whether negated attacks duplicate supports, and whether criticizing a link's relevance claim\ncan absorb criticism that belongs to its reason claim.\n\nOn the method: *\"running examples by this structure and mapping them out visually in RELATION to the\nontology and possible VARIATIONS of the ontology is crucial\"*. Machine microtasks and human interaction\nbelong in that examination; abuse and spam should not distract from establishing the reasoning needs\nfirst. This specifies the order of this design work, not the removal of security safeguards.\n\nThe aim is analytic clarity that helps people agree more over time. He explicitly retains the positive\nhive-mind ambition while doubting that one systematization can capture everything true and valuable\nfor everyone. Context-sensitive senses are therefore instruments for making disagreements tractable,\nnot grounds for declaring them permanently incomparable.\n\n### Agreement as outcome, never as measure (17 October; his question, a session's answer, ruled the same day)\n\nAsked about the new wording on the public overview (*increasing agreement through reasoning*), he\nwrote: *\"Well, what's so wrong about seeking agreement? What are some counterarguments to this? Why\ndo I feel slightly uncomfortable with that? Liberal modern society doesn't \"want\" more\nconformity-enforcement than our culture already has, it probably always wants less of that, and we\ndon't want thought-police and consensus by bullying. But if we are helped and held in a facilitated,\nsafe, free process (free as in everyone is fundamentally free to think and decide whatever they want)\nboth through software and culture around something like Deliberus, and this leads to increasing\nagreement about what is true and valuable and what should be done, then I'd celebrate that, who\nwouldn't?\"*\n\nThe answer put to him: nothing is wrong with agreement as an outcome; the hazard is agreement as\nthe measure. A system scored on agreement produces it the cheapest way, and the cheap ways are\nalready measured in this corpus: long-form debate summarised by a model drifts toward the consensus\nmiddle with the tail gone (arXiv:2606.01736), writers drift toward the model's voice (the variance\nof writing complexity falls 21 to 50 percent under assistance,\n[why-representations-converge-and-whether-minds-do.md](why-representations-converge-and-whether-minds-do.md)),\nand participants preferred the facilitator that had moved their allocations (the DeepMind\nfacilitation study, N=879). From inside the score, agreement reached by reasons and agreement reached\nby copying are the same number, which is the assembly lens's copy number against convergence. So the\ndiscomfort is the confession principle applied to the aim: a system whose success is defined as\nagreement cannot confess the failure that looks like success.\n\nThe counterarguments that carry weight, with their homes: Rawls's burdens of judgment, that free\nreason under free institutions yields permanent reasonable pluralism, already the wager's ceiling\n([convergence-wager-red-team.md](convergence-wager-red-team.md)); Mill's *On Liberty*, that a truth\nheld without contest decays into dead dogma, so dissent is what keeps its grounds alive (not cited in\nthe corpus before this; the human-side form of *nothing is permanently terminal*); Lorenz et al. 2011,\nthat showing people one another's estimates narrows the range and raises confidence without raising\naccuracy, so agreement reached through mutual visibility spends the independence that made the group\nworth asking ([what-counts-as-evidence.md](what-counts-as-evidence.md)); Sunstein's incompletely\ntheorized agreements, that people act together because the reasons stay unspoken, which he holds as a\nquestion for a debate inside Deliberus ([legibility-under-power-red-team.md](legibility-under-power-red-team.md),\nhole 2); and Mouffe's agonism, that every consensus has an outside, so the question to put to any\nagreement is who was not in the room (not cited before this).\n\nWhat his own sentence supplies is the discriminator: *free*, and agreement that the instruments can\ntrace to reasons rather than to influence. That is the stance of 21 August (a lean must be\nattributable to the contributed reasoning) and the ruling of 10 October (ledger S8: members keep\ntheir own position, the ability to check, and the freedom to leave). Proposed for Thread 8 of\n[conceptual-threads.md](../conceptual-threads.md), in place of the 16 October sentence, awaiting his\nruling: *Deliberus welcomes the agreement that clarity earns and never optimizes for agreement, since\na system scored on agreement produces it the cheap way, and from inside the score agreement by reasons\nand agreement by influence look the same.*\n\n**RULED 2026-10-17**, founder: *\"Great. Adopt it\"*. The sentence stands in Thread 8 of\n[conceptual-threads.md](../conceptual-threads.md) in place of the 16 October wording, and the thread's\nopening line was brought into line with it in the same edit (*seeks clarity and welcomes the reasoned\nagreement it earns*).\n\n## 1. What should be deduplicated?\n\nFour possible duplicates need different treatments:\n\n| Possible duplicate | Candidate treatment | What must remain distinguishable |\n|---|---|---|\n| Two renderings of the same assertion in the same scope and sense | Reversible sameness grouping, retaining the uses and their attribution | Different speakers, provenance, dates, and refusals of a rendering |\n| A proposition and its exact negation | Two opposed answers to one scoped question | Opposite commitments; neither is the other's paraphrase |\n| One objection reached through a reason, a critical question and a relevance panel | One contribution with explicit targets and views onto it | A challenge to a premise versus a challenge to the inference using it |\n| One finding influencing the same result through several paths | Shared provenance and an explicit account of the paths | Distinct inferential uses versus independent evidence; the aggregation remains an open design choice |\n\nThis is a proposed interpretation of his direction: **one semantic home for the same issue, reachable\nwhere it is relevant**. Several interface routes can open that home without creating several stores\nof reasoning. If a contribution makes two separable assertions, splitting it can reveal two legitimate\ntargets. If one assertion genuinely bears on two targets, it can be referenced at both without being\ncopied or counted as two independent observations.\n\nThe existing sameness work already preserves original members. The seven-key keeper also distinguishes\nopposition from sameness. However, its `opposes` includes both exact contradiction and contrariety:\ntwo alternatives that cannot both hold need not exhaust the possibilities. That key alone cannot justify\nturning a denial of either into an affirmation of the other. See\n[soft canonical clustering](soft-canonical-clustering-and-reversible-merge-semantics.md) and\n[the measured task-definition differences](microtasks-in-broad-daylight.md).\n\n## 2. Sense and context: separate what changes\n\nConsider the made-up source sentence:\n\n> The hall will be comfortable because it will be warm and quiet.\n\nThe earlier splitter proposal is **T9**, sharpened in [A part and a reason](a-part-and-a-reason.md)\n§ 11 and [Microtasks in broad daylight](microtasks-in-broad-daylight.md) § 3h. The whole utterance\nasserts the reasons as well as the conclusion. Its pieces should remain attributable to that utterance,\nand the reasons should be related to the separately addressable conclusion as reasons. The user's\npresent message reaffirms the intended behavior. It does not settle whether those two reasons make one\njoint argument, two independent arguments, or some more conditional combination.\n\nThree readings of *comfortable* make a useful diagram:\n\n1. **A stipulated criterion.** For this exercise, the speaker explicitly defines comfort as a particular\n   combination of warmth and quiet. The definition itself is recorded and its application inspectable.\n2. **An ordinary, defeasible indication.** Warmth and quiet are reasons for expecting comfort, but do not\n   exhaust what it means or guarantee it in this case.\n3. **A different relevant context.** The claim is about a wheelchair user's visit, and access or the\n   seating matters. This might expose a different sense, an omitted condition, a change of the person\n   or situation, or a disagreement over which consideration matters. It is not automatically a new sense.\n\nA sense per use can capture genuine meaning differences, but creating a new sense for every changed\noutcome would explain away disagreement. Distinguish **meaning**, **referent and occasion**,\n**criteria or thresholds**, and **the evidence for applying the concept**. Two people may mean the same\nthing and disagree over whether it holds. They may also use different terms and draw a valid inference\nthrough a stated bridge. Matching sense identifiers is neither a general relevance proof nor a demand\nthat every good inference share vocabulary.\n\nThe earlier design is [Retrofitting the reasons channel](retrofitting-the-reasons-channel.md) § 3,\nwith [Core, periphery and the concept layer](core-periphery-and-the-concept-layer.md) and\n[Store the uses, read the meaning](store-the-uses-read-the-meaning.md). This extends their per-use\ninterpretation to both ends of supporting and attacking links and to the steps between them.\nAcceptability concerns a claim under its reading; validity concerns an inference under its stated\ninterpretations and rules; relevance concerns the bearing on this target. Sense tracking helps specify\nall three questions and answers none by itself.\n\n### 2a. A shared case, and a reason that states its combination (17 October)\n\nThe founder challenges why grouping should obstruct contributing reasons to the shared graph:\n*\"if some says 'X because A and B' then it's just a boon if someone else comes up with C as a\nstrong reason for X. And if someone shoots down A, then so be it, that's our collective best\nguesstimate\"*. Asked whether A and B bear separately or together, he proposes: *\"Well then a reason\nshould argue that inside the text of its claim I guess?\"*\n\n**The shared case is the point.** A source supplies a contribution to the growing case for X;\nits author does not own or close that case. Another person's C can support the same conclusion,\nand criticism of A belongs in the same shared record. Attribution says who supplied which\nargument; it does not restrict what others may add. The grouping question concerns dependence\nwithin an argument, not permission to combine people's contributions.\n\n**A concrete candidate using ordinary claims.** In a toy example, let X be *the courtyard gate is\nopen now*. A supporting claim R says *today is Tuesday, and this gate is open every Tuesday*.\nThose two asserted components are genuine parts of R. They are separately disputable, and the\nlink R → X says what they jointly establish. A further reason C, *the current camera image shows\nthe gate open*, bears on X through another argument. Learning that it is Wednesday defeats R's\nTuesday argument; it leaves the camera argument to be evaluated on its own merits. The general\nTuesday rule alone should not continue contributing as if it still establishes today's opening.\n\n**Correction and lineage, 17 October.** The founder points out that B is the rule for this inference,\nnot an independent reason for X without A. He is right: the relevant distinction was already explicit\nin [support semantics](support-semantics-and-the-conjunction-problem.md), corrected for the\nparts/reasons confusion in [the retrofit](retrofitting-the-reasons-channel.md) § 1, and given its\nvertical floor in [the chain note](the-chain-inside-a-link.md) § 10.3 (AG14, ruled 14 October).\nThe [recovered Socrates drawing](../sketches.md#additional-recovered-sheet-socrates-inference-and-navigation) already joins a\ncase and a general rule beneath one validity check. The open question here is how to represent and\ndisplay that established distinction with the fewest duplicate homes. It is not whether dependent\npremises exist or whether the collective case may receive more reasons.\n\n![The Socrates sheet: case, rule and validity, with navigation (held privately)](../../.private/sketches/socrates-inference-and-navigation.jpg)\n\nFor the strict toy deduction, B means *at the time being discussed, if it is Tuesday then this gate\nis open*. The written \"every Tuesday\" needs that scope: *opens for some hours on Tuesdays* would not\nestablish *open now*. The general schedule is an ordinary, challengeable claim in its own right;\nits **role in this application** is the rule. Asserting the rule and checking its application are\ndifferent questions. When the step is sound under that rule and scope, A supplies the case and X\nis the conclusion. A multi-step path must expose intermediate conclusions until each application\nmeets the ruled floor. A composite sentence is a possible container for this structure, not a way\nto hide the case/rule distinction inside prose.\n\nThis can be drawn using existing kinds of object:\n\n```text\nA: today is Tuesday ───────────────┐\n                                  ├─ parts of R: A and B ── supports ──► X\nB: this gate is open Tuesdays ─────┘                                  ▲\n                                                                     │\nC: current camera image shows it open ─────────────────── supports ────┘\n```\n\nThis is a candidate representation, not an implemented or ruled grouping mechanism. It may avoid\na new bundle node type: an ordinary composite claim holds the joint premises, and the existing\nlink carries their bearing on X. It connects directly to [A part and a reason](a-part-and-a-reason.md)\n§ 5 and § 11: parts of a reason can be required for that reason without becoming required parts\nof its conclusion. The splitter must preserve that grouping when exposing the pieces; flattening\neverything into separate supports for X would lose what the sentence says about dependence.\n\nThe same candidate needs a deduplication check against the alternative *A supports X, conditional\non B*, with B under the link. They can express the same inference. Recording or showing both must\nnot create two independent arguments. The claim that *A and B imply X* also differs from asserting\nA and B: a warranted conditional supplies no evidence that its conditions hold. Merely writing the\nwhole argument in a sentence therefore does not settle its standing or its inference's relevance.\n\n**And/or is a clue, not the grouping rule.** *X because A and B* can give two separate good reasons\nor premises that only work together. *Either reason would suffice* describes resilience to the loss\nof one argument, not necessarily uncertainty. *I know A or B, but not which* is different again: the\nsource has not asserted either separately, so a split must not invent both. People can contribute\nin ordinary language; the machine proposes the structure and shows why, with correction available.\nThe design comparison still owes a test of this composite-claim representation in Lins, including\nthe risk of representing its dependency a second time under the relevance claim.\n\n**The two addresses, drawn (17 October, evening).** The tenth view, `09_gate`, puts the composite\nreason with its parts A and B beside the step under the link R → X, whose machine-written sentences\n(a claim the moment they are written, ledger AG12) restate A, B and the application. An attack on\n*today is Tuesday* then has two addresses: part A of R, and the step's own sentence. Where the same\nscoped rule is reused, the existing rule claim is referenced, and repeating its wording in a source or a\ndisplay creates no second argument; a step references existing claims rather than restating them. A\nsession's recommendation on where that claim sits, unruled (ledger AL4): under the link (the licence as the\ngrounds of the relevance claim, AD22), with a composite supporter holding only joint facts of the case\n(*warm and quiet*); whether a composite reason may instead include the rule by reference stays open with AL3. For the splitter that is one more thing beyond T9: among a sentence's\nreasons it tells a case from a rule, since a rule split off as an independent reason goes on\ncontributing after its case has failed. His verdict on the views, the same evening: cards that\ncomment on the ontology are not the ontology (*\"I can't for the life of me find a view that shows\npedagogically and exactly what the ontology would hold\"*), and the next build, accepted (*\"Yes please\nrebuild it the way you propose, so I can look at an actual realistic example of the different example\narguments and variations in ontology that we can ponder and try and poke holes in etc etc\"*): small\ngraphs holding the gate and comfort cases as real claims with the production link types, one graph\nper variation (as stored today; the composite reason; the rule under the link; an explicit\napplication node), so that what each variation stores can be read off the boxes.\n\n### 2b. Directly checkable facts: time is an instrument input\n\nThe founder requests automatic evaluation of facts such as *today is Tuesday*, including different\ntimezones. This is a requested capability; the following contract is a design proposal, not shipped\nbehavior. The existing `deliberus/staleness.py` compares evidence dates with age thresholds and proposes\nreview. That is not a calendar/clock evaluator for arbitrary claims, and its human-ratification rule\nshould not be mistaken for a requirement to vote on a computed weekday.\n\n- Resolve *today* and *now* to the intended occasion and timezone. In an imported utterance that may\n  be its speaker's then, not the ingestion date or the present reader's now. Keep that interpretation\n  inspectable; if it is missing, ask for it rather than silently choosing the server's zone.\n- Given an explicit instant and named timezone, code computes the date, weekday and local time.\n  The model may propose the interpretation; it should not supply the clock fact from memory or an\n  estimated confidence. A check records its inputs, result, time and instrument/source.\n- A live *now* view refreshes. A claim about a fixed past occasion does not become false at midnight.\n  Reuse the same scoped check across the applications depending on it; multiple views do not create\n  multiple independent observations. A missing or failed instrument returns an explicit unknown.\n- In the gate example this can settle A under its stated interpretation. It cannot verify the gate's\n  schedule B or that the physical gate follows it. Those remain separately challengeable claims.\n\nThe minimal demonstration should show the same instant in two zones on different dates, a local\nmidnight transition, an imported past utterance, and a failed check. It should distinguish the tool\nresult from the interpretation and from its eventual effect on the claim's displayed standing. The\nautomatic checking direction is his; a new strength bypass or numeric certainty rule is not decided\nby it. This complements the evidence instruments in [What counts as evidence](what-counts-as-evidence.md).\n\n### 2c. The relevance claim: a copy, a tie, and a reading (17 October, his questions)\n\nHis questions, on seeing the two-address view: *\"Are the relevance-claims keeping actual inline\ncopies of the claims it's about? What about placeholders that prefill based on the actual claims it\nrelates to, such that it never drifts? Can a relevance-claim be tied to a support/attack-link? What's\ntechnically possible to tie to an edge in FalkorDB/graph databases? Can a support-link get a link to a\nclaim (that argues about relevance)? And is arguing about relevance not very closely related - perhaps\noverlapping and as such being a duplicate point of entry into the argumentation (a gap to close, then)\n- to the concept tracking playing a very similar role, as argued in the recent Codex session?\"* The\nlast one is his own question from that session, 16 October: *\"Could the reasoning/challenging of the\nrelevance-claim (attached to the link that goes from a claim to its reasons) be a duplicate point of\nentry such that it could eat up some claims that might better belong as critique of the reason-claim\nitself?\"*\n\n**A copy, and it can drift.** The machine's relevance claim is written as *\"<first headline>\" bears\non \"<second headline>\".* from the two claims' short titles at the moment it is written\n(`machine_sentences.relevance_claim_for`). Nothing re-renders it: the title pass leaves machine\nsentences out, and no pass revisits a relevance claim when a headline it copied changes (checked\n2026-10-17, `scripts/apply_titles.py` and `deliberus/short_title.py`). And every relevance claim into\none target reads the same, *Bears on: <target>*, so three side by side cannot be told apart by their\ntitles; the example graphs title them by the link for that reason.\n\n**The tie exists, by ids and not by an edge.** The support or attack link carries `relevance_claim`\n(the claim's id) and `relevance_claim_at`; the relevance claim carries an `ARGUES_RELEVANCE` link to\nthe target whose `source_id` names the first claim (ledger Q33, built 2026-09-28). What a property\ngraph allows, probed on the embedded FalkorDB the same day: a link carries properties, arrays among\nthem, and has an internal id; two links of one type may run between the same pair; **a link cannot be\nthe endpoint of another link** (the query is refused). So a claim is tied to a link either by an id\nheld on the link, which is what the store does, or by giving the link a node that stands for it, which\nis what the relevance claim already is: the reification of the link, the thing a person attacks when\nthey attack the bearing. A statement-about-a-statement store (RDF-star) would allow the direct form;\nnothing here needs it. One wrinkle: with two links allowed between one pair, `source_id` plus target\ndoes not pick a link; the link's own `relevance_claim` does.\n\n**A placeholder that never drifts is the critical-question mechanism applied to relevance.** A critical\nquestion is a potential, rendered on read from scheme, template and link, and minted when somebody\nanswers it (ruled; `latent_questions`). A relevance claim can be the same object: a potential rendered\nfrom the link's two endpoints, minted the moment a machine writes steps under it (AG12) or a person\nargues it, and even then carrying the template and the two ids so that its text is read, never copied.\nNothing in AG12 is against this: the sentence becomes a claim when written, and a potential is not yet\nwritten. **His follow-up the same night names the condition**: *\"if it's ever minted and the claims it\ntalks about are rephrased then it drifts. Or are claims never rephrased after created? Can we make\nsure/guarantee they never are?\"* A minted relevance claim never drifts only if one of two things holds:\nits text is rendered on read from the two ids (what is proposed), or the texts it copies never change.\nThe second is the storage law, § 2d.\n\n**Relevance is a reading over three grounds, and only one of them is new.** What a link rests on is\ndrawn as the grounds of its relevance claim (AD22); those grounds are (i) the case fact, which is the\nreason's own standing (Q33's first factor), (ii) the rule or licence and its application to this case\n(the step, AG14), and (iii) that no shared term shifts its sense unnoticed between the reason and the claim\n(a difference in sense that is stated and bridged is allowed),\nwhich is concept tracking's job and a code-computed input to relevance (Q39, raised by him on\n2026-10-15). A bare *this does not bear* is the shrug the single-exit rule removed elsewhere; the real\nentry points are a factual challenge (home: the reason), a rule challenge (home: the step under the\nlink) and a definitional challenge (home: the concept layer). **So the overlap he suspects is real and\npartial, and it is the third ground**: today a definitional objection can be filed twice, as a\ndefinition of the word and as an attack on the relevance claim (the comfort example draws exactly\nthat, o3 defines the conclusion and attacks the relevance claim), and the two are not joined.\n**Proposed (ledger AL8): a definitional challenge to a term the reason shares with its claim lands\nonce, in the concept layer, and reaches relevance through the sense-alignment input; the relevance\nclaim is never argued bare, it is read from its three grounds, and a person who argues bearing does so\nby contributing a ground.** That closes his duplicate point of entry without removing the relevance\nclaim, which stays as the container the link's grounds hang under.\n\n**Can relevance-arguing go altogether, and is there an initial machine pass?** (his follow-up the\nsame night: *\"So does it make sense that we even have relevance-arguing or can we delegate all of that\nresponsibility to concept-tracking? Or is that not feasible? What if we had an initial machine pass of\nrelevance-judging - or do we have that already (at least proposed somewhere?)? How would that work and\nrelate to any other relevance judgments/arguments after that initial pass?\"*). Not all of it: no unnoticed\nequivocation is necessary and not sufficient, since two claims can share every term in one sense and\nstill have no bearing on each other; the rule that joins them is the other ground, and it is a claim\nof its own. What goes is relevance as a bare argued object. An initial machine pass is ruled (Q33's\n*proposed* tier: a second-family judge with an abstention) and built as a script that asks for a\nnumber (`scripts/propose_link_relevance.py`, dry-run by default; writing to production needs his\napproval), and that number was measured unsteady under rewording and premise order on 7 October, which\nis why the standing proposal is that the initial pass *writes the grounds* instead: the rule, the tie\nof this case to it, and the sense alignment, under the link, as machine sentences (AG12), checked by\nthe no-gap and nothing-removable tests (AG14), with the link's relevance read from them: the gate is\nthat a step with no gap exists, the magnitude is the rule claim's own standing and scope. Everything\nafter that is a contribution to a ground: attack the rule claim, file a definitional claim in the\nconcept layer, attack the case fact; the reading recomputes; and a person's own relevance claim takes\nthe container, which is built (`hand_over_relevance`). The link's connection questions become the\nstep's sentences shown as questions with a move beside each (AD17). Filed on AL8.\n\n### 2d. Nothing is rephrased in place, and the entry order is split, then sameness, then link (17 October)\n\nHis questions: *\"Or are claims never rephrased after created? Can we make sure/guarantee they never\nare? All claims need to be split using the splitter when entering Deliberus. Then they need to be\nsameness-clustered/merged, after splitting, not before, very important, right?\"* And, when the answer\nbegan from the code: *\"Not just about the code; do we foresee any NEED to do this or not?\"*\n\n**No need is foreseen, and every case that looks like one resolves to a new claim.** The storage law\nproposed as ledger AA1 already says it: a use (what somebody said) is kept and never changed, a machine\njudgment is recorded once, and a reading is computed and at most cached. Run the cases that tempt an\nedit through it: a clearer wording of the same proposition is a new rendering, grouped `same` with the\nold by the keeper, the group carrying the strength and the old rendering kept with its attribution\n(the merge design, and the first row of § 1); an author's *that is not what I meant* is a refusal of a\nrendering plus a new claim, never an edit, since the refusal is the act that cannot be wrong and the\nrestatement is a fresh assertion; a typo is the same case at small scale, and an edit window would\nsilently change what every attack, vote and relevance claim already on the node was about; a fresh\ncut by a better splitter supersedes the old parts and removes nothing (`scripts/recut_splits.py`); a\nmachine's sentence rewritten by a later pass is a new sentence with the old one superseded; a\ntranslation is another rendering; a redaction for privacy retires the claim and withholds its text\nrather than editing it. The decontextualiser does rewrite a source sentence, but before the node\nexists, and the source span is kept beside it. What *may* change is the headline (`short_title`): it\nis a reading of the text, recomputed by the title pass, which is exactly why nothing stored may copy\na headline (§ 2c). Until 2026-10-17 one route edited a claim's text in place: the split review's *reword*\naction, where a person rewords a machine-cut part of their own sentence (`proposed_split.record_action`,\nbuilt 2026-10-05 on his word). The creation functions set a text once, by a fresh id. **Ruled and\nbuilt 2026-10-17 (ledger AL10; his word: *\"Can you build AL10 then please?\"*):** the reword mints the\nreworded part as a new claim beside the old one, attributed to the author when the author rewords and\nto the reviewer otherwise, confirmed, with its own part link to the sentence; the old link is marked\n`superseded_by` the new id, and the old part moves to the `superseded` stage with its text and\neverything on it kept; the claim page lists the old part among the replaced ones.\n`tests/test_claim_text_is_never_rewritten.py` reads the code and fails on any in-place edit of a\nclaim's text; its exception list is empty. Saving unchanged wording writes nothing (before, it recorded a\nreword with the confirmation set; now it would also have minted a duplicate and stranded the old part's\nlinks), and an action on a part already replaced or dropped answers 409, so a second touch lands on the\npart that stands. The fault had been found once before: a reader of the vision pages named it on\n2026-10-11 against *contributions are added, never overwritten*, the split-review audit had recorded the\nconflict, and `TODO.md` held the repair; this session did not find either record until after the build. The row: a claim's text is never rewritten; a rephrase is\na new claim grouped with the old; the headline is a reading; nothing stored copies a headline (the\nlast clause rides on AL8, the relevance claim rendering from ids, and is unbuilt). The first clause\nentered the canon the same day, on *contributions are added, never overwritten* (founder: *\"Yes, add\nit.\"*; his note on the draft wording, *\"'beside the old one' is unclear tho maybe?\"*, is why the line\nsays the old claim stays, marked as replaced, with everything that was attached to it). The public vision page carries it in visitors' words on his *\"Sure\"* of the same evening (`docs/vision.md` § A corpus cannot be un-thought).\n\n**Split first, then sameness, then link: right, and the live order is the reverse.** The corpus holds\nthe principle in his own one-step invariant, which is *earned* and so runs leaf-first, and in the ruling\nof 13 September that a `contains` or `overlaps` verdict means the separability floor was not reached\nand raises a decomposition proposal, after which the pair re-labels at leaf grain. What enters today\nruns extraction (its own atomisation) → linking, where the keeper decides sameness on claims at\nextraction grain → the splitter later, by script, on chosen claims, with the reasons-off cut (T9)\nunbuilt. So a bundle can be judged `same` as, or linked to, another claim before its parts exist, and\nthe repair is a round trip. Proposed as ledger AL9: on entry a claim is cut to the floor first (parts,\nand reasons off), sameness is decided among the pieces, and links are written last; `contains` and\n`overlaps` then become rare, since both sides stand at the floor. Two things the order does not settle:\nthe splitter's measured error rate (his review of its cuts is pending) now sits in front of every\nclaim, which is the trust-ladder question for that pass; and a machine-cut part enters the sameness\npool as `extracted`, attributed to the author, which is as today.\n\n**Built the same night, behind a switch (his word: *\"Ok that needs to be fixed, can we guarantee\nsplitting is done first?\"*; AL9 ruled).** `deliberus/floor.py`, on `DELIBERUS_FLOOR_FIRST=1`: the\nshared post-store pass cuts every entering claim the splitter judges bundled, and its parts again\nwhile they look bundled, three levels deep at most, with no claim limit; and the linking step reads a\npair only when both its claims stand at the floor, holding and counting the rest (a bundle with live\nparts, a bundled-looking claim the splitter never read, one it read and refused to write, and one that\ngives its own reason inline that the reasons-off reading has not been through). The gate sits in\nthe one linking function every path calls, so a script that links one source, a cut that failed and a\npre-run that hit a limit are all held the same way, which is what *guarantee* means here. Off by\ndefault: a cut is one to two model calls a claim, which the production key's free tier cannot carry,\nso switching it on in production is his call. Twenty tests, `tests/test_floor.py`.\n\n**T9 built the same night (his word: *\"Build T9 then pls\"*).** The same pass reads a second shape\nsince 2026-10-19, the conditional (ledger AM5 and AM6, his *\"Build it\"*): *if P, then Q* becomes two\nclaims nobody asserts (kind `supposed`), a support link P → Q carrying the author's hedge as stance,\nand the sentence as the link's relevance claim by id; the entry order he asked for (cut first) is\nstill behind `DELIBERUS_FLOOR_FIRST`. The *because*-shape, as built: `deliberus/reasons_off.py`: a sentence\nthat gives its own reason inline (*because*, *since*, *therefore*, *hence*, *so that*, *given that*,\n*for this reason*) is read by one call that names its conclusion and its reasons, each reason tagged\n*case* (a fact of the situation) or *rule* (what licenses the step), and a separate coverage call\nchecks that the pieces repeat nothing and invent nothing, the check the splitter already uses. What\nis written: the conclusion and the reasons as the author's claims (`argument_label` *conclusion*,\n*reason (case)*, *reason (rule)*); the sentence decomposed into all of them, provenance `split`,\nconfirmed unset; a support link from each case reason to the conclusion, provenance `split`, with no\nscheme; and the rule as the author's claim under that link's relevance claim\n(`machine_sentences.relevance_claim_for`), never a support link of its own, his correction of 17\nOctober. A reading that refuses (a conclusion that repeats the whole, reasons with no conclusion, an\ninvented piece) writes nothing and holds the sentence for a person. The floor step runs it before the\nparts cut, so the splitter cuts the pieces and never the sentence. Two calls the build made without\nasking: the sentence decomposes into every piece, reasons included, which keeps the sentence as the\nauthor's argument (`a-part-and-a-reason.md` § 11) and not only its conclusion; and a rule with no\ncase reason beside it hangs nowhere and is counted. `scripts/reasons_off.py` reads one claim or one\nsource, dry by default (`--write` writes; on the production graph that is his call, and no person has\nread one of its cuts yet). Sixteen tests, `tests/test_reasons_off.py`. Not decided by the build: AL3\n(the composite reason) and AL4 (the rule's home).\n\n**What supersession is, and the coherence it guarantees** (his question the same night: *\"What is\nthis supersede thing, again? How does it work? What does it do to associated claims? How can we\nguarantee that everything is coherent?\"*). Nothing in the graph is deleted or rewritten; a thing\nthat is replaced is *marked*. A link gets `superseded_by` (what replaced it: a recut's marker, the id\nof the person's claim that took a machine's relevance claim up, or a timestamp) and `superseded_at`;\na claim that is replaced gets `maturity_stage = 'superseded'` and `superseded_by`. Every reader that\ncounts part links, strength inputs or a machine's sentences filters to the live ones, and a test reads\nthe code for a reader that forgot (`tests/test_part_of_readers.py`); the default public reads hold the\n`superseded` stage back (`graph/maturity.py`). **What it does to the claims linked to the old one:\nnothing.** An attack on the old wording stays on the old wording, visibly, with its history; it is\nnot moved onto the new claim, because the attacker never saw that sentence, which is exactly the\nmismatch he is afraid of. So coherence is guaranteed by three things: a text is written once (AL10,\nwith its test), a replacement is marked and never erases, and readers filter to the live. What is\n*not* guaranteed today is continuity of evidence across a rephrase: the attacks on the old wording\ndo not count for the new one. That is the sameness group's job, *the group carries the strength*,\nwhich is designed (stage four of the sameness ladder) and unbuilt; until then an old wording and its\nreplacement are two claims the keeper may mark `same`, and a reader sees both.\n\n**His follow-up: *\"What if a rephrase causes a semantic mismatch between the linked claims all\naround the superseded claim in question?\"*** It cannot, and the reason is where the links sit. A\nlink is written against the text of its two endpoints, and under AL10 that text never changes, so\nevery link in the graph still says what it said about the sentences it was written about; a\nrephrase creates a new node with no links but its own part link, and moves none of the old ones.\nThe mismatch he is afraid of needs a link whose endpoint changed under it, and no operation does\nthat. What a rephrase *can* do is put two claims that differ in meaning into one sameness group, if\nthe keeper calls a changed meaning `same`; then the group would carry strength across a gap, which\nis a false merge, the sameness bar's question (his labelling sitting, R12 and T5; the false-merge\nalarm, stage five of the ladder, unbuilt) and not a rephrase question. A reader who finds attacks on\nthe old wording and none on the new sees exactly what happened, which is the honest state.\n\n### 2f. A theory's claims: a party in reported speech, not a node (17 October)\n\nHis question, on a sameness pair whose second claim read *Global workspace theory proposes that conscious\nexperiences in humans occur when information is made widely available to other parts of the brain*:\n*\"When a person has claimed a bunch of various claims, we use reported-speech, right? And what does that\nresult in, in terms of how we store and annotate the claim? But if a theory claims different things,\nsimilarly, we just keep that as part of the claim? Is this good or bad?\"* and *\"Perhaps just like we\nhave Sources as an entity in the database (right?) we could have Theories or Hypotheses, which are\nlinked to the claims they make, claims can thus belong to a Theory or to multiple theories? Does that\nmake sense, is it useful? Then the claim would be linked and at the same time shortened to 'Conscious\nexperiences in humans occur when information is made widely available to other parts of the brain' if\nwe decide that.\"*\n\n**What reported speech does today.** When a source says *X argues P*, the sentence is stored as the\nauthor's claim that X holds P, and P is minted as its own claim, attributed to X and asserted by nobody\n(kind `minted_report`), the two joined by a link no strength computation reads (`REPORTS`); a source\nthat says X did *not* hold P stores P as nobody's with X in `not_held_by` (`deliberus/reported_speech.py`).\nMeasured 2026-10-17: the pass already treats *Aristotle asserts P* and *Research indicates P* as reports\nand mints P (attributed to Aristotle, to *Research*), while *Global workspace theory proposes P* was\nkept whole as one extracted claim. So the reading of what counts as a party is inconsistent, and the\nsameness sitting as served holds two pairs that are a report beside its own minted content.\n\n**Keeping the theory inside the claim is bad, for the reason this note exists.** *GWT proposes P* fuses\ntwo things a person might challenge, whether GWT proposes it and whether P holds, into one node, so an\nattack has two homes in one address; and the sameness question is put to *GWT proposes P* against\nsomebody's assertion of P, which is the comparison his pair 2 showed to be muddled.\n\n**The proposal, ledger AL14: a named theory, hypothesis, school or body of research is a party.** P is\nminted and attributed to the theory, the reporting sentence stays as the author's claim, and what a\ntheory claims is *read* from the claims attributed to it, the way a party's page would be, with the\ntheory's name a term in the concept layer. That yields exactly his shortened claim with no new node. Why\nnot a Theory node beside Source: by the four sorting questions of `claim-or-system.md`, *theory T holds\nP* is something two people can disagree about, so it is a claim and not form; a Source is a provenance\nrecord nobody disputes, which is why it is a node; and a stored tenet-set would be the kind of thing\nAA1 says is read, never set. Costs: attribution normalization must canonicalize theory names (*GWT*,\n*global workspace theory*, *Baars' theory*); a theory does not assert, its proponents do, which\n`attributed_to` already tolerates for groups; every such sentence mints one more claim asserted by\nnobody, counted separately as all minted claims are; and an author's own theory would be minted a\nsecond time until a source's author is stored (AB15). First step, if adopted: a dry reading over the\nstored reports and the extracted claims that name a theory, to count what it would mint.\n\n**His answer, and the count.** *\"Yes, adopt AL14 and run the dry count. Or wait, what exactly is AL14:\nmy proposal for a new node, or your take?\"* It is Claude's take, not his node (answered 18 October), so\nAL14 stays open until he chooses; a reconciliation of the two is also on the table, the theory's name as\na term in the concept layer and its tenets as claims attributed to it. The dry count: a narrow net (a claim naming a theory, hypothesis, theorem, thesis, doctrine or paradigm beside a verb of holding) finds 25 of the 2,223 extracted and minted claims, all of them extracted, about ten naming global workspace theory, and some of the 25 are evidence claims (*supports the theory*) rather than reports of what a theory holds; a looser net that also takes *model, framework, principle, law* finds 74, padded with legal frameworks and principles. So adoption would touch at most about 25 claims, mostly in one debate.\n\n### 2e. The forecasting connection, in his intended sense (16 October)\n\nHe pasted a forecasting post (good forecasts rest on a small set of invariants; a violated invariant is\nwhat contingency planning is for) and corrected the first reading of it: *\"That's not really what I\nmeant by pasting that screenshot. It's more that the thinking that will go on inside Deliberus and\nthrough it may be inspired by that idea.\"* Not advice to simplify Deliberus. The moves it inspires,\nread against the graph: compare cases to find which features stay stable and which merely coincided\n(the compared-cases origin of a rule, `what-a-rule-rests-on.md`); state a forecast compactly while the\nlonger argument stays addressable; separate the conditional prediction from confidence that its\nconditions hold; vary one invariant and see which route to the forecast fails and which survive;\nnotice when several forecasts lean on one generalisation, so their number is not independent\nconfirmation; use the outcome to revisit the rule's scope or the case's classification rather than\nonly marking the prediction right or wrong. A worked example (whether a community event sells out)\nis a candidate for the example graphs; nothing is built.\n\n### 2g. After a cut: the whole as a record, and a fresh cut that matches before it mints (19 October)\n\nHis question, after a reply said that cutting a claim again keeps the old cut: *\"But I hope the\ncompound statements that the splitter cuts don't remain in the graph so ppl start arguing about them\nand it all gets messy, since those interactions belong on the separately addressable points of entry\naround the resulting cut claims? And what happens to any reasoning around those claims when/if the\nsplitter's design is revised and the graph potentially recreates the cut parts/claims from scratch?\nHow do we robustly wisely handle this?\"* It is his direction of 16 October applied to a cut: *\"There\nneeds to not be multiple points of entry for reasoning/attack/support around a claim.\"*\n\n**What happens today, read in the code and on the live graph that day.**\n\n- A cut writes parts under the claim and leaves the claim itself untouched (`proposed_split.commit`).\n  Its standing is read from its parts, the weakest one governing, and also from any reason that points\n  at the claim itself (`graph/badge_inputs.py` reads both), so the whole is still a second place to\n  argue. The claim page still takes responses on it, and the linking step still links it unless the\n  cut-first order is switched on (`DELIBERUS_FLOOR_FIRST`, AL9, off by default).\n- Measured: 13 claims have live machine-cut parts. 19 reasons point at those wholes, all written by\n  machines (14 at extraction, 5 by auto-connect) and none by a person; 33 links run from them, 3 of\n  them written by people.\n- Cutting a claim again (`proposed_split.recut`) mints every part under a new id, even a part that\n  says exactly what an old one said, and then retires the old parts' links. A retired part's own claim\n  leaves circulation unless something else is linked to it, in which case it stays, detached from the\n  whole. So a reason on an old part stays where it was written, visibly, and stops counting toward the\n  whole, and the fresh part that says the same thing starts bare. The record keeps everything; what is\n  lost is the effect, and nothing says so.\n- A fresh cut stops only where a person kept or reworded a part, not where a part carries a reason.\n  No harm has come of it yet: of the 45 parts replaced on 6 October, none carried a reason beyond the\n  link to its claim, and no fresh part repeated an old one word for word.\n\n**Proposal one (ledger AL15): reasons about what a cut claim says go to its parts, and the whole\nstays as a record.** The whole keeps four jobs: the record of what was said (the use, which AA1 and\n*contributions are added, never overwritten* keep unchanged, and which the parts only render); a\nstanding read from its parts; a premise others may cite where a link needs the parts together, where\nits weakest-part standing is the right one; and one question of its own, whether the cut is faithful\nto what was said (the split review: keep, reword, drop, and *the parts leave something out*). It loses\none: a place to argue about its content. Each reason about its content goes to the part it bears on,\nso one issue has one home (AL1). Three routes:\n\n1. New machine links: the cut-first order already does it (AL9, built behind its switch).\n2. Machine links already on a whole: each is read once against the parts (*which of these does it\n   bear on: one, several, the parts together, none?*), written on that part and superseded on the\n   whole, the old link kept visible. Whether that may happen without a person depends on R12: the\n   reading is a machine judgment with no measured rate yet, so under R12's reading it is written as a\n   proposal beside the link, and the old link keeps counting until the rate is measured. ⚠ R12 was\n   ruled later that day, for retiring a machine link on a later reading: *\"I think that's fair. It's\n   always recoverable/reversible, right, since it's all logged…\"*. What still holds this move is T1,\n   since writing a machine link onto a machine-cut part is machine work resting on machine work.\n3. People: the page of a whole that has parts offers responses on the parts and says why, and a reason\n   a person already put on a whole moves only by a proposal to that person (*nothing a person wrote or\n   confirmed moves without a proposal*). A challenge to how the parts combine (*these two cannot both\n   hold*) is a link between the parts, not a reason on the whole.\n\n**Proposal two (ledger AL16): a claim cut again matches before it mints.** The ruled entry order (AL9:\ncut, then sameness, then link) applied to cutting again. Before a fresh part is written it is matched\nagainst the parts already there, word for word first and then by the keeper:\n\n- a fresh part the same as an old one keeps the old part, with every reason and vote on it, and adds\n  the fresh cut's readings to its link (`readings`, the older ones kept in `earlier_readings`, so the\n  record says the new version agrees with it);\n- an old part the fresh cut splits finer stays and gets the finer parts under it, and its own reasons\n  then move by proposal one;\n- an old part the fresh cut drops is retired as today when nothing hangs on it, and held for a person\n  when something does (*the new cut drops this part, which carries three reasons*);\n- a fresh part with no match is minted as today.\n\nUntil this is built, a fresh cut holds any claim one of whose parts carries a reason, as it already\nholds a part a person kept. The invariant, checkable over the graph beside `out_of_step`: no reason\npoints at a retired part without either a successor or an open question to a person.\n\n**Proposal two was ruled and built the same day** (his word, on it as put in chat: *\"Blessed, build it please.\"*). `proposed_split.recut` now matches before it mints (`match_parts`): the same words first, then the keeper, only a confident *same* counting, at most twelve readings a claim and only between parts that share a third of their words. A kept part's link records the version that agreed (`matched_version`, `match`) and rotates its readings; the version that cut it stays named on the link, so a person's review still sorts it by its cutter. An old part no fresh part says again is checked for what hangs on it (`what_hangs_on`: a reason either way, a part under it, another claim it is a part of, a vote, named as a vote and never by who cast it, or a person's keep, rewording or confirmation); if anything does, nothing is written and the fresh cut waits on the claim (`split_recut_held`) until a person applies it (`apply_held`) or keeps the split as it is (`dismiss_held`), each recorded with whose word it was. `scripts/recut_splits.py --held` lists them, and a run no longer asks again after a hold, which until then it did three times for a part a person had kept. Undo puts a kept part's readings back. **One clause was built differently:** an old part the fresh cut splits finer is held for a person when something hangs on it, and retired as before when nothing does; it is not nested automatically, since a nest needs a coverage check of its own and is machine work resting on machine work, which T1 caps by measurement (R12, ruled later that day, lets a later reading retire a link, not build a nest).\n\n**His answer on proposal one, the same day:** *\"I feel about this roughly how I feel about the conditionals (and may-hedging) yesterday - deleting (or at least fully hiding and preventing reasoning about) the whole/original claim is justifiable since the meaningful claims that could be extracted were extracted (well, should be) and so hiding or even deleting the whole wouldn't erase the meaningful content and reasons that the author expressed, only the form they originally, for a different medium, chose to string them together and express them as. Perhaps if the words were spoken they'd have been given extra gravitas or emphasis on some sentences or words, but that's lost in translation and similarly for this I say good riddance. Because that's not what Deliberus is focused on, it's focused on structured intersubjective scrutiny and careful reasoning. Except we might want to keep the original claim as telemetry more or less and to be able to recut it, if at all necessary. But provenance is ofc interesting - perhaps all cut/resulting claims (and the original) should bear/inherit knowledge about who the author/s were (for different keeper-clustered/merged versions of the \"same\" claim) and you should be able to trace it back to its source/s...\"* So the whole is to be telemetry, hidden and closed to reasoning, kept as the text a fresh cut reads; provenance already passes down a cut (author and source on every part), and across a sameness group it is the group's, unbuilt. What it leaves open: a link that uses the whole as a premise because it needs the parts together loses its source when the whole hides (33 links run from the 13 wholes, 3 written by people). Recorded in ledger AL15.\n\n**The joint premise, a session's recommendation the same day (unruled).** His question: *\"One question stays open: what happens to links that need the parts together. — Explain and recommend something\"*. A whole serves as the premise of a link when the reason is the parts together (*today is Tuesday, and this gate is open every Tuesday*, read as one reason for *the gate is open*; ledger AL3). There are three ways to keep such a link once the whole stops being a place to argue, and the strength code was run on each (a whole of two parts, a link the source asserted):\n\n| | Both parts at 0.9 | One part refuted (0.9 and 0.2) |\n|---|---|---|\n| The whole keeps serving as the premise, scored as the code scores it now | the claim moves to 0.500: the whole sits at 0.507 (the up-transmission whisper), so the link passes failure on and never success | nothing reaches the claim (right) |\n| The whole keeps serving, read as its weakest part when its cut passed the coverage check | 0.519, the same as one reason at 0.9 | nothing (right) |\n| Each link rewritten from each part | 0.569 with two parts, 0.632 with three: one joint argument counted as two or three | 0.519: the surviving part carries an argument that needed both |\n\nThe recommendation is the second row. The whole stays as the premise and is closed to reasons about what it says, so the person-written links keep exactly what their writers drew. A whole whose cut passed the coverage check says nothing its parts do not say together, which is the condition under which the coverage-gated lift (`weakest-link-arithmetic-and-the-merge-hunch.md` § 7, filed in `TODO.md`, never ruled) lets a whole inherit its weakest part upward. Every machine cut carries that check's verdict on its links (`r.coverage`): on the live graph that day 12 of the 13 cut claims read *clean* and one *dropped*, which a cut may still be written with (only invention and an unasserted part are held), so the lift would reach the twelve. A whole a person split by hand has no verdict, so the lift would need the check run on it first. Not measured: how many of the 33 links need all the parts rather than one; a link that needs only some would be narrowed to them by a later reading, unbuilt. The finding reaches ledger AL3 too, whose clause *the link carries what they jointly establish* holds only under the lift.\n\n**Considered and rejected.** Never cutting again where a part carries a reason: safe, and it fixes the\nolder splitter's mistakes in place exactly where people engaged, which is frame lock-in at the grain of\na part. Copying reasons onto the fresh parts: a reason was written against the text of the part it\nnames, so a copy is an assertion nobody made, and it would count twice (§ 2d, what supersession is).\nLetting both stand and relying on the sameness group to carry the strength (designed, stage four of\nthe sameness ladder, unbuilt): the right answer for a person's rewording, which is a new use; for a\nfresh machine cut that says what an old part said, the second node adds no use, so not minting it is\ncleaner than grouping it afterwards.\n\n**What it costs and what it does not settle.** One keeper call for each old and fresh part that look\nalike and are not word for word the same, a few per claim; under R8's brake a test reads three or\nfour. Two parts cut from one sentence by two versions of one pass are the easiest sameness case there\nis, and the keeper's reading of them is unmeasured. A wrong *same* carries a reason across a gap in\nmeaning; since every match is a named reading on its link (`proving-a-pass-reliable.md` § 11), the\nmatches one version made can be found and undone together. Proposal one is open; proposal two was ruled and built the same day (below).\n\n### What he argued for, 15 to 17 October, and where each stands\n\nIn his words, trimmed, in the order he said them; the state is as of 17 October at night.\n\n| His words | Where it stands |\n|---|---|\n| *\"read everything that is remotely relevant to the first chapter in the corpus … and read the claude.md of Deliberus first of all\"* | Done by that session; its memo is local-only. Two things it said were worth carrying, now ledger AL11 (understanding shown by use on a case the instructions never saw) and AL12 (what a correction reaches downstream when the passage names neither example nor ruling); actions in `TODO.md` |\n| *\"Anything here that wasn't taken up by Claude that you think is clearly worth pondering for Deliberus?\"* | The same two, AL11 and AL12; the rest it found already taken up (the shared manipulable case in the rehearsal, the question as working surface in the graph views) |\n| *\"the thinking that will go on inside Deliberus and through it may be inspired by that idea\"* (the forecasting post) | § 2e and ledger AL13; a worked example is a candidate, nothing built |\n| *\"add a todo to render diagrams in Lins of the Deliberus [architecture/codebase] and [ontology …] and [AI microtasks - including annotations about their prompting …]\"* | Built as the ten commentary views; whether the microtasks view names every pass the record of machine judgments knows is unchecked. Open, `TODO.md` |\n| *\"whether a reason is acceptable/valid/relevant depends on what concept sense is in use in the claim(s) in focus\"* | Ledger Q39 and AL8, open |\n| *\"It should be split by the splitter (this has been proposed at an earlier date!) so that the reasons are separated from the main claim\"* | T9, ruled and built 2026-10-17 (*\"Build T9 then pls\"*): the reasons-off reading runs before the splitter, which is itself not changed until his word |\n| *\"Could this be captured in different senses that apply in different contexts? It's all in the name of clarity.\"* | § 2 and N4 (a sense is somebody's difference from a shown reference); open |\n| *\"I want to create ways for humans to agree more and more over time … much closer to a hive mind in a positive sense\"* | Ruled on the public thread as AL7, 17 October: agreement welcomed as an outcome, never the measure |\n| *\"we need to 'deduplicate' all features of the optimal ontology … There needs to not be multiple points of entry for reasoning/attack/support around a claim.\"* | AL1, open |\n| *\"A negated attack could be offered as support - must be deduplicated somehow, perhaps by striving to have as few negations as possible in all claims?\"* | AL2, open; § 4 answers half (removing negation words is no normalisation); whether the decontextualiser should prefer positive phrasing at mint is the open half |\n| *\"Could the reasoning/challenging of the relevance-claim … be a duplicate point of entry such that it could eat up some claims that might better belong as critique of the reason-claim itself?\"* | AL6 (the sorting question) and AL8 (relevance read from three grounds), open |\n| *\"running examples by this structure and mapping them out visually in RELATION to the ontology and possible VARIATIONS of the ontology is crucial\"* | Built: ten views, then six example graphs of real claims; his walkthrough is next |\n| *\"thinking about what is amenable to machine microtasks and ofc thinking about human interaction while not getting DISTRACTED too early on by thinking about how all this can be abused\"* | § 6 and ledger AK1 (payoff removal last), open |\n| *\"Make sure key docs are up to date and congruent\"* | Done twice, 17 October |\n| *\"if some says 'X because A and B' then it's just a boon if someone else comes up with C … And if someone shoots down A, then so be it … 'X because A or B' is just unnecessary hedging also, no?\"* | AL3, open, which now carries his *or* sentence and the one thing missing (a source that asserts neither must not have both written as asserted) |\n| *\"Well then a reason should argue that inside the text of its claim I guess?\"* | AL3 (the composite reason) and AL4 (the rule under the link), open |\n| *\"'gate is open every Tuesday' is an inference-rule, right? Which is not an independent reason\"* | His correction, recorded in § 2a; drawn in the example graphs 02 to 04 |\n| *\"such things as what today's date and the current time is (in different timezones) should always be directly checkable in Deliberus … evaluate truthiness of such claims as A automagically\"* | AL5, open, unbuilt |\n| *\"Hasn't this been discussed ad nauseam in earlier docs about parts vs reasons and corroborative vs [whatever] … And also this relates to the chain within a link, which is vertically explicated to the 'floor' of one inference/step, right?\"* | Yes: support semantics, the retrofit § 1, the chain note § 10.3 and AG14; the lineage is in § 2a |\n| *\"Perhaps the object-model.md doc should more clearly say that it's contents are legacy? … harvest … Are these [sketches] saved, in fact? If not, can we harvest them?\"* | Done: the harvest note, the six drawings recovered and held privately on his word |\n| *\"Maybe I can remember which sketches these refer to, can you tell me about them briefly?\"* | The three unlocated drawings are described in `docs/sketches.md`; his memory of them is the open loop |\n| *\"Or are claims never rephrased after created? … split … then … sameness-clustered/merged, after splitting, not before\"* | § 2d; AL9 and AL10, both ruled and built 2026-10-17 |\n| *\"Perhaps just like we have Sources as an entity in the database (right?) we could have Theories or Hypotheses, which are linked to the claims they make\"* | § 2f and ledger AL14: a theory is a party in reported speech, its claims read from attributions; open, since his *adopt* came before he learned it was Claude's take and not his node; the dry count finds at most about 25 claims |\n\n## 3. The reason and the relevance claim: which objection goes where?\n\nUse the reading in which warmth is a defeasible reason for comfort:\n\n- C: the hall will be comfortable for these visitors at this event.\n- R: the hall will be warm at that event.\n- L: granted R, it bears positively on C in this situation.\n\n| Contribution | First target | Why |\n|---|---|---|\n| The heating is broken and the hall will be cold | R, with the prediction's own argument | It disputes the premise, rather than granting it |\n| These visitors find that temperature oppressive | The applicability of R → C, with a claim about these visitors | It can grant the stated temperature and dispute its positive bearing |\n| Here *comfortable* means accessible; warmth does not establish access | The interpretation and then the connection | It does not show that the hall is not warm |\n| The hall will be uncomfortable because the seating causes pain | C through a distinct argument | It offers grounds against the conclusion |\n| The thermometer is faulty | The measurement-to-temperature inference | It need not establish coldness or directly attack every conclusion about comfort |\n\nThe practical sorting question is: **if the reason were granted under the agreed interpretation,\nwould this objection still stand?** If yes, there is a connection-level issue. If the objection instead\ndenies that reason, its first home is the reason. This is a diagnostic for clear cases, not a complete\nclassifier: mixed objections need splitting, and changing an interpretation can change both questions.\n\nThe implementation has a real overlap worth inspecting. `relevance.py` describes acceptability as the\nreason's standing adjusted by the particular link's premise-question answers. `graph/qbaf.py` and\n`graph/badge_inputs.py` implement this distinction from connection questions. The corpus already flags\nthat a premise answer can then weaken a supporter on one link only\n([Session 41](session41-when-a-reason-fails.md), § 1). A factual challenge to a shared reason and a\nlink-local adjustment may thus represent the same underlying issue through different paths.\n\nThree alternatives are drawn for comparison, not silently adopted:\n\n- **A — current storage.** Claims and relationship properties, questions remembering their endpoints,\n  a relevance claim whose id the relationship stores, plus premise and connection answers.\n- **B — one question per scoped issue.** Premise questions resolve to the reason's issue; connection\n  questions resolve to the application of that reason to that target. Interfaces point to these homes.\n  Exact complements may share an issue with opposite answers. New identity and routing work is needed.\n- **C — explicit inference applications.** An argument application joins a premise set, a rule and a\n  conclusion, and has addressable questions of its own. The relevance claim is about that application,\n  or a view of its warrant, rather than an accidental second copy of it. This makes bundles explicit\n  but needs decisions about rule identity, joint premises and the relationship to existing link fields.\n\nB and C can compose. A graph node for an application is a representational option, not proof that\nevery epistemic distinction needs a new database label. The same tests should be run against an\nimplementation using ordinary claims plus explicit references.\n\nThis distinction has a close formal counterpart: ASPIC+ distinguishes attacks on premises, defeasible\ninferences and conclusions. It gives useful targets to compare, but does not decide Deliberus's numeric\ncalculus or resolve its duplicate paths. [Modgil and Prakken's tutorial](https://journals.sagepub.com/doi/full/10.1080/19462166.2013.869766).\n\n## 4. Negation: reducing duplication without changing the question\n\n**Removing negation words is not a safe normalization.** “Not all seats are occupied” differs from\n“all seats are unoccupied”; “not hot” differs from “cold”; “not established that P” differs from\n“established that not P”. Scope, modality and thresholds belong to the meaning that matching preserves.\nMacCartney and Manning distinguish exhaustive negation from non-exhaustive exclusion; the latter leaves\nroom for neither alternative. Their inference treatment also preserves the surrounding operators.\n[Natural logic and natural language inference](https://nlp.stanford.edu/~manning/papers/natlog-cm.pdf), § 2.\n\nThere are at least three different examples behind “a negated attack is support”:\n\n1. **A demonstrated complement.** Within a stipulated question with exactly two exhaustive answers,\n   establishing not-not-P establishes P. A system can show the positive wording while retaining the\n   original expression and the equivalence that licensed it. It should not award a second independent\n   evidential contribution for this redescription.\n2. **A defeated objection.** A witness says the hall is noisy; the witness turns out not to have visited.\n   That removes a reason to believe it is noisy. It does not establish quiet, let alone comfort.\n3. **Positive evidence that also answers an objection.** A suitable measurement supports quietness\n   and challenges a same-time, same-place noise claim. One measurement has two explained roles. The\n   paths must retain their common origin; drawing two paths does not create two measurements.\n\nThe second example is already connected to Q34, ruled for now: a failed reason drops out, and a negative\nconsequence needs a contribution explaining why. The converse is equally important for the design:\nan attack that fails does not automatically manufacture a new positive reason. The third is not solved\nby forbidding one of its legitimate roles. It needs identity, dependence and aggregation handled\ntogether. [Relevance and the two factors](relevance-and-the-two-factors.md) § 7 and\n[Retrofitting the reasons channel](retrofitting-the-reasons-channel.md) § 4 supply the existing work.\n\n### 4b. 20 October: the question again, beside the map sketch\n\nHe asked it again the day he drew the map view with supporters and attackers below the focused\nclaim ([../sketches.md](../sketches.md), the 20 October section):\n\n> *\"But this brought up a question for me again, the idea about having a single point of entry for\n> each type of interaction with a claim in the graph, so for example an attack could be a negated\n> support-claim, so shouldn't it simply be un-negated (the claim stripped of negations and certainly\n> any double negations, in an early extraction phase or other routine machine pass) and assigned to\n> the right \"pole\" (support/attack). Reading negations and especially double negations is more\n> cognitively taxing (for humans at least) and so maybe the ideal is to keep each claim as its\n> support-variant. But maybe there are exceptions and counterarguments to this approach, hehe, lil\n> bit meta-ironic here...\"*\n\nOne example, whole. The focused claim: *The town should build the new bridge.* Under it, an attack as\nits author wrote it: *The bridge is not affordable.* His proposal: store *The bridge is affordable*\nonce, set its pole to support, and record the author's sentence as a denial of it, so everything said\nabout affordability meets at one claim.\n\n**What holds.** One home for affordability is § 1's second row already: a claim and its exact\ndenial are two answers to one question, and a reader asking *is it affordable?* should arrive at one\nplace whichever answer a source gave. The reading cost is real: negated sentences take an extra\nstep to verify ([Clark and Chase 1972](https://doi.org/10.1016/0010-0285(72)90019-9), Cognitive\nPsychology 3, 472–517). And readers themselves un-negate exactly where an exact opposite exists: for\ntwo-valued predicates, *the door is not open* is held as *the door is closed* about 1.5 seconds after\nreading (Kaup, Lüdtke and Zwaan 2006, Journal of Pragmatics 38, 1033–1050, as Giora summarises it in\nthe same issue). The machine can do the same, in the same cases.\n\n**What breaks if the positive wording is stored rather than shown:**\n\n1. **A pole is relative to the claim in focus, so no claim has one support variant.** *Nuclear power\n   is dangerous* attacks *Sweden should build reactors* and supports *Sweden should phase out nuclear\n   power*. A claim used under both has no wording that supports everywhere; which way it points is\n   decided when a claim is focused, which is what the sketch's lower half draws.\n2. **Many negations have no exact positive** (*not hot* is not *cold*, above). And a common attack\n   shape, *there is no evidence that P*, denies P's support, not P: un-negated into *P* with a denial,\n   it would say its author asserted P false where they asserted it unshown.\n3. **Most double negations hedge.** *Not unlikely* and *not uncommon* are weaker than *likely* and\n   *common* (litotes); stripped, they claim more than their author did. They belong with the stance\n   rule (ledger AM6, a hedge is stance): the wording *likely*, with the weaker stance kept beside it.\n   A logical double negation (*it is not the case that the bridge is not affordable*) simplifies outright.\n4. **Whose words.** AA6 holds a claim to its author's words; rewriting an attack into its opponent's\n   positive wording puts the opponent's frame first.\n\n**The shape that keeps his aim, a session's recommendation, unruled:** the author's words stay\nstored as written; where an exact opposite exists, the claim and its denial are linked as exact\nnegations (a key the keeper's `opposes` cannot stand in for, since it covers contraries too) and share\none home; the display shows the positive wording of the pair, with the pole drawn relative to the\nfocused claim and the author's sentence one point away; where no exact opposite exists nothing is\nrewritten; hedging double negations go through the stance rule and logical ones are simplified. This\nanswers AL2's open half (should the decontextualiser prefer positive phrasing at mint) as *prefer it\non display, never in storage*. The display wording is one model call per pair, a pass keyed on pairs\nnot yet read, so it costs once per pair.\n\n## 5. What the diagrams need to let a person do\n\nThe Lins set relates architecture, current storage, proposed objects, worked claims, challenge targets,\nnegation cases and machine tasks. Cards carry their state, reasoning and source; links say what the\nrelationship means. The diagrams are authored design examples, not an extraction from live discussions.\nAn example illustrates a proposed behavior; it does not prove a classifier or a strength calculation.\n\nA walkthrough should let us make these changes and examine their consequences:\n\n| Change | What to inspect |\n|---|---|\n| Grant a premise while challenging its use | Does the premise remain available in other arguments? |\n| Refute a premise used in several arguments | Do its uses point back to one challenge, or need repeated objections? |\n| Change the sense of a term at one endpoint | Is the affected inference explained without changing other attributed uses? |\n| Add an inline reason | Is it separated from the conclusion and kept as an asserted part of the source utterance? |\n| Rephrase an objection as support for an opposite | Does the same scoped issue retain both expressions and avoid a second independent contribution? |\n| Replace an exact complement with a contrary | Does the system stop assuming there are only two possibilities? |\n| Reuse one observation under two arguments | Is shared origin visible all the way to the common conclusion? |\n| Add a new criterion for comfort | Can we distinguish a meaning proposal from evidence or a competing value judgment? |\n\nThe first implementation is a set of inspectable diagrams in the existing Lins viewer. It is not yet\nan executable alternate-ontology or strength simulator. Its cases state expected distinctions for a\nlater simulation; those expectations remain available for criticism.\n\nThe opening instructions and six-step walkthrough (`tools/lins/README.md` in the repository) accompany ten\nauthored views in `tools/lins/design_diagrams.py` (the tenth, `09_gate`, added 17 October). On 16 October their cards and connections were\nwritten and read back, both arrangements were checked through Lins, and the comfort inference was\ninspected visually. This verifies the views, not the proposed semantic judgments. Lins also needed a\nquery repair for lanes without progress counts: repeated zero columns require distinct aliases.\n\n## 6. Microtasks and human interaction\n\nThe inventory includes work outside extraction: sameness, sense assignment, reason routing, rewriting,\nrelevance, unwritten steps, review, corpus connections and consistency checking. Each has its own input,\noutput, prompt or design home, maturity and limitation. A deterministic operation is marked as such.\n\nPromising proposed questions are small and separable: identify the conclusion of an arguing utterance;\nidentify each reason span; preserve assertion versus supposition; propose the sense used at an endpoint;\nask whether an objection survives granting its premise; distinguish an exact complement from a contrary;\nfind an existing issue before creating one. These are candidate task definitions, not claims that a cheap\nmodel already answers them reliably. Jev's existing measurements bear on task shape and evaluation, not\na presumption that these new semantic tasks are solved.\n\nA person should be able to say “that is not true”, “even if true, it does not follow”, or “that is not\nwhat I mean” in ordinary language. The machine can propose the affected statement, inference or use,\nshow the interpretation and reuse an existing home where appropriate. The categories serve the\nconversation; they should not be an ontology exam before someone can contribute. Misrouting and\ngenuine multi-target contributions are cases to examine visually before choosing that interaction.\n\n## 7. What is decided, and what stays open\n\n**Requested direction:** build and use the diagrams; compare variations through detailed examples;\nconsider sense at both ends of positive and negative reasoning; separate inline reasons; reduce duplicate\nplaces where the same issue must be answered; pursue clarity and increasing agreement; establish these\nneeds before threat-model design. **Not selected by that direction:** a new sense schema, a canonical\npolarity for every claim, an inference-node implementation, the grouping of warm and quiet, or a numeric\nrule for shared evidence. Existing open rows T9, Q38 and Q39 are relevant; this note does not mark them\nwholesale settled or implement them in the claim graph. The note's own proposals are ledger Arc AL (AL1 to AL6,\nopen); the agreement sentence is AL7, ruled 17 October.\n\nThe diagram source and operator instructions live in `tools/lins/`; the deeper homes remain the linked\nnotes. [Object model](../object-model.md), [architecture reference](../architecture-reference.md),\n[convergence](../convergence.md) and [conceptual threads](../conceptual-threads.md) carry short pointers\nso a future reader can find this direction from the question they are working on.\n\n**See also:** [A part and a reason](a-part-and-a-reason.md) ·\n[Graph visualization](graph-visualization.md) · [Microtasks in broad daylight](microtasks-in-broad-daylight.md) ·\n[Core, periphery and the concept layer](core-periphery-and-the-concept-layer.md) ·\n[Store the uses, read the meaning](store-the-uses-read-the-meaning.md) ·\n[Claim or system](claim-or-system.md) · [The number in the world and the repertoire](the-number-in-the-world-and-the-repertoire.md).\n"}