{"path":"research/how-far-a-premise-is-from-its-claim.md","content":"# How far a premise may be from the claim it is linked to\n\n**Date**: 2026-10-07 · **Type**: a founder question, answered by reading the code, reading thirty\nlive links by hand, and two tests of whether a machine can say how far apart two claims are.\n**Status**: the measurements are done and are a model's reading, not a person's. Every proposal\nis Claude's and **unruled** (ledger Arc AD). Nothing here is built. It came ahead of the four open\nreviews because he asked.\n\n**Continued the same afternoon** in [what-a-link-rests-on.md](what-a-link-rests-on.md), on his\nrequest for a pass that handles this. Two more ways of having a machine judge a link were tried\nand failed, which makes four; asking for *one* further claim was measured (too few); and the\nproposals below gained seven more (AD6 to AD12), of which AD6 revises AD2.\n\n**Continued 2026-10-10** in [the-chain-inside-a-link.md](the-chain-inside-a-link.md), on his demand\nthat no leap be left too large: what a too-large leap is in terms that can be checked, and a count\nof what sits under each link today (1 of 1,569 has anything).\n\nHis question, after the linking step's part-of links were found to be premises stored backwards\nand were re-filed as support: ***\"Are we stringent about how far-removed a premise is from the\nclaim itself? Perhaps it's actually depending on a chain of inferences that's made up of more\ncomplex reasoning and a few different premises etc? Idk\"***\n\n## 0. The short version\n\n- **No.** The linking step (auto-connect, which links claims across texts) asks one thing: does the\n  first claim give *evidence, reasoning, or a premise* the second rests on. It writes the link when\n  its own rating is above 0.3. Nothing asks what lies between the two claims.\n- **It shows.** Thirty of its support links, read by hand: **8 are one step, 11 hold only with one\n  further claim a reader could dispute, 4 are several steps apart, 7 are not reasons in the\n  direction stored.** Of fifteen links written at extraction, inside one text, 11 are one step and\n  none is several steps.\n- **Where the far ones are.** Between two texts that argue with each other, no link in the sample\n  was several steps away. Between texts from different debates, none was one step.\n- **The step's own rating does not see it**: 0.67 on average for the one-step links and 0.68 for\n  the links that are not reasons.\n- **A machine cannot be asked how far.** Asked to list what else must be true, it listed two or\n  three further claims on nearly every link. Asked which standard pattern of argument the link\n  follows, it named one on 27 of 30, among them 6 of the 7 that are not reasons. A reader told that\n  a link exists finds a way for it to hold.\n- **What the machine did give is the content.** On nine of the ten links where the hand reading\n  had named one missing claim and the model answered, the model's list contained it. The same\n  missing claim came back on link after link, and it is the point the two texts disagree on.\n- **So the proposal is to store what lies between and not to grade the distance**: a link between\n  two texts is the linker's own argument, and it should carry the one claim it rests on, where a\n  reader can dispute it.\n\n## 1. What the step asks today\n\nRead in `deliberus/extraction/auto_connect.py` on 2026-10-07, after that morning's repair.\n\n- **Which pairs it reads.** For each claim of a new text, up to eight stored claims: the five\n  nearest by similarity of wording, plus pairs that share a contested word or have the same shape\n  of links around them.\n- **Sameness first.** A separate question asks whether the two are the same claim, one inside the\n  other, overlapping, opposed, or unrelated. *Unrelated* ends it.\n- **Then one question about the relation**, after a gate that asks whether there is a logical\n  relation at all: *\"Does A provide evidence, reasoning, or a premise that B rests on? → supports\"*,\n  with *\"A premise that the other claim depends on is `supports`, with the premise as the source.\"*\n- **What it stores**: the kind of link, a direction, a rating from 0 to 1 that the prompt calls\n  *strength of the relationship* and does not define, one sentence of explanation, and where the\n  pair came from. A link is written when the rating is above 0.3.\n- **What it does not store**: an argument pattern (the *scheme* that links written at extraction\n  carry, which is what gives a link its critical questions), or any reading of how much the first\n  claim bears on the second. Counted on the live graph that day: 622 live support links from this\n  step, none with a pattern. Of the 5,736 live support and attack links from all writers, none has\n  had its bearing read by a machine or argued by a person.\n\nSo the answer to the question as asked is plain: nothing in the step is strict about distance, and\nnothing records it.\n\n**What that costs today.** Nothing in the strength numbers. A link the machine guessed between\nclaims other people wrote is presumed to bear nothing until it is read or argued\n([relevance-and-the-two-factors.md](relevance-and-the-two-factors.md)), so these links move no\nclaim's strength. On a map they are drawn like any other support link, a far premise the same as a\nnear one.\n\n## 2. Thirty links, read by hand\n\n**The sample.** Thirty support links drawn with a fixed seed from the 569 the step has written\nbetween claims taken from the texts themselves, and fifteen from the 615 support links written at\nextraction inside one text. Each was read with both claims in full and the step's stored reason.\n\n**The four answers, fixed before reading.**\n\n| Answer | Meaning |\n|---|---|\n| one step | The first claim, with at most a commonplace nobody in the debate disputes, makes the second more likely |\n| one further claim | It does so only together with one more claim that a reader could dispute |\n| several steps | Two or more further claims or an intermediate conclusion lie between |\n| not a reason | No support in the direction stored: a restatement, an explanation, the wrong direction, or the first claim's author meant the opposite |\n\n**Who read.** Claude, in the session of 2026-10-07. That is a model's reading, from the same family\nas the model that wrote many of these links. No person has read them. The sample is small and\ndrawn once.\n\n| | Links | One step | One further claim | Several steps | Not a reason |\n|---|---|---|---|---|---|\n| The linking step, between texts | 30 | 8 | 11 | 4 | 7 |\n| Extraction, inside one text | 15 | 11 | 3 | 0 | 1 |\n\nOne example of each, from the thirty:\n\n- **One step.** *\"Claude's internal workspace … evolves over a single forward pass through Claude's\n  neural network, rather than through iterative loops over time as in the human brain\"* (the\n  paper), for *\"There is no recurrent activity … present in the AI language model Claude's internal\n  processing\"* (the reply). The paper's own statement is the reply's premise.\n- **One further claim.** *\"Anthropic researchers … report that there are several key differences\n  between that workspace and the global workspace model of consciousness\"* (the paper), for\n  *\"Anthropic's research findings … fall short of the criteria that global workspace theory …\n  typically requires\"* (the reply). It holds if the differences are among the things the theory\n  requires. Neither sentence says so.\n- **Several steps.** *\"Almost all artificial intelligence systems will develop inherent drives\n  toward … self-improvement\"* (a paper on the drives an AI would have), for *\"Recursive\n  self-improvement in artificial intelligence is a present-day phenomenon\"* (an essay on the\n  intelligence explosion). Between them: that today's systems are\n  among those with the drive, that a drive is enough to act on, and that what is observed today is\n  that.\n- **Not a reason.** *\"The relevant distinction between murder and state-sanctioned execution is\n  social and institutional rather than physical\"* (a defender of the death penalty), stored as\n  support for *\"The concept of justice is not furthered or advanced by the state taking a human\n  life via execution\"* (an opponent). Its author used it to defend executions.\n\n**Two things the table hides.**\n\n*The texts' relation predicts the distance.* Fifteen of the thirty join the two texts of one\nexchange (the machine-consciousness pair mapped the night before): 6 one step, 7 one further\nclaim, 2 not a reason, **none several steps**. The other fifteen join texts from different\ndebates: 2 one step, 4 one further claim, **4 several steps, 5 not a reason**. Cut by a fact the\ngraph already holds, whether the two texts attack each other at least twice: all 8 one-step links\nare inside such a pair, and all 4 several-step links are outside one. Of the step's 622 live\nsupport links, 386 are inside such a pair and 236 are not.\n\n*Why inside a text differs.* Inside one text the author wrote the premise for that conclusion.\nBetween two texts the step pairs a sentence written for one argument with a conclusion written for\nanother, and what would join them is what neither author wrote down.\n\n**The step's own rating does not see any of this.**\n\n| Read by hand as | Links | Mean of the step's rating |\n|---|---|---|\n| one step | 8 | 0.67 |\n| one further claim | 11 | 0.67 |\n| several steps | 4 | 0.62 |\n| not a reason | 7 | 0.68 |\n\n## 3. Two tries at having the machine say how far\n\nBoth ran after the hand reading and were set beside it. Same model as the step (`claude-sonnet-5`),\none call a link, nothing written to the graph.\n\n### 3a. Ask it to list what else must be true\n\nThe question: granting the first claim, is the second more likely (a gate, asked first); if so,\nwrite every further claim that must also be true and that a careful reader in the debate could\ndispute; then name the span from the length of the list.\n\n| | By hand | The model |\n|---|---|---|\n| The step's thirty links: one step / one further / several / not a reason | 8 / 11 / 4 / 7 | 0 / 1 / 25 / 4 |\n| The fifteen from extraction | 11 / 3 / 0 / 1 | 1 / 7 / 6 / 1 |\n| Same answer | | 5 of 30, and 2 of 15 |\n\nIt listed two further claims on 17 of the thirty, three on 7, four on 1. The links read by hand as\none step got two or three like the rest.\n\n**Why it fails as a measure.** The question has no place to stop. Any step of reasoning has\nsomething under it if one looks: that an example is typical, that two words name the same thing,\nthat a cause was enough. How many further claims one finds depends on how finely one cuts, which is\nthis project's own conviction that nothing is permanently atomic, met from the other side. Two\nreaders who cut at different depths will count differently on every link, and a count that depends\non the reader is not a property of the link.\n\n**What it is good for.** The lists themselves. Of the eleven links read by hand as needing one\nfurther claim, the model answered on ten. On nine its list contained the claim the hand reading had\nnamed, and on the tenth a neighbouring one (matched by the same reader, so this too is one model\nchecking another). It was the first claim listed on six of the nine; on the others an identity\nbetween the two texts' words came first (§ 4). The example above came back as *\"The 'key differences' the researchers identified are differences\nspecifically in the features or properties that GWT's criteria require.\"*\n\n### 3b. Ask the classifier the graph already has which pattern the link follows\n\nLinks written at extraction carry an argument pattern (example, cause to effect, analogy, and so\non), and a pattern brings its critical questions, which are a bounded list of what that kind of\nstep rests on. The pattern classifier has a no-fit answer. So the thought was: a link that is\nseveral steps, or no reason at all, should fit no single pattern.\n\nWritten down before it ran: *of the 11 links read as several steps or not a reason, no pattern fits\non at least 5; of the 8 read as one step, a pattern is named on at least 7.*\n\n| Read by hand as | Links | A pattern named | No pattern fits |\n|---|---|---|---|\n| one step | 8 | 7 | 1 |\n| one further claim | 11 | 10 | 1 |\n| several steps | 4 | 4 | 0 |\n| not a reason | 7 | 6 | 1 |\n\n**The first prediction failed: 1 of 11, where at least 5 were expected.** The second held. The\nclassifier named a pattern on 27 of the thirty. Its three no-fit answers were for pairs it read as\n*a plain restatement* (twice) and *a plain entailment* (once), which is closeness, the opposite of\nwhat was looked for.\n\n**What that shows.** The classifier is told the two claims are connected by support and asked what\nkind. A reader told that a link exists supplies a way for it to hold. The same shape has been\nmeasured here before on menus: an answer of *none* loses to the answers beside it\n([auto-connect-upgrade.md](auto-connect-upgrade.md) § 2c). So a pattern on a link is not evidence\nthat the link is one step, or that it is a reason at all, and the no-fit answer cannot be used as a\ncheck on the linking step's links.\n\n## 4. What the distance is made of\n\nRead across the model's lists for the thirty links, the further claims are of three kinds.\n\n**A bridge**: the claim that makes the first bear on the second. *The differences are among what\nthe theory requires.* *What held for machines in factories holds for AI across the economy.* The\ntradition of drawing arguments has a name for the place it fills. Two premises that only work\ntogether are *linked*, and Freeman's account ties each structure to the question a challenger would\nask; for the linked structure the question is *\"Why is that relevant to the claim you allege it to\nsupport?\"* The bridge is the answer to that question.\n\n**An intermediate conclusion**: a claim the first supports and the second rests on, so that the\nhonest drawing is a chain. Beardsley, in the 1950 textbook these diagrams come from, called a\nstatement that is both a conclusion and a reason for a further conclusion a *serial* argument. A direct link\nacross a chain hides the middle of it. Where the graph also holds the chain, the direct link repeats\nit: counted on the live graph, **74 of the step's 622 support links (12%) join two claims that are\nalso joined by a path of two support links**, against 25 of 2,685 (1%) for links written at\nextraction.\n\n**An identity between the two texts' words.** *The \"J-space\" of the paper is the \"mental\nworkspace\" of the reply.* The model listed a claim of this kind on six of the fifteen links\nbetween the consciousness texts. It is true and it is not an argument: it is the concept layer's\njob (the part of the graph that tracks contested words), and the map of that exchange stored the\ntwo texts' entries for one word as unrelated (ledger AB14).\n\n**The bridges repeat, and what repeats is the crux.** On four of the fifteen links between the two\nconsciousness texts the missing claim is the same one in different words: that what was found in\nthe model is the same kind of thing as the workspace of human conscious access. That is what the\npaper assumes and the reply denies. Two more repeat: that what was observed counts as thinking\nand not as passive correlation (three links), and that a single example is typical (four). The\nnight before, a different pass had read the two texts whole\nand proposed the premises neither wrote down; its second was that markers of access *\"are treated\nas relevant evidence bearing on phenomenal consciousness\"*\n([the probe's output](mapping-the-machine-consciousness-debate/unwritten-premises-between-the-pair.txt)).\nTwo passes, one from the whole texts and one link by link, arrive in the same place. An earlier\nfinding of this project is that the crux of a debate is usually left unsaid by at least one\nside ([the-load-bearing-unsaid.md](the-load-bearing-unsaid.md)); the far links are\nwhere the unsaid part sits.\n\n## 5. What the graph already has for this\n\n- **A link's bearing can be argued, as a claim.** Since the two-factor ruling of 2026-09-28, a link\n  moves its target by the supporting claim's own strength times the link's *relevance*, and a\n  person can argue the relevance: the argument is minted as an ordinary claim (*\"S bears on T: …\"*)\n  and the link's relevance follows that claim's standing (`POST\n  /claims/{id}/links/{source}/relevance`). **No link on the live graph has one.**\n- **A machine can read a link's bearing** on a five-step scale, from *settles* to *marginal*\n  (`relevance_judge.py`). No support or attack link on the live graph carries such a reading; the\n  190 readings stored under that name are on part-of links, from the census of 2026-09-22.\n- **A pass that proposes unwritten premises between two texts** (`cross_source_premises.py`), run\n  by an operator, one call a pair, proposing only.\n- **Potentials.** A critical question is not stored as a node: it is rendered from the link and\n  becomes a claim when somebody answers it\n  ([latent-scaffolding-and-the-type-token-split.md](latent-scaffolding-and-the-type-token-split.md)).\n\nSo the place for a bridge exists. What is missing is that nothing puts one there.\n\n## 6. Proposals (all unruled, ledger Arc AD)\n\n**AD1. A link between two texts is an argument neither author wrote, so it carries its bridge: the\none claim that would make the first bear on the second.** Inside a text the author made the step\nand it stands as written until somebody asks why. Between texts the linking step made it. An\nargument owes its premise, and a machine's argument owes it at the start. This does not forbid far\nlinks. It asks that what a far link rests on be on the page.\n\n**AD2. The bridge is stored on the link as a potential, in the linker's own call, and becomes a\nclaim when somebody answers it.** One more field in a call already being made: *the one further\nclaim this rests on, or none.* Shown with the link as *holds if: …*, with two moves, *dispute it*\nand *it is stated here*. Minted, it is the link's relevance claim, so the ruled machinery takes\nover and the link's bearing follows the bridge's standing. Stored as text and not as a node, it\nadds no claims to the graph until a person touches it. Measured once: asked for\neverything, the model put the bridge first on six links of nine and an identity between the two\ntexts' words first on the rest, so the field asks for the bridge and sends identities elsewhere\n(AD4). ⚠ *Unmeasured* when this was written: how it answers when asked for one. Measured the same\nafternoon: under two wordings it named the same claim on 6 links of 10, and on the four links\nwhose missing claim is the crux it named that claim once. One is too few; AD6 asks for up to\nthree ([what-a-link-rests-on.md](what-a-link-rests-on.md) § 2d).\n\n**AD3. A source records which debate it was mapped for, and across debates the step proposes and\ndoes not draw support.** In the sample no link between texts from different debates was one step\n(0 of 9 by the opposed-pair test), and 236 of the 622 live support links are in that class. The\ntest used here, two attack links between the texts, is a stand-in: two texts on the same side of\none debate fail it. Hence the first half. Low confidence: nine links.\n\n**AD4. Bridges that repeat are proposed as the debate's unwritten premises, with a count of the\nlinks resting on each.** The link-by-link lists and the whole-pair probe are two sources for the\nsame proposals, and the count says how much of the map hangs on each. Identities between the two\ntexts' words go to the concept layer and are not bridges.\n\n**AD5. A pattern on a link is no evidence that the link is a reason, and a machine told that a\nlink exists cannot be asked to check it.** A finding more than a proposal (§ 3). It bears on any\nlater plan to let the pattern classifier's no-fit answer act as a guard, and on how a link's check\nmust be posed: the reader is not told a link exists, and the answer *no relation* is asked first.\n\n**Considered, and not proposed.**\n\n- *A grade on every link (one step, one further claim, several steps).* The model and the hand\n  reading agreed on 5 of 30. A field that two readers fill differently records the reader.\n- *Refusing every link that is not one step.* It would have dropped the eleven links that need one\n  further claim, and those bridges are where the crux was found.\n- *Leaving it to the machine reading of relevance.* A far link would get *weak*, and so would a\n  near link that is simply a weak reason. One number cannot say which, and the number is not\n  something a reader can dispute.\n\n## 7. What this does not say\n\n- **That the linking step's links are mostly wrong.** 19 of 30 are a reason at one step or one\n  bridge away, by one model's reading.\n- **That within-text links are sound.** Eleven of fifteen read as one step, in a sample of fifteen.\n  The one read as not a reason joins two premises of a third claim.\n- **That distance is unreal.** A chain is longer than a step. What failed is counting it by asking.\n- **How a person would read the same thirty.** Under the project's own order of work a person reads\n  a machine pass before anyone proves how often it is wrong. This reading is a machine's, and it\n  belongs with the finding.\n\n## 8. What was read, and how far\n\nIn the corpus, in full this session: the linking step's code and prompt, the pattern classifier,\n`relevance.py` and `relevance_judge.py`, the relevance route in `api.py`, and the probe's output.\nBy heading and opening only:\n[linked-convergent-and-what-we-are-actually-asking.md](linked-convergent-and-what-we-are-actually-asking.md)\nand [the-load-bearing-unsaid.md](the-load-bearing-unsaid.md), which hold the theory of linked and\nconvergent support and of the unsaid premise; neither is restated here.\n\nOnline on 2026-10-07, as search extracts and not whole papers. **Verify before citing outward.**\n\n| Source | Read |\n|---|---|\n| Reed, Walton & Macagno, *Argument diagramming in logic, law and artificial intelligence* (Knowledge Engineering Review, 2007) | extracts: an arrow in a diagram stands for an inference; Beardsley's serial argument (*Practical Logic*, 1950, p. 19); that in Beardsley's diagrams the passage from reason to conclusion is itself never supported or evaluated; unstated premises drawn as dashed boxes |\n| A workshop paper on Freeman's approach to argument structure ([CEUR vol. 1811](https://ceur-ws.org/Vol-1811/paper3.pdf)) | extracts: the question behind each structure, quoted from Freeman's 2011 book (pp. 13–16); the book itself not opened |\n| Conway, *On the Distinction between Convergent and Linked Arguments* (Informal Logic) | extracts: Thomas as the source of the distinction and the criterion by relevance |\n\n**The scripts and their outputs** are kept beside this document: the\n[sample with both claims in full](how-far-a-premise-is-from-its-claim/premise_distance_sample.json),\nthe [reading by hand and the first test](how-far-a-premise-is-from-its-claim/premise_distance_trial.py)\nwith [its answers](how-far-a-premise-is-from-its-claim/premise_distance_trial.json), the\n[second test](how-far-a-premise-is-from-its-claim/premise_distance_schemes.py) with\n[its answers](how-far-a-premise-is-from-its-claim/premise_distance_schemes.json), the\n[counts](how-far-a-premise-is-from-its-claim/premise_distance.py), and\n[what each run printed](how-far-a-premise-is-from-its-claim/what-the-runs-printed.txt).\n\n## Cross-references\n\n- [auto-connect-upgrade.md](auto-connect-upgrade.md) § 2c (the linking step, its repair, its cost)\n- [relevance-and-the-two-factors.md](relevance-and-the-two-factors.md) (a link's bearing as the\n  second factor, and how it is argued)\n- [mapping-the-machine-consciousness-debate.md](mapping-the-machine-consciousness-debate.md) § 5d\n  (the probe for unwritten premises)\n- [weakest-link-arithmetic-and-the-merge-hunch.md](weakest-link-arithmetic-and-the-merge-hunch.md)\n  § 1d and § 8 (premises that only work together)\n- [self-similar-decomposition-and-claim-ontology.md](self-similar-decomposition-and-claim-ontology.md)\n  (nothing is permanently atomic)\n- [what-a-link-rests-on.md](what-a-link-rests-on.md) (the same afternoon: the pass he asked for,\n  two more failed ways of judging a link, and what recurs across links)\n- ledger: [session26-open-rulings.md](session26-open-rulings.md) Arc AD\n"}