{"path":"research/what-a-link-rests-on.md","content":"# What a link rests on: naming it as a standing property of the graph\n\n**Date**: 2026-10-07 · **Type**: a founder request, answered by reading the corpus and the code,\ntwo outside readings, five counts that need no model, and two paid trials with predictions written\nfirst. **Status**: every proposal is Claude's and **unruled** (ledger AD6 to AD12). Nothing is\nbuilt and nothing was written to the graph. It came ahead of the four open reviews because he\nasked.\n\n> ⚠ **Revised the same evening.** He heard these proposals and answered: *\"I think merely naming what\n> the links consist of would be weak sauce. I want to have all the implicit reasoning structure and\n> premises be explicated and exposed and scaffolded automagically. To the extent that process can be\n> trusted of course.\"* [writing-out-the-unwritten.md](writing-out-the-unwritten.md) is the answer: a\n> link is written out in one form and the writing is checked, tried on these same thirty links.\n> **What it replaces here**: § 4.1's *up to three further claims* (now one rule, with names and facts\n> as their own parts); § 4.8, which left an author's own steps unwritten; § 4.9 (the reading is better\n> done on the written-out steps); and the reading in § 2e that the joined names are unwritten, when\n> the paper's own claims state the joining. **What stands**: no machine reading of a link may write,\n> hold or weigh it (§ 4.2, § 4.7), names are settled once for a pair of texts (§ 4.5), and what recurs\n> is what is offered (§ 4.6).\n\nHis words, on reading the morning's finding that links between two texts often hold only with a\nfurther claim nobody stated ([how-far-a-premise-is-from-its-claim.md](how-far-a-premise-is-from-its-claim.md)):\n***\"This needs to be handled/explicated/scaffolded by some pass. Maybe a new pass that's more like a\n'standing property of the graph' in some ingenious way. Think deeply about what's robust and wise\nhere.\"***\n\n## 0. The short version\n\n- **The question *is this a reason for that* has no steady answer until the step says what it\n  rests on.** For any two related claims some further claim makes the step hold, so a reader\n  answers according to which further claim they supply, or whether they supply one. People do this\n  too: annotators agree only moderately on whether a reason implies a claim.\n- **Measured today, on the same thirty links, one model**: asked one way it refused 4, asked\n  another way it refused 20, and eighteen links changed side. A third question, put to the step\n  once its further claim was written in, changed with the order of the two premises on 14 of 44\n  and called the stated premise idle on most links, the sound ones included.\n- **Four ways of having a machine judge a link have now failed here in one day**, three of them\n  against a prediction written first. So the pass that is robust is the one that **never judges a\n  link**.\n- **What a machine does give is content.** It names what a step would rest on. One naming is not\n  steady either (the same claim under two wordings on 6 links of 10). **What recurs across many\n  links is the steadier thing**: in two separate askings the commonest thing under the links\n  between two texts was the same two names taken to be one thing (under twelve of fifteen links\n  in the second), and the claim the two texts disagree on came back under several.\n- **So the design is: name under every link, judge none, and read what recurs.** The standing\n  property is that no link a machine draws between two texts lacks a statement of what it would\n  rest on. A link without one owes one, and the debt is a query.\n- **And the safeguard is to remove the payoff, since no detector held.** A link a machine guessed\n  is a proposal. It is drawn as one, and it weighs nothing, until something outside the machine\n  gives its further claim standing: a text that states it, or a person who argues for it.\n- **A person's reading is owed.** The thirty links now carry three machine readings each that\n  disagree with one another. They are the sitting.\n\n## 1. Why the morning's question had no answer\n\nThe morning's note found that the linking step (the step that links claims across texts) is not\nstrict about how far a premise is from its claim, and that a machine cannot be asked how far. This\nnote asks why, and the answer is in the object.\n\n**A link from one claim to another says less than it seems to.** *A is a reason for B* is true\nrelative to something further: a rule, a classification, an identity. Informal logic has said so\nfor forty years, and every source the helper read that takes up the question agrees on one\npoint: **some added premise always makes a step valid**, at the least the sentence *if A then B*, or B itself. So that a\nfurther claim can be written under a link is no evidence the link is good\n([the helper's report](what-a-link-rests-on/outside-reading-argumentation-theory.md), question 1).\n\n**So two readers who disagree about a link are usually supplying different further claims.** A\nlenient reader asks whether some plausible claim would carry the step, and nearly always finds\none. A strict reader asks whether the first claim counts with nothing added, and nearly always\nfinds it does not. Neither is reading the link. Each is reading the link plus something they\nbrought.\n\n**People show the same spread.** In the corpus built for the best-known task on unstated premises,\nannotators agreed at a kappa of .42 on whether a reason implies its claim, and its builders gave\nup asking people to fill gaps freely because the answers varied too much. Asked to bridge two\nclaims with no constraint, three annotators wrote 2.7 premises on average, shared about a third of\ntheir words, and declined to bridge between 1% and 15% of pairs, depending on the annotator. Where\na single step is given whole, with its premises (steps of school science, in that study),\nagreement on whether it follows was .72\n([the helper's report](what-a-link-rests-on/outside-reading-machines-naming-premises.md), headline\nnumbers; a helper model read those papers, to the depths it states).\n\n**The formal traditions build the same point into their objects.** The interchange format for\narguments forbids an edge from one statement straight to another: every connection passes through\na node that says what kind of step it is. Structured argumentation names each defeasible rule so\nthat the rule itself can be attacked. Wigmore's charts, by contrast, state no rule under a step and\nchart the doubts about it. None of them treats a bare arrow between two statements as a finished\nobject.\n\n## 2. What was measured today\n\nEverything below is on the forty-five links read by hand in the morning: thirty the linking step\nwrote between two texts (by hand: 8 one step, 11 resting on one further claim, 4 several steps, 7\nnot reasons) and fifteen written at extraction inside one text. **The hand reading is Claude's, a\nmodel of the same family as the readers below. No person has read these links.** Scripts, answers\nand both sets of predictions are in [the folder beside this note](what-a-link-rests-on/).\n\n### 2a. Five counts that need no model\n\n| Count | Result | Use |\n|---|---|---|\n| How close a named further claim lies to its link's conclusion (cosine) | the same range for links read as one step (0.66 to 0.80) and as not reasons (0.55 to 0.79) | none: a claim that merely restates the conclusion cannot be told by arithmetic |\n| Whether the stored claim nearest to a named further claim is one the two texts fight over | 7 of 32, against 37 of the two texts' 225 claims (16%) | none |\n| The Rabbit Rule of argument-mapping teaching, counted: the share of the conclusion's content words the premise never uses | medians 0.74 to 0.82 for all four hand grades | none: every decontextualised claim rewords everything |\n| Links that repeat a path the graph already holds (two or three support links between the same two claims) | **191 of the linking step's 622 support links (31%)**, against 38 of 2,719 for every other writer | real, for double counting; it does not track the hand reading of distance (4 of 8, 4 of 11, 1 of 4, 1 of 7) |\n| Named claims that say two names are one thing, grouped by wording | one group spanning 6 of the 15 links between the consciousness texts at cosine 0.80 | **works**; substantive claims blur into one topic-wide group below 0.75 |\n\nFour of five gave no signal. The one that works counts sentences that are nearly the same\nsentence.\n\n### 2b. First trial: a reader told that a no is welcome\n\nOne call a link. The reader was told somebody proposes the link and that it is not asked to defend\nit, was shown how each claim's own text uses it (the claims it supports and rests on there, taken\nfrom the graph; nine of the thirty had none), and was asked a gate first: granting the first\nclaim, is the second more likely.\n\n| | Morning's question | This question |\n|---|---|---|\n| Links refused, of thirty | 4 | **20** |\n| Of the 7 read by hand as not reasons | 2 | 6 |\n| Of the 19 read by hand as reasons | 1 | 10 |\n| Of the 4 read as several steps | 1 | 4 |\n\n**Eighteen of the thirty changed side between the two questions.** Three were refused by both.\nNine were let through by both, and all nine are reasons by the hand reading (3 one step, 6 with\none further claim).\n\nIts refusals read well. Several corrected the hand reading: three links graded *one step* were\nrefused as two wordings of one fact, which is the sameness question and not support. The link from\na defender of the death penalty to an opponent's conclusion was refused as counting the opposite\nway for its own author. It also refused links whose further claim is plain and worth having, where\nthe morning's finding was that those are where a debate's crux sits. **A refusal ended the answer,\nso no further claim was written under twenty links.**\n\nOf the predictions written before it ran: the gate would refuse at least 4 of the 7 not-reasons\n(held, 6) and at most 2 of the 23 others (**failed, 14**); it would call at least 5 of the 8\none-step links direct (**failed, 0**: where it let a link through it always wrote a further claim,\nand inside one text it wrote one under 11 of 11); its further claim would be the one the hand\nreading had named on at least 8 of the 11 links graded that way (**failed**: it let only 6 of them\nthrough); it would point at a line of the texts at least 3 times (held, 3: one further claim the\nhand reading had called unsaid is stated by the paper itself). Two more held by the same reader's\njudgement: no further claim was a bare statement that two names are one thing, and none restated\nits conclusion.\n\n### 2c. Second trial: name always, then check the completed step\n\nTwo calls. The first always writes what the step would rest on, whatever the reader thinks of the\nlink, and answers separately whether the two claims say the same thing and whether the step runs\nthe other way. The second is shown only an argument, two premises and a conclusion, with nothing\nto say that a link was proposed or which premise a machine wrote. It is asked whether the premises\ntogether count, whether each is idle given the other, and whether one merely restates the\nconclusion. It ran twice, with the premises in opposite order. The three questions are the clauses\nof Hitchcock's definition of a relevant premise (2011): it suffices with some companion, the\ncompanion does not suffice without it, and the companion can be checked without assuming the\nconclusion.\n\n| What the check said | Result |\n|---|---|\n| The premises count together | 87 of 89 checks |\n| The same three answers in both orders | **30 of 44 links** |\n| The stated premise is idle, or the named claim restates the conclusion | in both orders on 31 of 44 links (21 of the thirty), in one order on 10 more, in neither on 3 |\n| A completed step (counts, premise not idle, nothing restated, both orders) | **2 of 45**; of the 19 reasons by hand, 1; of the 7 not-reasons, 1 |\n\nOf six predictions written before it ran, the four about the check failed: steadiness (30 of 44\nwhere at least 36 of 45 was predicted), separation (1 of 19 where at least 12), the idle premise\nas the mark of a link refused earlier (6 of 13 where at least 8), and agreement with the first\ntrial (0 of its 10 where at least 8). The one about the same-fact flag held (3 of the 5 links it\nnamed). The one about naming could not be scored as written: the morning's hand reading had\nrecorded the claim it named for six of the eleven links only (§ 2d).\n\n**Why the check fails.** It answers to how the named claim is worded. Told to write a general\nclaim, the reader mostly wrote one about the case in hand, and a claim about the case in hand\ncarries the premise inside it, so the premise looks idle. Counted on content words: where the\ncheck called the premise idle in both orders, the named claim repeats a median 32% of the\npremise's words; where in one order, 21%; where in neither (three links), 13%. Both links that\npassed have a named claim written as a general rule (*whatever holds the thoughts a system can\nreport holds them in the manner of short-term memory*), though others written that way did not\npass. Hitchcock, whose definition the check used, had himself given up redundancy of a premise as\na test in 1985. **The verdict again follows wording, this time the wording of the further\nclaim.**\n\n**What did hold, partly.** The flag for two wordings of one fact: of the four links the first\ntrial had refused on that ground, this one flagged two, and it flagged the link the hand reading\nhad called a restatement, and one more. That is the sameness question, which already has its own\nreader.\n\n### 2d. Naming, across three askings\n\n- **Asked for everything** (morning): two or three claims a link, the hand-named claim among them\n  on 9 of 10.\n- **Asked for one, twice** (the two trials): on the ten links named in both, the same claim 6\n  times and a different one 4 times, by the same reader's judgement. Where a step needs two\n  things, each asking picked one.\n- **On the four links whose missing claim is the point the two texts disagree on**, the single\n  answer was that claim once. The lists had it four times.\n- **A middle was offered on 27 of 45** when the question allowed one *only if the honest drawing\n  is a chain*, including 6 of the 8 links read by hand as one step.\n\nSo one claim is too few, an offered slot gets filled, and a single naming is one reader's choice\namong several true things. This is the human result again (2.7 premises, a third of the words\nshared).\n\n### 2e. What recurs\n\nCounted on the second trial's namings for the fifteen links between the two consciousness texts,\nwith no model:\n\n- **Twelve of fifteen name the same joining of names**: what the paper calls the *J-space* is what\n  the reply calls the *mental workspace*, or the *space to think*.\n  ⚠ *Corrected the same evening: this joining is not unwritten. The paper's own claims state it\n  (the J-space as the model's internal workspace, six times over); only the word* mental *is the\n  reply's. Looked for in the texts, 17 of 27 joinings of names stood in the paper's claims\n  ([writing-out-the-unwritten.md](writing-out-the-unwritten.md) § 8c).*\n- Three smaller groups of substantive claims come back under two to four links each: that a\n  feature found in two very different systems is a general solution; that the differences found\n  are ones the theory's criteria turn on; that the space found does what conscious access does.\n  The third is what the paper holds and the reply denies.\n\nThe joining of names is not innocent. *Mental*, *workspace* and *think* are words people want, and\nthe reply puts them in quotation marks. Whether the thing found may be called by those names is\nmost of the exchange ([prime-real-estate-words.md](prime-real-estate-words.md) § 5f). The commonest\nthing under the links between these two texts is a claim on a word.\n\n### 2f. What it used\n\n179 model calls on the keeper's plan, $12.96 at API list prices. **No money was paid**: a dollar\nfigure here measures plan allowance. A naming call with the texts' own lines shown is 7.8 cents at\nlist prices, a check 6.6.\n\nThe second trial was started with a time limit shorter than its work and was stopped before it\nwrote its answers. Nothing was lost: every call was in the record of machine judgments with its\ninputs and its reply, and the answers were rebuilt from there\n([the script](what-a-link-rests-on/bridge_step_recover.py)). The record has been live since\n6 October, and recovering a stopped run is one of the uses it was built for.\n\n## 3. What follows\n\n**A single machine reading of a link is not information about the link.** Four designs, one day\n(the last three against predictions written before they ran):\n\n| Design | What it did |\n|---|---|\n| List what else must be true and count | two or three on nearly every link; agreed with the hand reading on 5 of 30 |\n| Name the pattern of argument; no pattern means no reason | a pattern on 27 of 30, among them 6 of the 7 not-reasons |\n| A gate, with refusal made welcome | 20 refusals where the morning's gate gave 4 |\n| Write the further claim in, then check the step | unsteady across premise order on 14 of 44; premise called idle on most |\n\nThe outside record says the same from other tasks: unprompted, strong models flag a flawed premise\n4% to 40% of the time, and told to look, 73% to 78%; a model made to defend an assigned side does\nworse than one that chooses; a proof generator of an earlier generation returned a proof for every\ngoal it was given, false ones included.\nWhat a model says about a step depends on what it was asked to do with it.\n\n**This is not a fault to be tuned away.** The project's own rule is that an error rate belongs to\nthe task as written before it belongs to the model. Here four wordings of the task were tried and\nthe answer moved with each. The task as posed, *judge this link*, has no steady answer to converge\non, for the reason in § 1.\n\n**Naming is a different kind of act.** It produces something a reader can dispute, and a poor\nnaming costs little: a further claim that is false or beside the point is easier to reject than a\nbare arrow, because the reader no longer has to say what is wrong with the step, only to say no to\na sentence. A paper read earlier in this corpus reports people who detect an error and follow the\nadvice anyway because they cannot put the error into words\n([cognitive-commons-and-deliberus.md](cognitive-commons-and-deliberus.md)). A named claim under a\nlink is the words.\n\n**The known hazard of naming** is in the literature and applies here: a further claim written by\nthe analyst can make a bad step look reasoned, and can be put in an author's mouth. The answers\nthe field has converged on are to mark the claim as the analyst's and never as the author's, and\nto write a claim the step *needs*, not one the author is said to have *used* (the\npragma-dialectical rule against attributing unexpressed premises; Walton and Reed; Ennis's\ndistinction). For a link between two texts this is simpler than for one text: neither author made\nthe step, so the further claim belongs to whoever proposed the link, which here is a machine.\n\n## 4. The design (proposed)\n\n**4.1 The object is a step.** A link a machine proposes between claims of two texts is stored\nwith what it would rest on: up to three further claims, each either a pointer to a claim one of\nthe texts already has or a new sentence; the pairs of names it took to be one thing; and, where\noffered, a middle. All of it is marked as the proposer's and as asserted by nobody. One claim was\nmeasured to be too few. Three is the size of what people and the model both write unasked.\n⚠ *Replaced the same evening by one general rule, with the names joined and any further fact as\ntheir own parts: a free form drifts to the two claims joined by* if *and* then\n*([writing-out-the-unwritten.md](writing-out-the-unwritten.md) § 5, § 6).*\n\n**4.2 The pass names and does not judge.** It returns no verdict on the link, no grade of\ndistance, no degree. Where its naming call notices that the two claims are two wordings of one\nfact, or that the step would run the other way, it says so as a note for the reader whose question\nthat already is (the sameness reader), and decides nothing.\n\n**4.3 The standing property, and its debt.** *Every live support or attack link written by a\nmachine between claims of different texts has a naming made at the current version of the\nquestion, on the current wording of both claims.* A link without one owes one. The debt is a query\nover the graph and the record of machine judgments, in the pattern already used for embeddings and\ntitles ([machine-work-as-a-standing-property.md](machine-work-as-a-standing-property.md) § 3). A\nscript drains it under the pair ceiling built this morning. Nothing runs by itself. On the live\ngraph today the debt would be 871 links.\n\n**4.4 What the graph reads with no model.** Per pair of texts: which joinings of names recur, and\nunder how many links; which named claims recur (grouped by wording as candidates, confirmed by the\nsameness reader, since wording groups whole topics together); which links repeat a path already\nthere. Per link, later: whether a named claim has since been stated by some text, or disputed.\nThese are readings, recomputed as the graph changes, and none of them is stored as a verdict.\n\n**4.5 Names are settled once for a pair of texts, as somebody's claim.** Twelve links resting on\none joining of names is one question, asked once. Its answer is stored as a claim by whoever\njoined the names, and when one of the names is a word people want, it is a move on that word and\ngoes to the concept layer (the part of the graph that tracks contested words) as such. This is the\nruling of the same day applied to links: a definition of a wanted word is somebody's claim on it\n([prime-real-estate-words.md](prime-real-estate-words.md) § 3a, § 5b).\n\n**4.6 What recurs is offered as the debate's unwritten premises.** A claim named under several\nlinks between two texts is proposed as a premise neither wrote, with the number of links resting\non it. This is where single namings stop mattering: no one link's naming is trusted, and a claim\nthat comes back under several links, from as many separate readings, is a finding. It joins the pass that\nalready reads a pair of texts whole for the same purpose; the two reached the same claim on the\nconsciousness pair by different routes.\n\n**4.7 A machine's link stays a proposal until something outside the machine backs it.** It is\ndrawn as a proposal, with *holds if* and its further claims, and it moves no strength. Both are\nalready so in the arithmetic: a link a machine guessed between other people's claims is presumed\nto bear nothing until it is read or argued\n([relevance-and-the-two-factors.md](relevance-and-the-two-factors.md) § 5). What is proposed is\nthat this never changes on a machine's say. It changes when a text is found to state the further\nclaim, or when a person argues for it or accepts it. The claim then becomes the link's relevance\nclaim, and the link's bearing follows that claim's standing, by the machinery already ruled. The\nproject already prefers removing a payoff to building a detector wherever a detector can be\noutrun. Here no detector held at all, so the payoff is what is left to remove: a link no machine\nreading can promote cannot be made to count by a reading that happens to be lenient.\n\n**4.8 Links inside one text are left as they are.** There the author made the step, and it stands\nas written until somebody asks why. Such a link already carries its pattern of argument, and the\npattern's questions are the same thing in standing form: asked for a further claim under eleven\nsuch links, the reader wrote what the pattern's question would have asked (*is the example\ntypical*, *is it common enough to count*). Wigmore's practice and the mapping teachers' caution\npoint the same way: write out what a step rests on where there is a doubt, not everywhere.\n⚠ *Not his request, which is for all of it. The revised design starts with the links between texts\nand goes on to an author's own steps, where the pattern on the link already supplies the most\ngeneral rule ([writing-out-the-unwritten.md](writing-out-the-unwritten.md) § 9.1).*\n\n**4.9 A person reads the thirty.** Each now has a hand grade by a model, two gates that disagree\non eighteen, a check, and two or three namings. Reading them is the measurement nothing here\ncould make. Until then no rate is claimed for any of this.\n\n## 5. What it would cost\n\nNaming all 871 links the step has written would be about $68 at list prices, on the plan and so in\nallowance, not money. For new links it adds about one naming call for each support or attack link\nwritten (277 between the two consciousness texts, where 1,735 calls were made to find them). It can also be\nasked in the linker's own call; whether that changes the linker's answers is unmeasured, and this\nmorning's attempt to merge two of its calls made them agree less.\n\n## 6. Proposals (all unruled; ledger AD6 to AD12)\n\n| | Proposal |\n|---|---|\n| AD6 | **A link a machine draws between two texts is stored as a step: with up to three further claims it would rest on, the names it joined, and a middle where offered, all marked as the proposer's.** Replaces *one further claim* in AD2, which was measured to be too few. |\n| AD7 | **No machine reading of whether a link is a reason, how far it reaches, or whether its premise is idle may write, hold, retire, weigh or rank a link.** Four designs measured, each moved by its wording. Strengthens AD5. |\n| AD8 | **Naming is a standing property with a debt query**: machine-written links between texts with no naming at the current version on the current wording. Drained by a script, never by itself. |\n| AD9 | **Names joined are settled once per pair of texts, as a claim by whoever joined them**, and go to the concept layer when a wanted word is among them. |\n| AD10 | **Further claims that recur across the links between two texts are offered as that debate's unwritten premises, with the count.** Wording proposes the groups and the sameness reader confirms them. Sharpens AD4. |\n| AD11 | **A machine's link is drawn as a proposal and weighs nothing until a text states its further claim or a person argues for it or accepts it; no machine reading changes that.** Display is new. The weight is already so. |\n| AD12 | **The thirty links are read by a person before any rate is claimed**, with the machine readings beside each. |\n\n**Considered, and not proposed.**\n\n- *A second, stricter reader that refuses links, with the linker writing only what survives.* Its\n  nine survivors were all reasons by the hand reading. But it refused five links whose further\n  claim is plain, a third question then called five of its ten survivors idle, and both readers\n  are one model. Nothing here shows it would refuse the same links tomorrow.\n- *The check of the completed step as a gate.* Unsteady, and it answers to the wording of the\n  named claim.\n- *A lexical screen for missing terms.* No signal on decontextualised claims.\n- *Retiring the links that repeat a path.* A link over a path can be an argument of its own. They\n  are counted, and the count matters when strength flows; nothing is retired on it.\n- *Leaving the far links unnamed and waiting for readers to object.* A reader cannot object to\n  what is not on the page, and the morning's finding was that the unsaid claim is where the\n  disagreement sits.\n\n## 7. What this does not say\n\n- **That the linking step's links are worthless.** By one model's hand reading 19 of 30 are\n  reasons at one step or one further claim. What failed is every way tried of telling which.\n- **That a machine can never judge a step.** Four designs, one model, thirty links, one day. A\n  reader from a second family of models was not available without spending the keeper's money and\n  was not tried. Trained people reach .72 on single steps given whole.\n- **That naming is reliable.** One naming is one choice among several true things. The design\n  leans on recurrence across links, measured on one pair of texts.\n- **That the hand grades are right.** The stricter reader corrected several, and the grades mixed\n  how far a link reaches with whether its further claim is believable.\n- **Anything about attack links.** Only support links were sampled.\n\n## 8. What was read, and how far\n\nIn the corpus, in full this session: the morning's note, the standing-work proposal, the relevance\nruling and what was built on it, the research on the unsaid premise, the note on parts and\nevidence, the linking step's prompt and reply form, the debt module for embeddings, and the pass\nthat proposes unwritten premises between two texts.\n\nOutside, on 2026-10-07, by two research helpers (models), each returning a report with its own\nstatement of how far every source was read. Both reports are kept unedited:\n[argumentation theory](what-a-link-rests-on/outside-reading-argumentation-theory.md) and\n[machines naming premises](what-a-link-rests-on/outside-reading-machines-naming-premises.md).\n**Nothing in them was checked against the sources by Claude or by a person. Verify before citing\noutward.** The figures used above, with what the helper says it read:\n\n| Used here | Source, as the helper cites it | The helper's read depth |\n|---|---|---|\n| Some premise always makes a step valid; the bare conditional says nothing | Govier (2018 [1987]); van Eemeren & Garssen (2026); Hitchcock (1985) | full text |\n| A relevant premise: suffices with a companion, the companion does not suffice without it, the companion is checkable without the conclusion | Hitchcock, *Inference claims* (2011), p. 198 | substantial extract |\n| Redundancy of a premise given up as a test | Hitchcock, *Enthymematic arguments* (1985), p. 88 | full text |\n| The further claim marked as the analyst's; needed and used assumptions | Walton & Reed (2005); Ennis (1982, at second hand) | extract; not opened |\n| No edge from statement to statement | Chesñevar et al., the argument interchange format (2006) | substantial extract |\n| No warrants stated, doubts charted | Goodwin on Wigmore (2000) | search extracts |\n| The Rabbit Rule | Rider & Thomason (2008); the Rationale tutorials | extract; summary |\n| 2.7 premises, a third of the words shared, 1% to 15% declined | Boltužić & Šnajder (2016) | full text |\n| Kappa .42 on reason implies claim; free gap-filling abandoned | Habernal et al. (2018) | full text |\n| Kappa .72 on single steps; a generator proves every goal | Bostrom et al. (2022) | substantial extract |\n| An assigned side does worse than a chosen one | Ku et al. (2025) | full text |\n| 4% to 40% unprompted, 73% to 78% told to look | a 2025 benchmark of premise critique | search extracts only |\n\n**The scripts and their outputs** are in [the folder](what-a-link-rests-on/): the lookup, the free\ncounts, the Rabbit Rule count, both trials with their predictions, the rebuild from the record,\nthe scoring, the recurrence count, and what each run printed.\n\n## Cross-references\n\n- [how-far-a-premise-is-from-its-claim.md](how-far-a-premise-is-from-its-claim.md) (the morning's\n  measurements and proposals AD1 to AD5; this note revises AD2 and strengthens AD4 and AD5)\n- [machine-work-as-a-standing-property.md](machine-work-as-a-standing-property.md) (debt as a\n  query; the record of judgments, used here to recover a stopped run)\n- [relevance-and-the-two-factors.md](relevance-and-the-two-factors.md) (a link's bearing follows\n  its relevance claim; a machine's guess presumes none)\n- [the-load-bearing-unsaid.md](the-load-bearing-unsaid.md) (the crux is usually unwritten)\n- [prime-real-estate-words.md](prime-real-estate-words.md) (a joining of names can be a claim on a\n  wanted word)\n- [auto-connect-upgrade.md](auto-connect-upgrade.md) § 2c (the linking step, its repair and cost)\n- [earning-the-right-to-write.md](earning-the-right-to-write.md) (more places to disagree, as the\n  test for what a machine may add)\n- ledger: [session26-open-rulings.md](session26-open-rulings.md) Arc AD\n"}