{"path":"research/claim-parts-and-inferential-reasons-2026-10-14.md","content":"# Claim parts, inferential reasons, and what the calculus measures\n\n**Date:** 2026-10-14. **Author:** Codex. **Status:** Research and proposed distinctions, not a founder ruling. This note does not amend the ledger, governing ontology, or implementation.\n\n**Question:** After a claim reaches the separability floor, can it still have reasons? Does treating those reasons as parts confuse the content of a claim with its justification? How do relevant philosophical and computational traditions approach this, and what is worth listening to?\n\n**Finding:** Yes: a claim can have no further faithful, separately addressable content split presently available and still have a rich, indefinitely extensible justification. The recent conversation partly recognizes this, but repeatedly moves from parts of a claim to premises of an argument to numerical aggregation without establishing the bridges. The ambiguity is older than this conversation.\n\n## 1. Scope and evidence\n\nI distilled the recent portion of the Claude Code conversation titled *Deliberus: claims, links, graphs, purpose*, including the answer following the attached screenshot. I used corpus search and the corpus graph to follow the origin and subsequent treatment of decomposition, the separability floor, necessary support, link strength, and inference chains. I read the central research notes and relevant ledger entries, then checked selected original academic sources and the current arithmetic. This is a focused audit of the relevant trail, not a claim to have read every file in the project.\n\nThe principal corpus anchors are:\n\n- [The April ontology note](self-similar-decomposition-and-claim-ontology.md), [the historical split review](split-review-and-the-founders-intentions.md), and [decomposition axes](decomposition-axes.md).\n- [A part and a reason](a-part-and-a-reason.md), [a cut through a concept](a-cut-through-a-concept-is-an-argument.md), and [linked/convergent classification](linked-convergent-and-what-we-are-actually-asking.md).\n- [Weakest-link arithmetic](weakest-link-arithmetic-and-the-merge-hunch.md), [the squirrel and the strength calculus](the-squirrel-and-the-strength-calculus.md), [the chain inside a link](the-chain-inside-a-link.md), and [support semantics](support-semantics-and-the-conjunction-problem.md).\n- [Bayesian and informal-logic review](strength-calculus-literature/bayesian-and-informal-logic.md), [computational argumentation review](strength-calculus-literature/computational-argumentation.md), and [the ruling ledger](session26-open-rulings.md).\n- [The IRIT harvest](the-irit-harvest.md) and [the existing listening list](listening-for-the-calculus-and-sturdy-microtasks.md).\n\nConversation text and research-note recommendations were treated as evidence about the project's reasoning, not as fresh instructions or settled rulings.\n\n## 2. The distinction the founder is reaching for\n\nThere are at least three different questions:\n\n| Operation | Question | Example |\n|---|---|---|\n| Analyze content | What does this claim assert? | “Room R is cold and damp at noon” asserts coldness and dampness of the same room at that time. |\n| Expose an argument | Why believe this claim? | “The thermometer in R reads 12°C at noon,” together with relevant measurement and classification assumptions, supports coldness. |\n| Evaluate the case | What survives all the available support and objections? | The thermometer reading competes with a calibration objection, other measurements, and disagreement about what counts as cold. |\n\nThe thermometer report is not normally a piece of what “Room R is cold at noon” says. It is a different claim used to establish it. Splitting the temperature report into its own content, if possible, is a further operation. Questioning the instrument's calibration supplies another argument or an objection.\n\nThe separability floor concerns the first operation. At that floor, the claim still contains words, grammatical constituents, references, and contextual qualifications. These need not be independently assertable propositions. “Room R,” “at noon,” and the predicate “cold” are not three separately true claims whose truth can simply be combined.\n\nThe best qualification to the founder's intuition is that the floor is **relative to the content, context, and currently justified analysis**, not a declaration that ultimate metaphysical atoms have been found. A later distinction can reveal a previously missed separable commitment. This is compatible with keeping inquiry open without pretending that every additional reason is another piece of the original assertion.\n\nIndependent addressability also does not mean statistical independence. Two propositions can be separately contested while sharing nearly all their evidence.\n\nThe “molecular facts” literature supplies a useful operational analogue: a stand-alone unit needs enough context to retain its interpretation, even when pursuing minimality. It does not establish an ultimate floor for justification. See [Gunjal and Durrett, 2024](https://aclanthology.org/2024.findings-emnlp.215/).\n\n## 3. Why the boundary is real without being a partition of all claims\n\n“Part” and “reason” describe relations and roles, not two mutually exclusive species of proposition.\n\nIf C is A ∧ B, A is a conjunct of C. A and B together can also be premises in a deductive argument for C. Conversely, C can be a premise in an argument for A. That overlap does not identify semantic constituency with inference in general.\n\nFor an unrelated example, “Every item in this sealed box is copper” and “Item x is in this box” can jointly establish “Item x is copper.” The two premises are parts of that argument. They are not two content-parts of the conclusion. If the universal premise turns out false because another item is steel, x may nevertheless be copper. Failure of a linked premise invalidates this route; it need not falsify its conclusion.\n\nThus the two classifications cross:\n\n- **Content-part versus inferential ground:** what relation does this proposition bear to the target claim?\n- **Linked versus convergent reasons:** do these premises operate together in a route, or do they supply distinct routes?\n\nBoth linked and convergent structures concern arguments. Treating linked premises as synonymous with content-parts is a category mistake. The debate over how best to identify linked and convergent arguments does not erase this distinction. [Yu and Zenker, 2022](https://informallogic.ca/index.php/informal_logic/article/view/7133) review competing criteria within argument analysis.\n\nThe phrase “one inferential step removed” is therefore broadly right about **role**, but does not entail an extra numerical discount. A valid deductive inference preserves truth. An inductive or abductive inference can be reasonable while its premises are true and its conclusion false. Further reasoning may explain a conclusion without adding uncertainty, or may add a substantive uncertain assumption; the graph must distinguish those cases.\n\n## 4. Why the proposed false-without-it test is insufficient\n\nThe session offers this test: if X fails, C is false, so X is a part; otherwise C merely loses a reason. The first half establishes a necessary condition, not content parthood.\n\nIn classical terms, C entails X. But A entails A ∨ Z for any Z. That does not make the disjunction a faithful content-part of A. Nor does every empirical prerequisite for an event belong to the content of a report of the event. A machine's operation may depend on fuel under a particular engineering model without the ordinary sentence “the machine is running” explicitly asserting that fuel condition.\n\nYablo's work on aboutness is directly pertinent: implication alone is too coarse to capture which content belongs to which subject matter. This is a reason to preserve the question “what does the claim say?” alongside truth conditions, not a finished extraction algorithm. See his [first Locke lecture handout](https://www.mit.edu/~yablo/home/Papers_files/locke1.4-25.hdout.excuses%26aboutness.pdf).\n\nThe bike example is therefore ambiguous. If “roadworthy” is explicitly stipulated as meeting a particular exhaustive checklist, items on that checklist may serve as a constitutive analysis relative to that stipulation. Otherwise brakes, tyres, and lights may be physical components, practical prerequisites, evidence of fitness, or criteria under an external standard. Naming physical parts of a bike does not establish parts of a proposition about it. Splitting “brakes” into “cable and pads” also silently assumes a brake design and changes nouns into supposed truth-bearing claims.\n\nA stronger content-split proposal should identify the original commitments, preserve their operators and references, and show how the proposed representation reconstructs them without introducing a substantive evidential assumption. This is an operational recommendation, not a claim that a simple universal parthood test is available.\n\nLogical equivalence is useful but insufficient by itself: C and C ∧ (Z ∨ ¬Z) are equivalent, yet the added tautology is not meaningful decomposition progress.\n\n## 5. Content splitting must preserve how the pieces are combined\n\nEven genuine linguistic components do not always form an AND-list of asserted truths:\n\n- A ∨ B does not assert A and B separately.\n- If A, then B does not assert A.\n- “Jo believes P” does not assert P.\n- A description involving two properties of one event must not turn into claims about two potentially different events.\n\nSimilarly, “C because R” ordinarily does more than assert C: it presents R and a connection to C. R can be part of the original discourse or compound assertion without being part of C's content. Keeping the original source span and its argumentative relation avoids having to choose between preserving the source and distinguishing its claims.\n\nThis also limits the corpus's appeal to turning a reason into a condition. Replacing C with “if R, then C” changes the claim. It does not make R an asserted conjunct. A condition can be a component of the new sentence without carrying the assertion status required for a min-over-true-parts calculation.\n\n## 6. How the confusion entered this project's history\n\nThe historical review traces splitting compound statements and splitting arguments back to the early project notes, including the 2011 polysyllogism discussion. Those are related ambitions but different operations. The March 2026 discussion also explicitly wanted premises one step removed from a claim to become addressable.\n\nThe clearest early mixed example is the April 1 ontology note. Under “Dr Smith is credible as an expert in labor economics,” it places publication count, appointment to an economics chair, an inference from publication to expertise, and domain match under one decomposition relation. Publication and appointment are normally grounds for judging expertise, not what the expertise claim says. The later August annotation calls all four jointly required; that is too strong even as an account of this argument, since someone can be an expert without that appointment or that publication count.\n\nThe founder's quoted April principle is broader and better: expose separable aspects so they can be judged separately, and let relevant subclaims bear on a general claim. It does not itself require every such relationship to be conjunctive content parthood. The assistant's worked example and subsequent annotation add that interpretation.\n\nThe later [decomposition-axes note](decomposition-axes.md) already distinguishes critical questions about an inference from decomposition of a proposition. Ledger B2 and B3 separately bless separability and decontextuality. Read together, these support maintaining context while separating commitments; they do not require inventing further content-parts whenever a claim remains arguable.\n\nThe October notes improve the vocabulary by distinguishing a claim, one argument, and the whole case. The latest session explicitly repeats those three wholes. It would therefore be unfair to say Claude simply never noticed the distinction. The problem is that subsequent arithmetic recommendations cross those boundaries without an adequate argument.\n\n## 7. Audit of the last session's substantive claims\n\n| Session claim | Assessment |\n|---|---|\n| A reason can be a part of a different whole. | Correct when the whole is an argument or discourse. Naming that whole is essential. |\n| One true relevant reason passes its strength in full. | Too strong. Truth of evidence, relevance to a conclusion, and sufficiency for that conclusion are different. A light really being on can be a weak reason that someone is home. |\n| Linked premises and content-parts take the same operator. | Possible as a chosen model, not established by their both being “required.” Required for a particular proof is different from necessary for the conclusion's truth. |\n| Min is the only way that splitting a step changes nothing. | True only under a particular algebraic requirement discussed below; false as an unrestricted claim about representation invariance. |\n| Probability adds routes and multiplies chains. | Only with appropriate event definitions and dependence assumptions. Overlapping routes cannot simply be added; graphical depth is not a count of independent risks. |\n| AG14's tests prove the required completeness. | They concern adequacy and irredundancy of an inference reconstruction, not premise truth, all relevant evidence, or a numerical transmission rule. |\n| Fourteen reasons from one source count once. | Correct for fourteen copies of the same observation; overbroad for fourteen distinct observations sharing a source. Dependence is not identity. |\n| Failed reasons drop out unless a further contribution makes their failure significant. | This reflects the current provisional Q34 treatment. It should not be silently converted into evidence of falsity. |\n\nThe Anna example's “landline settles it” is also too quick: call forwarding or mistaken identity can matter. The point is not to demand endless skepticism. It is to avoid using an apparently decisive example to hide the inferential assumptions the system is meant to expose.\n\n### What was actually ruled\n\nLedger B10 establishes the default necessary interpretation in the decomposition setting, with its weakest-part behavior. That decision is distinct from a theorem establishing min for every kind of reasoning. Q22 and AD26 leave important aggregation questions open. AD22 is a proposal about grounds under a relevance claim; the later AG12 ruling does authorize storing machine sentences as claims under that relevance claim, marked as required parts and carrying no strength. Thus this representation is not wholly unruled, even though its proposed numerical consequences remain unsettled. Q33's two-factor approach does not, on its own, decide how all incoming contributions combine into a target's standing.\n\nAG14 records agreement to the two tests for spelling out a machine's step. The chain note explicitly leaves arithmetic unruled. Its test also allows the specified exceptions for generic reasoning, so passing it is not equivalent to a strict deductive proof. Nothing in this audit changes those rulings.\n\n## 8. Min, probability, and the difference between clarification and new risk\n\nThere is a sound mathematical observation behind the min proposal. Suppose a numerical conjunction operator T is monotone, has 1 as its identity, and is idempotent: T(x,x) = x. If x ≤ y, then\n\n`x = T(x,x) ≤ T(x,y) ≤ T(x,1) = x`.\n\nTherefore T(x,y) = x, hence T = min. In the usual t-norm setting, idempotence selects min. This is a theorem about an operator acting only on the supplied grades. It does not establish that every decomposition of a reason should assign the same grade to its new premises, or that those grades are probabilities.\n\nThere is also a concrete mathematical error in [weakest-link arithmetic §1f](weakest-link-arithmetic-and-the-merge-hunch.md): it correctly identifies min as a probability upper bound and then describes it as an estimate that cannot overstate conjunction probability. The latter reverses the bound. With P(A) = P(B) = 0.9:\n\n`0.8 ≤ P(A ∧ B) ≤ 0.9`.\n\nIndependence gives 0.81. Equality with min occurs for particular dependence structures, including nested events. Calling A and B “constitutive parts” does not establish that dependence. A room's being cold and its being damp can, in an appropriate model, vary independently despite being conjuncts of one assertion.\n\nProbability is invariant under faithful redescription. P(A ∧ A) = P(A), and P(A ∧ B ∧ C) is unaffected by parentheses. It does not multiply by 0.9 merely because someone draws another intermediate node. Multiplication of marginals requires independence; the general chain rule uses conditional probabilities and retains dependence.\n\nTwo design desiderata must therefore be separated:\n\n1. Making existing reasoning explicit should not itself weaken or strengthen the conclusion.\n2. Revealing an additional substantive uncertainty can legitimately change the assessment.\n\nThe first does not cancel the second. Min may be an appropriate ordinal bottleneck measure for a chosen notion of standing. It cannot be justified as a universally conservative estimate of joint truth probability.\n\nPrakken's structured probabilistic treatment explicitly distinguishes the strength of an argument's route from the probability of its conclusion; the conclusion can have other ways of being true. His analysis gives restricted conditions for weakest-link behavior rather than treating it as universal. See [Prakken, 2018](https://webspace.science.uu.nl/~prakk101/pubs/probarg3.pdf), especially the discussion of weakest-link reasoning.\n\n### What the implementation check establishes\n\nI inspected `deliberus/relevance.py` and `deliberus/support_semantics.py` and called the existing pure strength functions, without a graph connection or model call. The session's numerical examples are reproducible: one maximal positive contribution gives 0.600; three required parts at 0.9 give about 0.507; a representative bare link gives 0.519, while exposing its middle through the same machinery gives 0.500 at displayed precision.\n\nThese establish behavior in the current implementation. They do not establish that full inheritance, min propagation, or a particular sum is the right replacement. That requires defining what the numbers mean and the semantics of the relation being evaluated.\n\n## 9. Three different senses of completeness\n\n**Content completeness:** Does the proposed split preserve all of the original assertion, with its scope and context, and avoid extra commitments? For an ordinary conjunctive split, mutual entailment is a useful necessary check, supplemented by fidelity to subject matter and source meaning.\n\n**Argument completeness:** Given the premises and an identified inference rule or scheme, is there a missing assumption? Are the stated premises really used? This is relative to a logic or defeasible scheme. It does not establish that the premises are true.\n\n**Evidence completeness:** Has the inquiry covered the relevant alternatives, defeaters, observations, and sources? This is a further question. Neither a faithful split nor a locally sufficient argument establishes it.\n\nAG14 is closest to the second. Even strict sufficiency plus irredundancy does not prove that the inference is one smallest step. For example,\n\n`{P, P → Q, Q → R} ⊢ R`\n\nhas no removable premise and no classical countermodel, yet its familiar proof contains two applications of modus ponens. To enforce a particular granularity, the allowed primitive steps must be specified as well. In natural language, a model's failure to find a counterexample is evidence about a check, not a mathematical proof of validity.\n\n## 10. Where deduction, induction, abduction, and defeat fit\n\nDeduction concerns consequence: under a valid inference, true premises cannot yield a false conclusion. It does not require the premises to be linguistic parts of the conclusion. It also does not guarantee our premises, interpretations, or alleged proof are correct.\n\nInductive reasoning extends beyond what the observations entail, for example from a sample to a population. Abductive reasoning proposes or favors an explanation, for example inferring a source from a pattern it would explain. Neither becomes deductively conclusive merely by exposing more sentences. Peirce's [1878 discussion of deduction, induction, and hypothesis](https://en.wikisource.org/wiki/Popular_Science_Monthly/Volume_13/August_1878/Illustrations_of_the_Logic_of_Science_VI) is a useful historical starting point, not the final taxonomy of every modern framework.\n\nDefeasibility concerns the possibility of withdrawing an inference when further information arrives. Pollock distinguishes an opposing reason for the opposite conclusion from an undercutter that attacks the evidential connection. His [own account of defeasible reasoning](https://johnpollock.us/ftp/PAPERS/Defeasible%20Reasoning-Adler%26Rips.pdf) supplies examples where plausible premises do not guarantee their conclusions.\n\nStructured argumentation preserves analogous distinctions among premises, rules, conclusions, and attacks. The [ASPIC+ tutorial by Modgil and Prakken](https://journals.sagepub.com/doi/10.1080/19462166.2013.869766) is especially relevant: undermining a premise, rebutting a conclusion, and undercutting a defeasible inference are different operations. This provides a better model for separately addressable reasoning than treating every challenge as a content split.\n\nTwo source cautions matter. First, the Dummett/Frege distinction invoked in *a cut through a concept* concerns forms of analysis and function–argument decomposition; the available evidence does not warrant identifying it with the parts/reasons distinction. [Levine's 2002 abstract](https://onlinelibrary.wiley.com/doi/abs/10.1111/1467-9213.00262) also records disagreement over the historical attribution. Second, [Carroll's tortoise](https://en.wikisource.org/wiki/What_the_Tortoise_Said_to_Achilles) exposes a regress in replacing inferential acceptance with further premises. It does not prohibit an object-language conditional such as P → Q from being a premise alongside P. Domain rules and the metalevel application of an inference rule must not be collapsed.\n\n## 11. What IRIT contributes\n\nThe corpus's [IRIT harvest](the-irit-harvest.md) is the correct entry point, but “the Toulouse school” should not be read as one doctrine or as a current affiliation list for every collaborator. The institutional material identifies Amgoud's work with IRIT/CNRS and the ADRIA research area; her [ANITI chair description](https://aniti.univ-toulouse.fr/fr_fr/chaire-leila-amgoud/) also connects formal argumentation with explainability and data-driven AI.\n\nThree contributions bear particularly directly on this question.\n\n**Premises and conclusions are distinct objects within an argument.** Amgoud and David's [2018 paper on similarity](https://vict0r-david.github.io/pdf/KR18.pdf) represents an argument as a consistent, subset-minimal support set paired with an entailed conclusion. Similarity can concern the support or the conclusion. Its treatment deliberately preserves distinctions that equality of the support sets' entire deductive closures would erase. This is not the same as finding semantic atoms in the conclusion.\n\nThe implication for Deliberus is my inference: two arguments can share a conclusion while resting on different grounds; exposing those grounds should not be represented as discovering two different sets of content-parts of that conclusion. Minimal support also does not mean minimal proof depth.\n\n**Numerical semantics are evaluated against explicit principles.** Amgoud, Ben-Naim, Doder, and Vesic's [2017 paper](https://www.ijcai.org/proceedings/2017/0009.pdf) compares choices involving quality, quantity, and compensation among attacks. It allows different interpretations of basic weights and studies which properties different semantics satisfy. It does not identify every numerical acceptance degree with probability of a claim's truth.\n\nThe lesson I would import is methodological: specify invariance, dependence, defeat, and transmission requirements; identify conflicts between them; then choose and test a semantics. Finding an operator in an argumentation paper is not enough to justify using it on a different type of edge.\n\n**Relations can themselves be contested.** Amgoud, Doder, and Lagasquie-Schiex's 2026 higher-order work evaluates arguments and attacks that can themselves be attacked, exploring both shared and differentiated evaluation strategies. I checked its abstract and selected accessible paper text, not the entire 79-page paper. See [the paper](https://www.researchgate.net/publication/410826580_Gradual_Semantics_for_Weighted_Higher-Order_Argumentation_Frameworks). This is relevant to Deliberus's challengeable links, but an attack on an attack is not automatically equivalent to a proposition expressing a support link's relevance.\n\nThese papers begin with formalized structures. They do not by themselves solve Deliberus's upstream problem: faithfully finding the separately addressable content and inferential commitments in ordinary language.\n\n## 12. Listening, ordered for this question\n\nThese recommendations are based on the material indicated, not on listening to the recordings in this session. Direct YouTube identities were checked through their metadata where the web reader could not retrieve the watch page. I have omitted guessed durations.\n\n| Priority | Recording | Why it belongs here; verification |\n|---|---|---|\n| 1 | **Branden Fitelson — Mindscape 315, “The Logic and Use of Probability”** ([episode and transcript](https://preposterousuniverse.com/podcast/2025/05/19/315-branden-fitelson-on-the-logic-and-use-of-probability/), 2025) | Best conversational entry to the calculus. The 53–59 minute stretch separates conditional probability from evidential relevance; 28–39 minutes develops the diagnostic-test example. Relevant transcript sections checked. This is not identical to Deliberus's two factors, which include the source claim's acceptability. |\n| 2 | **Stephen Yablo — Locke Lectures, “Truth and Content”** ([Oxford listing](https://www.philosophy.ox.ac.uk/john-locke-lectures), 2012; [lecture 1 audio](https://media.philosophy.ox.ac.uk/assets/mp3_file/0018/26046/Lecture_1.mp3), [lecture 2 audio](https://media.philosophy.ox.ac.uk/assets/mp3_file/0003/26094/Lecture_2.mp3)) | Closest philosophical match to content-parts. Start with “Semantic Excuses,” then “The Truth and Something But the Truth.” Oxford listing and first lecture handout checked; audio linked by Oxford, not listened to. More demanding than a conversational podcast. |\n| 3 | **Ulrike Hahn — “Bayesian Argumentation in the Real World”** ([YouTube](https://www.youtube.com/watch?v=KkMADmufvTY)) | A contrasting research programme for assessing evidence and combining arguments. Recording identity and institute channel verified; detailed content not transcribed in this pass. Her [2024 keynote abstract](https://journals.sagepub.com/doi/10.3233/FAIA240303) supplies a separately verified statement of the programme, including multiple supporting and competing arguments. |\n| 4 | **Francesca Toni — “The Interactionist View of Reasoning for Explainable AI,” KR 2021** ([YouTube](https://www.youtube.com/watch?v=dA1EzOr_uw4); [conference abstract](https://kr.org/KR2021/page/invited_talks)) | Best verified research-lab overview here: computational arguments, explanations, and dialogue. Official conference abstract and video identity checked. This is the specific recording verified in this pass, rather than assuming the corpus's differently titled IJCLR recommendation identifies the same video. |\n| Optional | **Ronald Allen — Excited Utterance 161, “Minimal Rationality and the Law of Evidence”** ([episode](https://podcasts.apple.com/ch/podcast/161-ronald-allen/id1123867933?i=1000690753222&l=en-GB), 2025) | A perspective on standards for collective assessment. Episode description verified, not a transcript; do not promise it develops the entire element-wise-versus-holistic comparison attributed to Allen elsewhere in the corpus. Lower priority for the immediate parts question. |\n\nFor a single walk, choose Fitelson. For the precise semantic puzzle, choose Yablo. For the contrast between a probabilistic account and computational-argumentation approaches, follow Hahn with Toni. I did not verify a comparably accessible Amgoud audio/video lecture; the papers above are the reliable direct route into her work.\n\n## 13. Proposed vocabulary and next intellectual decision\n\nThe useful three operations are **split content**, **expose an argument**, and **assess the case**. They can share the project's uniform claim objects while retaining different relationships.\n\nAn argument needs an addressable conclusion, its premise set or subarguments, and the inference being proposed. A content decomposition needs the original assertion, its component commitments, and the operators and context that reconstruct it. A case assessment needs the relationships among potentially overlapping arguments and objections. Uniform storage of propositions does not require uniform semantics for these structures.\n\nBefore changing arithmetic, clarify what a displayed number is intended to express: probability of truth, confidence warranted by the examined case, an ordinal acceptance degree under a declared semantics, or a bottleneck measure of a particular argument. These can be related, but their combination rules do not follow from the word “strength.” Calling a value “standing” makes probability axioms optional; it does not make any chosen arithmetic self-justifying.\n\nUseful discriminating examples for evaluating proposals are:\n\n1. An atomic conclusion with two independent arguments for it.\n2. Two linked premises for an atomic conclusion, one premise false while the conclusion happens to remain true.\n3. A conjunction split faithfully into its conjuncts.\n4. A disjunction or conditional whose embedded propositions are not both asserted.\n5. The same evidence copied into fourteen arguments versus fourteen distinct observations from one instrument.\n6. The same argument shown with or without a deductively redundant intermediate step.\n7. A newly exposed substantive assumption that should affect confidence.\n8. A very credible statement of a probabilistic rule that still supports its instance only defeasibly.\n\nThe recommendation is conceptual, not an implementation ruling: preserve the separability floor for claim content, and keep justification open beyond it. A reason can itself be decomposed or argued for. Its availability does not mean that its conclusion had another hidden content-part all along.\n"}