{"id":"97d17055-ced0-46aa-baeb-80658607a18b","arxiv_id":"2509.08433","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"The paper defines a similarity score S* that subtracts a contradiction ratio from a shared-property ratio, and organizes knowledge entities into threshold-based paraconsistent super-categories.","lead":"This paper proposes a similarity score for knowledge bases that subtracts a contradiction penalty from the usual shared-properties ratio, plus a hierarchy and repair mechanism for conflicting facts. It is a conceptual framework with toy examples, not yet an empirically validated system.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Repair mechanism is undefined for pairs and its only worked example is arithmetically wrong; the framework's consistency-preservation claim is unsupported, although the bare S* formula may be salvageable.","rationale":"The central S* formula is simple, and the reflexivity, symmetry, and boundedness results are plausible under a charitable reading of Pshared, Pcontradictory, and Ptotal. So the basic measure is not worthless. However, the paper's advertised repair mechanism is one of its four stated contributions and is the part that justifies the abstract's phrase 'garantissant la cohérence des évaluations.' That mechanism fails in the only example where it is demonstrated: Definition 4 is explicitly single-entity, yet Example 2 uses E on a pair, and the repaired similarity value in Eq. (19) is arithmetically inconsistent with Eq. (3)-(4). These are not stylistic issues; they mean the repair algorithm is either underspecified or incorrect, and Theorem 2 and Corollary 3.1 merely restate that repairs preserve categories without a valid construction of the repaired entities. The also-noted absence of real-data validation (Sec. 4.3) further undercuts the abstract/conclusion claims of medical and legal validation. Because the repair/hierarchy guarantees are load-bearing for the paper's novelty, the reader's REJECT verdict is appropriate; no verdict adjustment is needed, though the bare S* formula might be salvageable in a revised version.","tokens_in":13213,"tokens_out":8287,"duration_ms":90911,"concrete_test":"Independently re-derive the repair step of Example 2 from Definitions 1 and 4 and Eq. (3)-(4). First apply Eq. (11) to K2 alone: since K2 contains no literal toux, E(K2)=∅, so Eq. (18) requires a pair-based E that is not specified. Then compute S*(K1,K'_2) with K'_2={fièvre, maux de tête}: Pshared={fièvre}, Pcontradictory contains at least the pair (¬maux de tête, maux de tête), and |Ptotal|=4, so Eq. (3)-(4) give S*≤0, not 0.25. If the authors intend E to be pair-level, the check is to rewrite Definition 4 and all dependent theorems for pairs and recompute Eq. (19); if the arithmetic still disagrees with Eq. (3)-(4), the repair mechanism is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing weakness is the contradiction-extraction/repair pipeline advertised as a core contribution. Definition 4 (Eq. 11) defines E(K) for a single entity K, and the text explicitly says it identifies contradictions within the same entity. But Example 2 applies E to the pair (K1,K2), returning E(K2)={toux, ¬toux} even though K2={fièvre, ¬toux, maux de tête} contains no positive literal toux, so E(K2)=∅ under Eq. (11). A cross-entity, pair-level extractor is needed but never defined. The only worked repair computation is also inconsistent with Eq. (3)-(4): for K1={fièvre, toux, ¬maux de tête} and K'_2={fièvre, maux de tête}, we have |Pshared|=1, |Pcontradictory|≥1 (the pair ¬maux de tête / maux de tête remains), and |Ptotal|=4. Hence S*(K1,K'_2)=0 (if pairs are counted) or −1/4 (if contradictory literals are counted), not the claimed 0.25 in Eq. (19). Since Proposition 5, Theorem 2, and Corollary 3.1 all depend on this repair operation, the 'gestion des contradictions' contribution has no valid demonstration. Section 4.3 also concedes that real-data validation remains to be done, contradicting the abstract and conclusion.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a paraconsistent framework for comparing knowledge-base entities represented as sets of literals. Its main measure is S*(K1,K2) = (|Pshared| - |Pcontradictory|) / |Ptotal| (Eqs. 3-4), intended to reward shared properties and penalize contradictions. The framework also defines paraconsistent super-categories Ξ*_K (Eq. 9), a contradiction extractor E (Eq. 11), and a repair mechanism, and it claims reflexivity, symmetry, boundedness, category preservation, and post-repair coherence. Two worked examples (Section 3.7) and a comparison with Jaccard similarity are provided. The paper is written in French and has an abstract, contributions, theory, examples, discussion, and conclusion.","tokens_in":13535,"tokens_out":11439,"duration_ms":120668,"significance":"If the basic S* formula were the only contribution, it would be a simple, interpretable, parameter-free normalized difference that reduces to Jaccard when no contradictory pairs are present. However, the additional advertised contributions—the hierarchical super-categories, the contradiction extractor, and the repair mechanism—are not correctly defined or demonstrated. The formal results are either immediate consequences of the chosen normalization or are unsupported. There is no empirical validation, and Section 4.3 explicitly concedes that validation on real datasets remains to be done. The underlying idea is salvageable, but in its current form the manuscript does not support the main claims.","major_comments":[{"comment":"The sets Ptotal and Pcontradictory are never formally defined, and Proposition 1 (Eq. 5) is false for inconsistent entities. If K={p,¬p}, comparing K with itself yields at least one contradictory pair (p,¬p) under Definition 1, so Pcontradictory≠∅ and S*(K,K)≠1. The proof 'par définition, Pcontradictory=∅' is not supported by any definition and conflicts with the existence of internal contradictions recognized by E in Eq. (11). Reflexivity cannot be guaranteed as stated.","section":"§3.3, Definition 2 and Proposition 1"},{"comment":"The definition of Ξ*_K includes Ki only if S*(Ki,Kj)>θ for all j≠i. In Example 2, K1 and K3 do satisfy S*(K1,K3)=0.5>0.4, but K1 fails against K4 (S*=-0.17) and against K2, so under Eq. (9) neither K1 nor K3 belongs to any super-category. The displayed partition {{K1,K3},{K2},{K4},{K5}} (Eq. 17) is not the set defined by Eq. (9). Either the quantifier is wrong (e.g., intended ∃j≠i or a transitive-closure rule) or the example is inconsistent. Theorem 1 also lacks a precise definition of 'different super-categories' under Eq. (9).","section":"§3.5, Definition 3 (Eq. 9)"},{"comment":"Definition 4 defines E as an intra-entity operator: E(K) is the set of pi∈K contradicting some qj∈K. In Example 2, E(K2) is claimed to be {toux, ¬toux}, but toux∉K2 and K2 contains no complementary pair, so Eq. (11) gives E(K2)=∅. Detecting the contradiction between K1 and K2 requires a pair-level extractor, which is never defined. Hence the repair step 'supprimons ¬toux de K2' is not licensed by the framework.","section":"§3.6.1 and Example 2 (Eq. 18)"},{"comment":"The only numerical repair calculation is arithmetically wrong. With K1={fièvre,toux,¬maux de tête} and K'_2={fièvre,maux de tête}, the shared set is {fièvre}, the contradictory pair (¬maux de tête / maux de tête) remains, and the union has 4 literals. Hence S*(K1,K'_2)=(1-1)/4=0, not 0.25. The value 0.25 ignores the remaining contradiction. This invalidates the illustrative support for the repair mechanism and, consequently, for Proposition 5, Theorem 2, and Corollary 3.1.","section":"§3.7, Eq. (19)"},{"comment":"The characterization 'K réparable si et seulement si E(K)≠K' is false as stated. If E(K)=K, e.g., K={p,¬p}, choosing R=K yields K\\R=∅, which is contradiction-free; Definition 5 is satisfied. The proof's claim that E(K)=K leaves no coherent subset is incorrect because the empty subset is coherent under the formal definition. The minimal-repair claims (Corollary 1.1, Corollary 3.1) therefore lack a valid basis.","section":"§3.6.5, Proposition 5"},{"comment":"These results are asserted rather than proved. Theorem 2 claims that repaired similarity preserves original similarity categories, but in Example 2 the category changes from negative (S*=-0.2) to S*=0 (or claimed 0.25) after repair; the categories are not preserved. Theorem 3's proof is a restatement and does not show how repairing a super-category yields coherence in 'all sub-categories.' Corollary 3.2 inherits this gap. These should be stated as conjectures or proved from the repair definition.","section":"§3.6.5, Theorems 2-3 and Corollary 3.2"}],"minor_comments":[{"comment":"For pairs with no overlap and no contradiction, the denominator is given as 5 instead of 6: (K1,K5), (K2,K5), and (K3,K5). The scores are 0 either way, but the computed fractions are inconsistent with the definition of Ptotal.","section":"§3.7, Example 2"},{"comment":"The paper concedes that 'une évaluation sur des ensembles de données réels restent à réaliser,' which contradicts the abstract's and conclusion's wording that the framework has been validated and is promising for practical applications.","section":"§4.3"},{"comment":"The abstract and conclusion claim validation by examples from medical and legal domains, but Section 3.7 contains only medical examples; no legal example is provided.","section":"Abstract and conclusion"},{"comment":"The citation key [?] in the first paragraph of the Introduction is unresolved. Additionally, Definition 6 equates ΞRP(K1,K2) with ΞP(K'_1,K'_2), mixing a similarity measure with a property-space notation; this type mismatch makes the definition hard to interpret. The notation K'≈, K'≠ in Theorem 2 is not defined.","section":"Introduction and Definition 6"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be a very early draft. The unresolved citation, the incorrect example calculations, and the unsupported theorem statements suggest that it is not ready for journal review. I would encourage the author to rework the formal definitions—especially the pair-level contradiction extractor, the super-category definition, and the repair operation—and to add a genuine empirical or formal case study before resubmitting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on Vilchis-Medina's arXiv:2509.08433. The headline: the S* formula is a real, if elementary, variant of Jaccard similarity, and the three basic properties are proved correctly. But the repair mechanism — one of the paper's advertised contributions — is undefined for pairs and the only worked example contains a clear arithmetic error. That sinks the paper as it stands.\n\nWhat's new and good: S* = (|Pshared| - |Pcontradictory|) / |Ptotal| is a legitimate extension of set-based similarity. When Pcontradictory is empty, it reduces to Jaccard under the paper's counting. Reflexivity, symmetry and boundedness (Propositions 1-3) follow immediately from the normalization but they are correct. The small comparison with Jaccard on the p1/p2/p3 example is transparent and makes the point that contradictions can flip a positive similarity to negative. Fine.\n\nThe soft spots are serious. The extractor E is defined on a single knowledge base (Definition 4, Eq. 11), returning contradictory literals within that same entity. In Example 2, the author applies E to a pair and writes E(K2) = {toux, ¬toux} even though K2 = {fièvre, ¬toux, maux de tête} contains no positive literal toux. Under the given definition, E(K2) is empty. So the repair step is undefined. And the recomputed similarity after the supposed repair is wrong: for K1 = {fièvre, toux, ¬maux de tête} and K'2 = {fièvre, maux de tête}, you get |Pshared|=1, |Pcontradictory|=1, |Ptotal|=4, so S* = 0, not 0.25 as claimed in Eq. (19). Since the repair mechanism, Proposition 5, Theorem 2, and Corollary 3.1 all lean on this step, the paper's paraconsistent handling of contradictions has no valid demonstration.\n\nLess severe but worth noting: Theorems 2 and 3 are asserted with sketchy proofs that essentially restate the definitions. Section 4.3 concedes that real-data validation remains to be done, yet the conclusion says the framework was \"validated\" by medical and legal examples — those examples are illustrative only.\n\nWho is this for? A reader looking for a quick formula might photocopy the first three propositions. But as a paper, the current version is not reliable. I would desk-reject it in its current form. If the author trims it to a short note on S* and its basic properties, dropping the repair/hierarchy machinery, it could be salvageable.","headline":"A genuine but elementary S* formula surrounded by a broken repair pipeline; reject as is, but salvageable as a short note.","tokens_in":14077,"tokens_out":3917,"would_cite":false,"duration_ms":37751,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["03B53","68T27","68T30"],"pacs":[],"model":"deepseek-v4-flash","headline":"A paraconsistent similarity measure S* subtracts a contradiction ratio from a shared-property ratio, so conflicting knowledge pairs can score negative while staying reflexive and symmetric.","keywords":["similarity measure","paraconsistent logic","knowledge bases","contradiction handling","Jaccard similarity","hierarchical clustering","knowledge repair","multi-agent systems"],"falsifier":"Recompute S*(K1,K'2) for Example 2: K1 = {fièvre, toux, ¬maux de tête} and K'2 = {fièvre, maux de tête}. Shared = {fièvre}; contradictory = {maux de tête}; Ptotal = {fièvre, toux, ¬maux de tête, maux de tête}, so S* = 1/4 − 1/4 = 0, not 0.25 as printed. Also check whether E(K2) for K2 = {fièvre, ¬toux, maux de tête} alone can contain toux; under the single-entity Definition 4 it cannot, so the extraction step used in the example is not reproducible.","tokens_in":13056,"feed_emoji":"⚖️","tokens_out":9354,"duration_ms":101758,"temperature":0.7,"pith_summary":"This paper argues that similarity between knowledge-base entities should account for contradictions rather than ignore them. It defines S*(K1,K2) as the fraction of shared properties minus the fraction of contradictory property pairs, so a pair that shares one trait but clashes on two gets a negative score. The measure is reflexive, symmetric, bounded between -1 and 1, and collapses to a Jaccard-style overlap ratio when no contradictions are present. A contradiction extractor and a minimal-repair mechanism are added so inconsistent entities can be compared after removing only what is needed to restore coherence. The practical stakes: in medicine, law, or multi-agent systems, two descriptions can be superficially close yet logically opposed, and a scalar measure that encodes that opposition is more honest than one that ignores it.","feed_headline":"Similarity formula subtracts contradictions from shared traits","feed_subtitle":"S* = shared ratio minus contradiction ratio, so conflicting pairs can score negative instead of falsely similar.","key_machinery":"The load-bearing object is the pair of ratios S+ and D± sharing the same denominator |Ptotal|. S+ = |Pshared|/|Ptotal| rewards common ground; D± = |Pcontradictory|/|Ptotal| charges for contradictions. Their difference S* turns similarity into a signed quantity: positive means agreement dominates, zero means a tie, and negative means conflict dominates. The shared denominator keeps the two terms comparable, and when Pcontradictory is empty S* becomes a literal-level Jaccard ratio. Supporting machinery includes the contradiction extractor E, which identifies contradictory literals, the minimal repair Rmin, which removes the smallest set needed to restore consistency, and the super-category def","core_discovery":"The paper's discovery is that classical overlap-based similarity can be extended without abandoning its simple set-ratio form: replace 'shared over total' by the difference between the shared-property ratio and the contradictory-property ratio. With Ptotal the set of literals appearing in either entity, Pshared the literals occurring in both, and Pcontradictory the contradictory pairs, S* = |Pshared|/|Ptotal| − |Pcontradictory|/|Ptotal|. The paper proves reflexivity, symmetry, and the [-1,1] bounds, and shows on medical-style examples how negative scores separate logically opposed entities. It also defines paraconsistent super-categories as threshold-based clusters and adds a contradiction e","pith_inferences":["A probabilistic or fuzzy extension of S* could replace exact counts by expected cardinalities, making the same ratio-difference interpretable when property membership is uncertain rather than definite.","In multi-agent systems, the sign of S* could act as a conflict signal: negative scores flag pairs that need repair or negotiation before cooperation, an application the paper names but does not develop in detail.","The threshold-θ hierarchy could drive active data collection: pairs scoring near θ are exactly those whose contradictions most affect the clustering, so querying them first would sharpen the hierarchy at minimal cost."],"forward_implications":["A pair that shares one property and contradicts on two gets S* = −1/5 in the paper's example, so logical conflict can outweigh overlap in the final score.","When no contradictory pair exists, S* equals |Pshared|/|Ptotal|, making the measure backward-compatible with Jaccard-like similarity on consistent knowledge.","S*(K,K) = 1 and S*(K1,K2) = S*(K2,K1), so clustering by thresholds is well-behaved and does not depend on the order in which entities are compared.","After applying minimal repair Rmin, the repaired entities are compared with the same formula, yielding a consistency-restored similarity score.","Super-categories group entities with S* > θ and separate different groups at S* ≤ θ, producing a dynamic hierarchy from a single threshold parameter."],"supporting_citations":[{"why":"Supplies the classical feature-based model of similarity (shared features) that S* extends to signed, contradiction-aware scores.","marker":"[26]"},{"why":"Represents the classical semantic-similarity baseline that ignores contradictions and that the paper argues overestimates similarity.","marker":"[21]"},{"why":"Defines the ontology-alignment and entity-matching setting where conflicting properties must be handled, motivating the contradiction term.","marker":"[23]"},{"why":"Provides a logic oriented to concepts and similarity, cited for the point that a similarity logic must not trivialize under contradictions.","marker":"[22]"},{"why":"Supplies the paraconsistent-logic background of reasoning in inconsistent theories, which licenses comparing entities containing both p and ¬p.","marker":"[20]"},{"why":"Foundational paraconsistent-logic reference for inconsistent formal systems, grounding the non-explosion premise of the framework.","marker":"[7]"},{"why":"Previous framework for similarity-based intelligent databases that the super-category hierarchy and property-selection machinery build on.","marker":"[27]"}],"fun_headline_variants":["Similarity equals shared minus contradictory traits","New similarity measure punishes contradictions","Paraconsistent similarity: reward shared, penalize conflicts","S* = shared ratio - contradiction ratio","Negative scores separate contradictory entities"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The repair mechanism assumes a contradiction extractor that can see contradictions between two entities being compared, but the formal definition only lets it look inside one entity; if it cannot pair literals across entities, the repair step has no defined input.","fun_headline_variants_meta":{"raw":{"variants":["Similarity equals shared minus contradictory traits","New similarity measure punishes contradictions","Paraconsistent similarity: reward shared, penalize conflicts","S* = shared ratio - contradiction ratio","Negative scores separate contradictory entities"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000911,"raw_usage":{"total_tokens":3702,"prompt_tokens":647,"completion_tokens":3055,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":391,"completion_tokens_details":{"reasoning_tokens":2992}},"tokens_in":391,"tokens_out":3055,"duration_ms":24846,"temperature":1.0,"reasoning_tokens":2992,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T20:39:35.897120+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute S*(K1,K'2) for Example 2: K1 = {fièvre, toux, ¬maux de tête} and K'2 = {fièvre, maux de tête}. Shared = {fièvre}; contradictory = {maux de tête}; Ptotal = {fièvre, toux, ¬maux de tête, maux de tête}, so S* = 1/4 − 1/4 = 0, not 0.25 as printed. Also check whether E(K2) for K2 = {fièvre, ¬toux, maux de tête} alone can contain toux; under the single-entity Definition 4 it cannot, so the extraction step used in the example is not reproducible.","supporting_citations":[],"review_version":1}