{"id":"03b9d5ae-c525-4906-8279-be0f54f55e92","arxiv_id":"2608.09748","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Decentralization is defined as the multiplicity of realizations of a specific subject on a graph, separate from distribution, with two new graph metrics for quantifying it.","lead":"This paper proposes a formal, graph-based ontology that defines decentralization as a property of specific subjects, such as data, training, or authority, in a computer communication system, and separates it from mere distribution. It offers a common language and two new metrics, Void Tolerance and Imperviousness, for comparing how decentralized different systems are, which could make claims about decentralized AI or blockchain architectures more precise.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The formal definition of decentralization as δ(p_u)>1 is not graph-relational: δ is an unconstrained analyst-supplied realization count, so the same topology and subject can flip between centralized and decentralized by changing only realization granularity.","rationale":"The paper's central contribution is the δ>1 criterion; everything else, metrics, implementation, and cross-system comparison, hangs off it. If δ can be changed without changing the graph or the declared subject, then the criterion formalizes a modeling choice rather than a discovered structural property. I checked Section IX's axioms and Section XII's Eq. 14: no realization-individuation axiom is stated, and the implementation stores c_u(v) as raw input. This is more specific than the reader's weakest assumption about subject selection: even with subjects fixed, the multiplicity function is not grounded. The reader's formula-level concern about Void Tolerance is real and should be fixed, but it sits in the analytical layer and does not threaten the definitional core as directly as the δ granularity issue. I do not think this forces rejection. The paper explicitly embraces contextual operationalism, so a conditional accept with required revisions is appropriate: add identity criteria for realization particulars, correct the metric formulas, and temper the universality claim so that realization granularity is acknowledged as part of the analyst-supplied modeling layer.","tokens_in":34790,"tokens_out":12747,"duration_ms":80037,"concrete_test":"Run the companion browser tool on a single-vertex topology with one subject and c_u(v)=1, then rerun with c_u(v)=2 (same graph, same support vertex, same edge set, same declared subject). The tool will compute δ=1 vs δ=2 and, under CQ1 and Section XII, classify the first subject CENTRALIZED and the second DECENTRALIZED, both with λ=1. If both outputs are accepted as valid analyses of the same topology, then δ>1 is a modeling input rather than a graph-relational property, and the claimed domain-independent foundation requires an independent realization-individuation rule.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that δ(p_u)>1 gives a formal, graph-relational definition of decentralization (Section IX, CQ1). The load-bearing gap is that δ is not determined by the graph or by any ontology axiom; it is an analyst-supplied count of \"realization particulars.\" In the implementation (Section XII, Eq. 14), δ(p_u)=Σ_v c_u(v), where c_u(v) is a free, per-vertex positive integer entered by the user. The axioms in Section IX only require R_p finite and non-empty; they give no identity criterion for when two realizations are distinct. Consequently, the same system topology and the same declared subject can be classified either way. In Instantiation 2, aggregation authority is called DECENTRALIZED because the authors count three realizations, one at each of v1, v2, v3. Nothing in the ontology prevents an equally faithful model in which the same committee is one joint realization of the subject; that model has δ=1 and is CENTRALIZED under CQ1. Conversely, a single data node can be made DECENTRALIZED by declaring two logical realizations at that vertex (δ=2, λ=1), the paper's own cryptocurrency example. Section XIV.E discloses that subject selection and graph abstraction are analyst-dependent, but it does not address realization granularity; Section XIV.D treats δ as foundational while admitting only the metrics are ad hoc. Thus the claimed universal, domain-independent foundation is not established: the definition reduces to a threshold over a modeling choice, and the Decentralization Problem reappears at the realization-individuation level. This is a load-bearing weakness of the strongest claim, though it is patchable if the ontology supplies realization-identity criteria.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that 'decentralization' lacks a transferable formal definition, and it proposes a four-layer graph-based ontology to fill this gap. The ontological layer introduces systems, topologies, and subjects; the logical layer defines a subject as centralized or decentralized via the realization multiplicity δ(p_u) of its unique projection particular, and as distributed via the vertex support λ(p_u); the modeling and analytical layers bind these notions to graphs and introduce two metrics, Void Tolerance and Imperviousness, together with a browser-based implementation. The paper validates the framework on federated learning and blockchain instantiations and provides proofs of consistency, disjointness, satisfiability, and entailment for the ontology.","tokens_in":35074,"tokens_out":9748,"duration_ms":79002,"significance":"If the framework worked as claimed, it would provide a subject-relative, graph-grounded separation of decentralization from distribution and a basis for comparing heterogeneous systems. The paper has real strengths: the logical-layer proofs are explicit and structurally clear; the paired instantiations usefully demonstrate that the same topology can host different decentralization profiles; the browser implementation with deterministic simulation is a concrete, reusable artifact; and Section XIV is unusually candid about limitations. However, the central definition depends on an unconstrained analyst-supplied realization count, and the main metric formula has a concrete technical inconsistency. The universal, domain-independent claims are therefore not yet established, although the framework may still be salvageable as a conditional modeling tool.","major_comments":[{"comment":"The central dichotomy CENTRALIZED(u) versus DECENTRALIZED(u) reduces to δ(p_u)>1, but δ is not determined by the graph or by the ontology axioms. The axioms in Section IX (Relation to Multiplicity) require only that R_p be finite and non-empty, and Eq. (14) defines δ(p_u)=Σ_v c_u(v), where c_u(v) is a positive integer entered by the user for each vertex. There is no identity criterion for when two realization particulars are distinct. Consequently, the same topology and the same declared subject can be classified either way: the aggregation committee in Instantiation 2 can be modeled as three realizations (δ=3, DECENTRALIZED) or as one joint realization (δ=1, CENTRALIZED) without any change in the graph or the declared subject. Section XIV.D concedes that the metrics are ad hoc but does not address this granularity, and Section XIV.C's claim that δ is the 'furthest context-independent distinction' is unsupported. This is load-bearing for the main contribution, so the paper must either provide an identity criterion for realization particulars or explicitly restrict the definition to a chosen realization model and drop the universal framing.","section":"Section IX, CQ1 and Section XII, Eq. (14)"},{"comment":"The displayed formula for Void Tolerance is inconsistent with the stated range T_L∈[0,1] and with the worked examples. Under the stated conditions |G_t|>|G_s| and ε>0, the denominator |G_s|−ε|G_t| is negative for ε=1, so for any r_v>0 the exponent is positive and T_L>1. In the Blockchain example for u_ledger, the text reports T_L=e^{−8/7}≈0.319 with r_v=2, |G_s|=2, and |G_t|=7, which does not follow from the displayed formula as written. Similar concerns apply to the FL examples, where the reported value ≈0.67 for r_v=1, |G_s|=2, and |G_t|=5 also does not match Eq. (9) as typeset. Please correct the formula, check the boundary cases, and recompute all reported analytical values.","section":"Section XI, Proposition 1, Eq. (9)"},{"comment":"The comparative statements that one system is 'more decentralized' than another depend on the arbitrary choices ε=1 and equal-weight vector averaging. Eq. (9) and Eq. (10) both contain the subject weight ε, and Eq. (21) fixes equal weights; Section XIV.D itself acknowledges that alternative functional forms could produce different numerical scales or comparative orderings. Since Section II (contribution 5) and Section XIII.E present the metrics as enabling direct comparison, the paper should either justify a principled default parameterization or consistently frame the comparative results as illustrative and conditional on the declared weights. As written, the evaluations in Section XI demonstrate internal computational consistency but do not support a general, parameter-independent ordering claim.","section":"Section XI and Section XII, Eq. (21)"}],"minor_comments":[{"comment":"There are several typos that should be corrected: 'singature' (Section VIII), 'Intantiations' (Section IX), 'decentarlized' (Appendix A1), and '100,000systems' (Section XII).","section":"Throughout"},{"comment":"The reference list is inconsistent: many entries provide volume and pages but omit the publication venue or year (e.g., [26], [27], [29]–[31], [34], [36], [38], [39]), and some entries lack author names (e.g., [14]). Please normalize the bibliography to a single style.","section":"Reference list"},{"comment":"The distinction between logical protocol nodes and physical deployment locations is central to the δ vs λ example, but the paper does not specify how an analyst should choose the vertex abstraction between logical and physical levels. A brief modeling guideline would make the example reproducible.","section":"Section IX, cryptocurrency example"},{"comment":"The sentence that the star-topology vulnerability is not detectable by 'any node-centric metric' is overstated, since classical articulation-point and vertex-cut measures are graph-theoretic and depend on the relational structure; please soften this claim.","section":"Section XIII.D"}],"recommendation":"major_revision","confidential_remarks":"I am sympathetic to the project, and the logical-layer material is coherent, but the realization-granularity issue is a genuine gap in the central definition rather than a presentation flaw. The authors could address it by adding an identity criterion for realization particulars or by reframing the contribution as a conditional framework for analyzing decentralization relative to a chosen realization model. The Eq. (9) inconsistency must also be corrected before publication. I would not reject outright if the universal claims are appropriately scoped."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper earns its keep: it cleanly separates decentralization (realization multiplicity δ) from distribution (vertex support λ), builds a description-logic ontology with serviceable consistency/disjointness/satisfiability proofs, and ships a working browser tool with reproducible metric computations. The literature analysis is careful, and the paired FL and blockchain instantiations show the framework can track subject-specific differences that node-counting metrics miss. Second, the central definition is less universal than claimed, and one equation is internally inconsistent.\n\nThe Void Tolerance formula in Proposition 1 (Eq. 9) as printed—with |G_s| − ϵ|G_t| in the denominator—does not reproduce any of the worked examples. The examples match e^{−r_v²|G_s|/(ϵ|G_t|)} with ϵ = 1. The printed form can also produce values above 1 when ϵ|G_t| > |G_s|. This looks like a typo, but it sits at the top of the analytical layer and must be fixed before the paper is defensible.\n\nThe deeper issue is that δ(p_u) is not graph-grounded. The ontology requires R_p finite and non-empty but gives no identity criterion for when two realization particulars are distinct. In the implementation, c_u(v) is a user-entered per-vertex count. So the same committee can be modeled as three realizations (decentralized, δ=3) or one joint realization (centralized, δ=1), and nothing in the ontology says one is wrong. The paper discloses in Section XIV.E that subject choice and graph abstraction are analyst-dependent; it does not disclose, let alone constrain, realization granularity. That makes the headline claim—a domain-independent, formal definition—conditional on an unstated modeling convention. The fix is straightforward: add realization-individuation criteria, or explicitly scope the definition to a declared realization model. As written, though, this is a load-bearing gap.\n\nThe metrics are honestly labelled ad hoc in Section XIV.D, and the ϵ and equal-averaging choices do affect the comparative ‘more decentralized’ ordering. That is a limitation, not a hidden flaw, but it means the analytical comparisons are illustrations, not theorems.\n\nBottom line: this deserves a serious referee and, likely, conditional acceptance after the formula is corrected and the realization-identity problem is addressed. It is a thought-provoking foundation for the decentralization literature, but not yet the universal definition it claims.","headline":"A genuinely useful δ/λ separation and a careful literature dissection, but the central definition rests on an unconstrained analyst-chosen realization count and the headline metric equation contradicts its own worked examples.","tokens_in":35660,"tokens_out":5919,"would_cite":false,"duration_ms":57836,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper defines decentralization as a subject-specific relational property: a subject is decentralized exactly when its projection multiplicity exceeds one, and distributed exactly when its realizations span more than one vertex.","keywords":["decentralization","ontology","graph representation","formal semantics","federated learning","blockchain","distribution vs decentralization","decentralization metrics"],"falsifier":"If two analysts choose equally defensible but different subject sets for the same graph, the framework returns different system-level classifications; observing such a disagreement in practice would show that the definition is not actually independent of the analyst's modeling choices.","tokens_in":34545,"feed_emoji":"🌐","tokens_out":8250,"duration_ms":65225,"temperature":0.7,"pith_summary":"Decentralization has been used for half a century in computer science without a transferable definition, so the same architecture can be called centralized in one paper and decentralized in another. The paper proposes to close this gap with an ontology for graph-representable communication systems in which decentralization is always decentralization of some subject—data, model training, aggregation authority, ledger replication—rather than a property of a whole system. A subject is decentralized when its projection multiplicity is greater than one, and distributed when its realizations occupy more than one vertex, with the two properties formally separated. The framework derives system-level categories (centralized, partially decentralized, fully decentralized), two graph metrics (Void Tolerance and Imperviousness), and instantiates them on federated learning and blockchain architectures to show consistent, comparable assessments.","feed_headline":"Decentralization gets a formal graph-based definition","feed_subtitle":"A new ontology separates decentralization from distribution and classifies federated learning and blockchain systems","key_machinery":"The load-bearing object is the projection particular $p_u$, an individual in the ontology that binds a system $s$, a topology $T$, and a subject $u$ through the relations hasProjection, ofSubject, and inTopology. Each projection carries a finite set of realization particulars, mapped by realizedAt to vertices of the graph; $\\delta(p_u)$ counts realizations and $\\lambda(p_u)$ counts the distinct supporting vertices, and the axiom $1 \\leq \\lambda(p_u) \\leq \\delta(p_u)$ keeps the two notions ordered. The logical layer axiomatizes these relations and derives the centralized/decentralized and distributed/undistributed predicates from $\\delta$ and $\\lambda$. The analytical layer adds two subject-specific graph metrics—Void Tolerance, which measures vertex-removal resilience of the subject-induced subgraph, and Imperviousness, which measures the edge-deletion effort needed to isolate a subject-supporting vertex—combined into a per-subject decentralization vector in $[0,1]^2$.","core_discovery":"The central claim is that decentralization can be defined without reference to any application domain: fix a graph-representable system, choose a subject of decentralization, and let $p_u$ be the unique projection particular linking the system, its topology, and that subject. The subject is decentralized iff the number of distinct realization particulars $\\delta(p_u)$ is greater than one; it is distributed iff the number of vertices supporting those realizations $\\lambda(p_u)$ is greater than one. Centralization is the degenerate zero-dimensional case in which every projection has multiplicity exactly one, and a system's dimensionality is the count of its decentralized subjects. Full, partial, and centralized system classes follow from whether all, some, or none of the declared subjects are decentralized, so the same topology can legitimately yield different classifications for different subjects. The paper argues this resolves the Decentralization Problem by making the previously implicit choice of dimension explicit and formal.","pith_inferences":["Beyond the paper, the framework turns subject selection into a reporting requirement: any decentralization claim becomes comparable only after the declared subjects and weights are stated, which the paper gestures at but does not develop into a standard.","The snapshot ontology suggests a natural extension to temporal decentralization, tracking how $\\delta(p_u)$ and $\\lambda(p_u)$ change across protocol phases, which the paper treats only as separate static topologies.","Because the exponential forms of the two metrics are acknowledged as ad hoc, replacing them with alternative functions over the same $r_v$, $r_e$, $\\lambda$, and $\\delta$ quantities could change comparative orderings; the numbers are not natural constants.","The $\\delta$-versus-$\\lambda$ distinction yields a testable prediction that node-counting proxies miss: two systems with identical node counts can differ in decentralization if one co-locates multiple subject realizations on single vertices."],"forward_implications":["Vanilla federated learning with one aggregation server is classified as partially decentralized: data and training are decentralized across clients, but aggregation authority is centralized.","A blockchain with a single block-ordering node but replicated ledger and client-side transaction submission is partially decentralized; re-assigning consensus to the core nodes makes it fully decentralized under the framework.","A subject can be decentralized and undistributed at the same time when several realizations sit on one vertex, so decentralization does not imply distribution.","Identical graph topologies can have different decentralization profiles when the same functions are assigned to different nodes, so topology alone does not determine decentralization.","Systems can be ordered as more or less decentralized by comparing the lengths of their averaged subject vectors, provided the declared subject sets are identical."],"supporting_citations":[{"why":"Supplies the early structural typology of centralized, decentralized, and distributed networks that the paper extends into a formal graph-based definition.","marker":"[1]"},{"why":"Provides the vanilla federated learning architecture used as the first instantiation and the motivating contrast with the structural view.","marker":"[2]"},{"why":"Gives the multidimensional treatment of decentralization and the distribution/decentralization distinction that the paper formalizes.","marker":"[17]"},{"why":"Supplies the formal-semantics account of multidimensional gradable adjectives used to model decentralized as dimension-relative.","marker":"[21]"},{"why":"Provides the notion of ontological commitment and formal ontology that the paper's layered definition follows.","marker":"[41]"},{"why":"Supplies the competency-question methodology used to constrain and evaluate the ontology.","marker":"[46]"}],"fun_headline_variants":["Decentralization gets a formal graph-based ontology","New metrics quantify decentralization in any system","Graph definition separates decentralization from distribution","Ontology resolves the decentralization problem for AI and blockchains","Decentralization defined as a subject-specific graph property"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework's classifications depend on the analyst's choice of subjects of decentralization and on the graph abstraction; different analysts may select different subjects, and the same topology can be classified fully, partially, or non-decentralized depending on that choice.","fun_headline_variants_meta":{"raw":{"variants":["Decentralization gets a formal graph-based ontology","New metrics quantify decentralization in any system","Graph definition separates decentralization from distribution","Ontology resolves the decentralization problem for AI and blockchains","Decentralization defined as a subject-specific graph property"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000318,"raw_usage":{"total_tokens":1805,"prompt_tokens":962,"completion_tokens":843,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":774}},"tokens_in":578,"tokens_out":843,"duration_ms":5549,"temperature":1.0,"reasoning_tokens":774,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:31:44.789215+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If two analysts choose equally defensible but different subject sets for the same graph, the framework returns different system-level classifications; observing such a disagreement in practice would show that the definition is not actually independent of the analyst's modeling choices.","supporting_citations":[{"cited_title":"On distributed communications: I. introduction to distributed communications networks","cited_arxiv_id":null,"evidence_quote":"Supplies the early structural typology of centralized, decentralized, and distributed networks that the paper extends into a formal graph-based definition."},{"cited_title":"Towards a theory of digital network de/centralization: Platform-infrastructure lessons drawn from blockchain","cited_arxiv_id":null,"evidence_quote":"Gives the multidimensional treatment of decentralization and the distribution/decentralization distinction that the paper formalizes."},{"cited_title":"A typology of multidimensional adjectives,","cited_arxiv_id":null,"evidence_quote":"Supplies the formal-semantics account of multidimensional gradable adjectives used to model decentralized as dimension-relative."},{"cited_title":"Formal ontology and information systems,","cited_arxiv_id":null,"evidence_quote":"Provides the notion of ontological commitment and formal ontology that the paper's layered definition follows."},{"cited_title":"Methodology for the design and evaluation of ontologies,","cited_arxiv_id":null,"evidence_quote":"Supplies the competency-question methodology used to constrain and evaluate the ontology."}],"review_version":1}