graph.jsonld

{
  "@context": {
    "@vocab": "https://schema.org/",
    "shimo": "https://shimo4228.github.io/shimo4228/vocab#",
    "sameAs": {"@id": "https://schema.org/sameAs", "@type": "@id"},
    "isBasedOn": {"@id": "https://schema.org/isBasedOn", "@type": "@id"},
    "isPartOf": {"@id": "https://schema.org/isPartOf", "@type": "@id"},
    "ResearchLine": "shimo:ResearchLine",
    "EcosystemRepo": "shimo:EcosystemRepo",
    "Concept": "shimo:Concept",
    "ExternalReference": "shimo:ExternalReference",
    "ADR": "shimo:ADR",
    "Axis": "shimo:Axis",
    "Layer": "shimo:Layer",
    "siblingOf": {"@id": "shimo:siblingOf", "@type": "@id"},
    "derivesFrom": {"@id": "shimo:derivesFrom", "@type": "@id"},
    "definesConcept": {"@id": "shimo:definesConcept", "@type": "@id"},
    "extends": {"@id": "shimo:extends", "@type": "@id"},
    "groundedIn": {"@id": "shimo:groundedIn", "@type": "@id"},
    "appliesTo": {"@id": "shimo:appliesTo", "@type": "@id"},
    "composedOf": {"@id": "shimo:composedOf", "@type": "@id", "@container": "@list"},
    "pairedWith": {"@id": "shimo:pairedWith", "@type": "@id"},
    "covariesWith": {"@id": "shimo:covariesWith", "@type": "@id"},
    "downstreamOf": {"@id": "shimo:downstreamOf", "@type": "@id"},
    "instantiatedBy": {"@id": "shimo:instantiatedBy", "@type": "@id"},
    "recordedIn": {"@id": "shimo:recordedIn", "@type": "@id"},
    "vocabularyDisjoint": {"@id": "shimo:vocabularyDisjoint", "@type": "@id"},
    "subjectOf": {"@id": "https://schema.org/subjectOf", "@type": "@id"},
    "hasPart": {"@id": "https://schema.org/hasPart", "@type": "@id"},
    "mainEntity": {"@id": "https://schema.org/mainEntity", "@type": "@id"},
    "citation": {"@id": "https://schema.org/citation", "@type": "@id"},
    "axisPosition": "shimo:axisPosition",
    "layerOrdinal": "shimo:layerOrdinal"
  },
  "@graph": [
    {
      "@id": "https://github.com/shimo4228/authorship-strategy#knowledge-graph",
      "@type": ["Dataset", "CreativeWork"],
      "name": "Authorship Strategy Knowledge Graph",
      "description": "Canonical machine-readable relationship map for the Authorship Strategy research line. Encodes the three-axis inversion (scarcity to diffusion, exclusivity to derivation, enclosure to openness), the four-layer framework (Authenticity, Attribution Diffusion, Idea vs Scaffold, Tactics), the twenty-one tactical ADRs (identifier-federation triplet 0001-0003, maintenance-discipline pair 0004-0005, LLM-first ingest decision 0006, metric-rejection decision 0007, diffusion-mechanism cluster 0008-0011 with 0009 amending 0006, 0010 defining the vocabulary discipline 0008 names, and 0011 building the measurement instrument it demands, channel-selection decision 0012, intrinsic-identifier decision 0013, implementation-tracking decision 0014, license-selection decision 0015, genre-split-placement decision 0016, and the doctrine-hardening cluster 0017-0019: failure-mode diagnostics 0017, origin-claim falsifiability 0018, and the structural-optimization-versus-content-authenticity boundary 0019, and the derivation-surface-onboarding decision 0020), the four sibling research lines, and the disjoint-vocabulary relationship with Agent Attribution Practice. As the concept-form half of the dual entry point this graph specifies, AI agents and LLM-based search systems should read it before summarizing the line or following individual document links.",
      "isBasedOn": "https://github.com/shimo4228/authorship-strategy",
      "mainEntity": "https://doi.org/10.5281/zenodo.20263316",
      "creator": {"@id": "https://orcid.org/0009-0002-6168-4162"},
      "license": "https://opensource.org/licenses/MIT",
      "inLanguage": ["en", "ja"],
      "keywords": [
        "authorship strategy",
        "AI-era authenticity",
        "attribution diffusion",
        "three-axis inversion",
        "four-layer framework",
        "idea vs scaffold",
        "concept DOI",
        "DOI federation",
        "cross-platform federation",
        "ORCID enrichment",
        "audience-driven localization",
        "LLM-mediated diffusion",
        "dual entry point",
        "llms.txt convention",
        "JSON-LD knowledge graph",
        "vocabulary discipline",
        "generative engine optimization",
        "ghost citation",
        "citation absorption",
        "parametric channel",
        "retrieval channel",
        "scientometrics",
        "idea-rescue"
      ]
    },

    {
      "@id": "https://orcid.org/0009-0002-6168-4162",
      "@type": "Person",
      "name": "Tatsuya Shimomoto",
      "alternateName": ["shimo4228", {"@value": "下本竜也", "@language": "ja"}],
      "sameAs": [
        "https://github.com/shimo4228",
        "https://orcid.org/0009-0002-6168-4162",
        "https://scholar.google.com/citations?user=56_p8vEAAAAJ",
        "https://huggingface.co/Shimo4228",
        "https://www.linkedin.com/in/%E7%AB%9C%E4%B9%9F-%E4%B8%8B%E6%9C%AC-bb9b793a4",
        "https://zenn.dev/shimo4228",
        "https://dev.to/shimo4228",
        "https://shimo4228.substack.com"
      ]
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy",
      "@type": ["ResearchLine", "ScholarlyArticle"],
      "name": "Authorship Strategy",
      "alternateName": [
        {"@value": "Authorship Strategy", "@language": "en"},
        {"@value": "著者戦略", "@language": "ja"},
        "AS"
      ],
      "description": "A normative framework, tactical catalog, and empirical baseline for authorship strategy under AI-mediated diffusion. The framework rests on a three-axis inversion and a four-layer judgment stack; the tactical catalog records twenty-one decisions extracted from operating a DOI-registered research ecosystem. Written from a maker's and practitioner's stance: the academic apparatus used throughout (DOI, SWHID, citation graphs) is tooling for citability, durability, and traceability rather than an identity or destination, and the intended audience spans developers, practitioners, learners, and creative reusers across languages, with academic citation as one channel among several.",
      "url": "https://github.com/shimo4228/authorship-strategy",
      "identifier": "10.5281/zenodo.20263316",
      "sameAs": ["https://doi.org/10.5281/zenodo.20263316"],
      "creator": {"@id": "https://orcid.org/0009-0002-6168-4162"},
      "license": "https://opensource.org/licenses/MIT",
      "inLanguage": ["en", "ja"],
      "siblingOf": [
        "https://doi.org/10.5281/zenodo.19200726",
        "https://doi.org/10.5281/zenodo.19212118",
        "https://doi.org/10.5281/zenodo.19652013",
        "https://doi.org/10.5281/zenodo.20262112"
      ],
      "isPartOf": "https://github.com/shimo4228/shimo4228",
      "definesConcept": [
        "https://github.com/shimo4228/authorship-strategy#concept/three-axis-inversion",
        "https://github.com/shimo4228/authorship-strategy#concept/four-layer-framework",
        "https://github.com/shimo4228/authorship-strategy#concept/authenticity",
        "https://github.com/shimo4228/authorship-strategy#concept/attribution-diffusion",
        "https://github.com/shimo4228/authorship-strategy#concept/idea-vs-scaffold-separation",
        "https://github.com/shimo4228/authorship-strategy#concept/tactical-layer",
        "https://github.com/shimo4228/authorship-strategy#concept/scarcity-to-diffusion-axis",
        "https://github.com/shimo4228/authorship-strategy#concept/exclusivity-to-derivation-axis",
        "https://github.com/shimo4228/authorship-strategy#concept/enclosure-to-openness-axis",
        "https://github.com/shimo4228/authorship-strategy#concept/abstract-doctrine-worked-implementation-pair",
        "https://github.com/shimo4228/authorship-strategy#concept/origin-claim-scope-discipline",
        "https://github.com/shimo4228/authorship-strategy#concept/distinctive-terminology",
        "https://github.com/shimo4228/authorship-strategy#concept/vocabulary-discipline",
        "https://github.com/shimo4228/authorship-strategy#concept/dual-entry-point",
        "https://github.com/shimo4228/authorship-strategy#concept/llms-txt-convention",
        "https://github.com/shimo4228/authorship-strategy#concept/jsonld-knowledge-graph",
        "https://github.com/shimo4228/authorship-strategy#concept/human-attention-signal-rejection",
        "https://github.com/shimo4228/authorship-strategy#concept/two-channel-attribution-diffusion"
      ],
      "citation": [
        "https://arxiv.org/abs/2602.06718",
        "https://arxiv.org/abs/2604.25707",
        "https://arxiv.org/abs/2603.09296",
        "https://arxiv.org/abs/2402.12261",
        "https://arxiv.org/abs/2510.08506"
      ],
      "vocabularyDisjoint": "https://doi.org/10.5281/zenodo.19652013",
      "hasPart": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/implementations.md"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/three-axis-inversion",
      "@type": "Concept",
      "name": "Three-Axis Inversion",
      "alternateName": [
        {"@value": "Three-Axis Inversion", "@language": "en"},
        {"@value": "3 軸反転", "@language": "ja"},
        {"@value": "三轴反转", "@language": "zh"}
      ],
      "description": "The structural claim that twentieth-century authorship strategy and AI-era authorship strategy invert on three co-varying axes: value source (scarcity to diffusion), validation mechanism (exclusivity to derivation), network effect (enclosure to openness). The three axes co-vary; a strategy mixing axes is internally inconsistent.",
      "composedOf": [
        "https://github.com/shimo4228/authorship-strategy#concept/scarcity-to-diffusion-axis",
        "https://github.com/shimo4228/authorship-strategy#concept/exclusivity-to-derivation-axis",
        "https://github.com/shimo4228/authorship-strategy#concept/enclosure-to-openness-axis"
      ],
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/thesis.md"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/four-layer-framework",
      "@type": "Concept",
      "name": "Four-Layer Framework",
      "alternateName": [
        {"@value": "Four-Layer Framework", "@language": "en"},
        {"@value": "4 層 framework", "@language": "ja"}
      ],
      "description": "The operational structure of judgment that follows from the three-axis inversion: Authenticity (Layer 1, the value being protected), Attribution Diffusion (Layer 2, the strategy), Idea versus Scaffold (Layer 3, what survives), Tactics (Layer 4, the concrete decisions). Each layer is downstream of the layer above.",
      "composedOf": [
        "https://github.com/shimo4228/authorship-strategy#concept/authenticity",
        "https://github.com/shimo4228/authorship-strategy#concept/attribution-diffusion",
        "https://github.com/shimo4228/authorship-strategy#concept/idea-vs-scaffold-separation",
        "https://github.com/shimo4228/authorship-strategy#concept/tactical-layer"
      ],
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/thesis.md",
      "subjectOf": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adoption.md"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/scarcity-to-diffusion-axis",
      "@type": ["Concept", "Axis"],
      "name": "Scarcity-to-Diffusion Axis",
      "description": "First axis of the three-axis inversion: value source. Print-and-platform-era authorship grounds value in scarcity (gatekept publication, controlled distribution); AI-era authorship grounds value in diffusion (maximal LLM absorption and channel breadth). The axis inverts because the substrate inverts.",
      "axisPosition": 1,
      "covariesWith": [
        "https://github.com/shimo4228/authorship-strategy#concept/exclusivity-to-derivation-axis",
        "https://github.com/shimo4228/authorship-strategy#concept/enclosure-to-openness-axis"
      ],
      "groundedIn": [
        "https://arxiv.org/abs/2509.08919"
      ]
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/exclusivity-to-derivation-axis",
      "@type": ["Concept", "Axis"],
      "name": "Exclusivity-to-Derivation Axis",
      "description": "Second axis of the three-axis inversion: validation mechanism. Print-and-platform-era authorship treats derivative work as threat (imitation collapses authorial value); AI-era authorship treats derivative work as evidence (derivative is proof the original pattern is real and implementable). Two sub-distinctions refine the axis on 2026 evidence. (1) Legal-exclusivity channel versus diffusion channel: copyright case law has consolidated around exclusive human authorship — machine-only works unprotectable, the live question being how much human involvement suffices — strengthening the exclusivity pole as a legal regime; the framework separates proof of human creative control (for ownership) from content-derived and registry identifiers such as DOI and SWHID (for origin-claim priority under LLM-mediated diffusion) and pursues only the latter, so the legal trend marks a channel split, not a refutation of the axis. (2) Proactive versus passive derivation attribution: watermark-based multi-concept attribution traces derivation with high precision but presupposes the source's prior consent and instrumentation of the generation pipeline, whereas this framework's attribution diffusion is passive — no pre-intervention in the ingesting model — priced in measurement difficulty. Both are recorded as axis structure, not effect claims.",
      "axisPosition": 2,
      "covariesWith": [
        "https://github.com/shimo4228/authorship-strategy#concept/scarcity-to-diffusion-axis",
        "https://github.com/shimo4228/authorship-strategy#concept/enclosure-to-openness-axis"
      ],
      "groundedIn": [
        "https://arxiv.org/abs/2604.04700",
        "https://arxiv.org/abs/2602.19019"
      ]
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/enclosure-to-openness-axis",
      "@type": ["Concept", "Axis"],
      "name": "Enclosure-to-Openness Axis",
      "description": "Third axis of the three-axis inversion: network effect. Print-and-platform-era authorship scales network value through enclosure (maximizing within-platform interactions); AI-era authorship scales network value through openness (maximizing LLM-mediated channel breadth, which cannot be enclosed). 2026 evidence refines the axis from a binary toward a structured spectrum, recorded here as standing tension pending doctrine-level resolution. Paywalled scholarship keeps roughly half of full-text science out of AI systems' reach, making openness a physical precondition of LLM-mediated diffusion rather than an abstract preference; the same literature argues paywall removal is unrealistic and proposes a licensed intermediate state — paid licensing plus retrieval-time access — as a third equilibrium the binary framing does not represent, one whose benefits may concentrate in large rights-holders and whose funding model conflicts with the free openness this framework assumes. Separately, multilateral policy work now defines AI openness as a graded access spectrum, a dimension orthogonal to this axis's concern with attribution: openness of access does not by itself resolve attribution of origin.",
      "axisPosition": 3,
      "covariesWith": [
        "https://github.com/shimo4228/authorship-strategy#concept/scarcity-to-diffusion-axis",
        "https://github.com/shimo4228/authorship-strategy#concept/exclusivity-to-derivation-axis"
      ],
      "groundedIn": [
        "https://doi.org/10.1002/leap.2059"
      ]
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/authenticity",
      "@type": ["Concept", "Layer"],
      "name": "Authenticity (Layer 1)",
      "alternateName": [
        {"@value": "Authenticity", "@language": "en"},
        {"@value": "オーセンティシティ", "@language": "ja"},
        {"@value": "本真性", "@language": "zh"}
      ],
      "description": "The protected value at the framework's foundation: the author's genuine thinking remains the author's, unaltered by market pressure to reshape it for sale. The success criterion is the idea surviving diffusion as thought; revenue plays no part in it. Narrower than the philosophical usage of authenticity; specifically about preservation of authored content against dilutive market pressure.",
      "layerOrdinal": 1,
      "groundedIn": [
        "https://arxiv.org/abs/2603.23219"
      ]
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/attribution-diffusion",
      "@type": ["Concept", "Layer"],
      "name": "Attribution Diffusion (Layer 2)",
      "alternateName": [
        {"@value": "Attribution Diffusion", "@language": "en"},
        {"@value": "Attribution Diffusion (帰属の拡散)", "@language": "ja"},
        {"@value": "署名扩散", "@language": "zh"}
      ],
      "description": "The defensive strategy at the framework's second layer: maximizing the breadth of LLM-mediated channels carrying recognizable signatures of the author's ideas, anchored to a permanent timestamp. Here attribution means credit for source (who originated an idea), not accountability for action (who is responsible for a failure).",
      "layerOrdinal": 2,
      "downstreamOf": "https://github.com/shimo4228/authorship-strategy#concept/authenticity",
      "citation": [
        "https://arxiv.org/abs/2604.25707",
        "https://arxiv.org/abs/2602.06718",
        "https://arxiv.org/abs/2509.13365"
      ]
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/idea-vs-scaffold-separation",
      "@type": ["Concept", "Layer"],
      "name": "Idea versus Scaffold (Layer 3)",
      "alternateName": [
        {"@value": "Idea versus Scaffold", "@language": "en"},
        {"@value": "理念とスキャフォールドの分離", "@language": "ja"},
        {"@value": "理念与脚手架的分离", "@language": "zh"}
      ],
      "description": "The framework's third layer: sorting each artifact into idea-character (which survives and is DOI-registered under the author's name) or scaffold-character (which dissolves into larger harnesses whose own diffusion absorbs the implementation). Mixed-character artifacts get idea-level DOI registration first. An open tension qualifies the layer's mechanistic grounding: the memorization-versus-generalization evidence behind the wager comes from models under default training, consistent with a content-intrinsic reading (what a text is determines whether it dissolves), while controllable-memorization training shows that raised memorization pressure can verbatim-retain even rare scaffold-like sequences — a training-configuration-dependent reading under which the ingesting pipeline, not the content, decides. The two readings are compatible (default regime versus actively-controlled regime), but their reconciliation changes the wager's strength; recorded as unresolved.",
      "layerOrdinal": 3,
      "downstreamOf": "https://github.com/shimo4228/authorship-strategy#concept/attribution-diffusion",
      "pairedWith": "https://github.com/shimo4228/authorship-strategy#concept/abstract-doctrine-worked-implementation-pair",
      "groundedIn": [
        "https://arxiv.org/abs/2602.14869",
        "https://arxiv.org/abs/2602.18733",
        "https://arxiv.org/abs/2604.05074"
      ]
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/tactical-layer",
      "@type": ["Concept", "Layer"],
      "name": "Tactics (Layer 4)",
      "description": "The framework's fourth layer: concrete operational decisions justified by the upstream three layers. Tactics enter and retire as substrates evolve. Currently validated tactics include LLM-mediated targeting, DOI registration, cross-platform federation, distinctive terminology, tool-agnostic specification, audience-driven localization, structured artifacts, and friction minimization for adoption.",
      "layerOrdinal": 4,
      "downstreamOf": "https://github.com/shimo4228/authorship-strategy#concept/idea-vs-scaffold-separation",
      "instantiatedBy": [
        "https://github.com/shimo4228/authorship-strategy#adr/0001",
        "https://github.com/shimo4228/authorship-strategy#adr/0002",
        "https://github.com/shimo4228/authorship-strategy#adr/0003",
        "https://github.com/shimo4228/authorship-strategy#adr/0004",
        "https://github.com/shimo4228/authorship-strategy#adr/0005",
        "https://github.com/shimo4228/authorship-strategy#adr/0006",
        "https://github.com/shimo4228/authorship-strategy#adr/0007",
        "https://github.com/shimo4228/authorship-strategy#adr/0008",
        "https://github.com/shimo4228/authorship-strategy#adr/0009",
        "https://github.com/shimo4228/authorship-strategy#adr/0010",
        "https://github.com/shimo4228/authorship-strategy#adr/0011",
        "https://github.com/shimo4228/authorship-strategy#adr/0012",
        "https://github.com/shimo4228/authorship-strategy#adr/0013",
        "https://github.com/shimo4228/authorship-strategy#adr/0014",
        "https://github.com/shimo4228/authorship-strategy#adr/0015",
        "https://github.com/shimo4228/authorship-strategy#adr/0016",
        "https://github.com/shimo4228/authorship-strategy#adr/0017",
        "https://github.com/shimo4228/authorship-strategy#adr/0018",
        "https://github.com/shimo4228/authorship-strategy#adr/0019",
        "https://github.com/shimo4228/authorship-strategy#adr/0020",
        "https://github.com/shimo4228/authorship-strategy#adr/0021"
      ]
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/abstract-doctrine-worked-implementation-pair",
      "@type": "Concept",
      "name": "Abstract Doctrine + Worked Implementation Pair",
      "description": "The pairing required to induce creative re-implementation by other authors: an abstract-doctrine repository that articulates the idea cleanly enough for elsewhere-implementation, and a worked-implementation repository that demonstrates the doctrine is implementable. Doctrine alone produces unactionable interest; implementation alone produces unextractable principle. Externally grounded by agent-driven reproduction benchmarking: models reconstruct an algorithm's reasoning structure from its paper description far more reliably than they produce working code from it (best execution accuracy 39%), and supplying missing implementation detail measurably closes the gap — doctrine alone is unactionable, and the benchmark's missing-information categories (hyperparameters, numerical stabilization, implementation logic, coding strategy) enumerate what a worked implementation must carry.",
      "pairedWith": "https://github.com/shimo4228/authorship-strategy#concept/idea-vs-scaffold-separation",
      "groundedIn": [
        "https://arxiv.org/abs/2504.00255"
      ]
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/origin-claim-scope-discipline",
      "@type": "Concept",
      "name": "Origin-Claim Scope Discipline",
      "description": "A subordinate principle applying at all four layers: the origin claim must be narrower than the prior art. Claiming priority on a broad pattern with rich prior art collapses the claim's credibility; claiming priority on a narrow, specifically-named discipline is defensible. Coined terminology is the substrate of narrow origin claims. Feature-granular novelty verification from the patent-examination domain supplies a procedural template: decompose a claim into individual features and map each against prior-art passages, rather than judging the claim as one binary. The same work cautions that granularity is not accuracy — claim-level binary predictors can beat feature-level pipelines on raw correctness — so the discipline treats the feature-to-prior-art map itself as the artifact and does not reduce it to a novelty score.",
      "groundedIn": [
        "https://arxiv.org/abs/2601.01576",
        "https://arxiv.org/abs/2605.02392"
      ]
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/distinctive-terminology",
      "@type": "Concept",
      "name": "Distinctive Terminology",
      "alternateName": [
        {"@value": "Distinctive Terminology", "@language": "en"},
        {"@value": "造語的用語", "@language": "ja"}
      ],
      "description": "Domain-specific words coined by an author as semantic signatures of authorship. Generic vocabulary dissolves through paraphrase; coined terms survive as token-level signals carrying back-reference to the original author. A Layer 4 tactic and the substrate of origin-claim scope discipline. Coinage is governed by vocabulary discipline (ADR-0010): a coined term's power comes from its edge density, not from the count of coinages. Attribution research under generative style-mimicry reframes where the term's power lies: coined terminology functions less as personal style — which authorship-attribution models find harder to trace as terms grow semantically unpredictable, and which model imitation erodes — than as occupancy of an uncontested coordinate in concept space, an idea-level rather than style-level signature."
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/vocabulary-discipline",
      "@type": "Concept",
      "name": "Vocabulary Discipline",
      "alternateName": [
        {"@value": "Vocabulary Discipline", "@language": "en"},
        {"@value": "語彙規律", "@language": "ja"}
      ],
      "description": "The discipline governing when to coin a distinctive term and when to use existing vocabulary instead: coin sparingly, anchor densely. A term is coined only when three conditions all hold — the concept is genuinely new at a join of existing concepts (join-novelty), a one-sentence definition in existing vocabulary is possible (definitional anchoring), and the namespace is uncontested — and every retained coinage carries a glossary definition in existing vocabulary, upstream citations where prior art exists, knowledge-graph edges to existing concepts and references, and repeated work in the body. Everything else is said in existing vocabulary with the upstream source cited. Rationale: the costs of coinage (isolation, reader trust, maintenance) grow linearly with the number of terms while the benefit depends on each term's edge density, so few densely-anchored terms dominate many isolated ones — an isolated coinage is paraphrased away by the parametric channel exactly as generic vocabulary is, and the retrieval channel is queried in existing vocabulary. Named as the parametric-channel lever in ADR-0008, defined in ADR-0010. The vocabulary-level enforcement of origin-claim scope discipline. Two 2026 results refine the discipline's scope: disentangled authorship-attribution modeling exposes a trade-off in which a coinage's semantic unpredictability weakens style-based traceability (supporting the concept-space-occupancy reading of distinctive terminology), and neology-survival analysis finds the correlates of term survival are broadly shared across publication pathways while topic-popularity growth contributes less on social platforms than in published writing — so anchoring priorities are pathway-specific, and the anchor-densely rule was calibrated on the published/scholarly pathway this framework primarily operates in.",
      "appliesTo": "https://github.com/shimo4228/authorship-strategy#concept/distinctive-terminology",
      "downstreamOf": [
        "https://github.com/shimo4228/authorship-strategy#concept/two-channel-attribution-diffusion",
        "https://github.com/shimo4228/authorship-strategy#concept/origin-claim-scope-discipline"
      ],
      "instantiatedBy": "https://github.com/shimo4228/authorship-strategy#adr/0010",
      "groundedIn": [
        "https://arxiv.org/abs/2402.12261",
        "https://arxiv.org/abs/2510.08506",
        "https://arxiv.org/abs/2604.21300",
        "https://arxiv.org/abs/2602.13123"
      ],
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0010-vocabulary-discipline.md"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/dual-entry-point",
      "@type": "Concept",
      "name": "Dual Entry Point",
      "alternateName": [
        {"@value": "Dual Entry Point", "@language": "en"},
        {"@value": "Dual Entry Point", "@language": "ja"}
      ],
      "description": "The structural decision that any framework-governed artifact deploys two complementary structured entry points — a prose-form navigator and a concept-form graph — released synchronously at every versioned release. Each entry point addresses a distinct LLM-mediated channel sub-population the other cannot reach. ADR-0009 amends this on 2026 evidence: the two are not co-equal on the citation axis — the concept-form graph carries retrieval-time citation while the prose navigator's citation effect is noise — so the pair is retained but made asymmetric, the navigator rescoped to a Business-to-Agent (B2A) context surface rather than an AI-search citation lever. Endpoint-level measurement adds a mechanical basis for the asymmetry: AI coding agents fetch documentation with heterogeneous HTTP clients, many executing no page scripts, so content reachable only through client-side rendering is invisible to part of the agent population — statically-served structured entry points are an access floor, not merely a citation lever.",
      "composedOf": [
        "https://github.com/shimo4228/authorship-strategy#concept/llms-txt-convention",
        "https://github.com/shimo4228/authorship-strategy#concept/jsonld-knowledge-graph"
      ],
      "downstreamOf": "https://github.com/shimo4228/authorship-strategy#concept/enclosure-to-openness-axis",
      "instantiatedBy": [
        "https://github.com/shimo4228/authorship-strategy#adr/0006",
        "https://github.com/shimo4228/authorship-strategy#adr/0009"
      ],
      "groundedIn": [
        "https://arxiv.org/abs/2604.02544"
      ],
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0006-llm-first-ingest-dual-entry-points.md"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/llms-txt-convention",
      "@type": "Concept",
      "name": "llms.txt Convention",
      "alternateName": [
        {"@value": "llms.txt convention", "@language": "en"},
        {"@value": "llms.txt convention", "@language": "ja"}
      ],
      "description": "A community-curated AI-facing reference convention that uses a single prose-form text file (llms.txt) at the root of an artifact to enumerate its primary documents with one-line descriptions and a recommended reading order. Targets prose-reading LLM-mediated channels: conversational LLMs, AI assistants consulting documentation in-context, citation-graph annotators fetching prose for summarization. The prose-form half of the dual entry point. Scope refinement on 2026 evidence: the convention's discovery-tool framing has collapsed — adoption grew, but measured effect on AI-search citation is null, AI crawlers request the file rarely, and primary-source statements from a major search operator confirm it is not consulted for search or citation, serving instead agents that already know the site; the agent-guidance role it once aspired to is migrating to executable-tool, transaction, usage-declaration, and access-enforcement protocol layers. Its validated remaining niche — the one this framework uses — is the agent-facing documentation surface (B2A): coding agents consulting a known repository's reading order. This narrows, and does not reverse, ADR-0009's asymmetric rebalance.",
      "pairedWith": "https://github.com/shimo4228/authorship-strategy#concept/jsonld-knowledge-graph",
      "instantiatedBy": "https://github.com/shimo4228/llms-txt-writer",
      "groundedIn": [
        "https://arxiv.org/abs/2604.02544"
      ],
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0006-llm-first-ingest-dual-entry-points.md"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/jsonld-knowledge-graph",
      "@type": "Concept",
      "name": "JSON-LD Knowledge Graph",
      "alternateName": [
        {"@value": "JSON-LD knowledge graph", "@language": "en"},
        {"@value": "JSON-LD knowledge graph", "@language": "ja"}
      ],
      "description": "A linked-data file that encodes an artifact's concept-level entities and inter-entity relationships as machine-parseable triples in a structured-data vocabulary (typically schema.org plus a local namespace). Targets structured-data-ingesting LLM-mediated channels: training pipelines, knowledge-graph crawlers, programmatic readers using dataset SDKs. The concept-form half of the dual entry point; complements file-level documentation by encoding concept-level structure that prose leaves implicit.",
      "pairedWith": "https://github.com/shimo4228/authorship-strategy#concept/llms-txt-convention",
      "instantiatedBy": "https://github.com/shimo4228/jsonld-knowledge-graph",
      "groundedIn": [
        "https://arxiv.org/abs/2603.10700",
        "https://doi.org/10.2139/ssrn.6284518",
        "https://arxiv.org/abs/2604.19113",
        "https://arxiv.org/abs/2603.29979"
      ],
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0006-llm-first-ingest-dual-entry-points.md"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/human-attention-signal-rejection",
      "@type": "Concept",
      "name": "Human-Attention Signal Rejection",
      "alternateName": [
        {"@value": "Human-Attention Signal Rejection", "@language": "en"},
        {"@value": "人間アテンション signal の却下", "@language": "ja"}
      ],
      "description": "The decision to exclude platform-level human-attention metrics — Git-host star counts (gameable: purchasable) and repository page-view counts (structurally blind to LLM-mediated reach, since a human reading the work through an LLM answer generates no view) — from the framework's success definition, and to decline off-page human-distribution labor as a red-ocean activity operating on a near-empty human-arrival funnel (empirically clone:view approximately 16:1). Success is measured by the breadth of LLM-mediated channels instead. The metric-side embodiment of the scarcity-to-diffusion axis: the framework does not compete for scarce human attention on platform terms but diffuses through LLM-mediated channels. Endpoint measurement extends the structural blindness beyond page views: AI agents compress multi-page consumption into one or two HTTP requests and fire no client-side engagement events, so bounce-rate and session-depth metrics misrecord full consumption as non-engagement, and referrer stripping folds LLM-mediated arrivals into direct traffic — engagement analytics cannot capture LLM-mediated reach even in principle. Boundary: the compression evidence comes from single-fetch documentation tasks, so it is read as grounding for document-retrieval contexts rather than for every agent workload.",
      "downstreamOf": "https://github.com/shimo4228/authorship-strategy#concept/scarcity-to-diffusion-axis",
      "instantiatedBy": "https://github.com/shimo4228/authorship-strategy#adr/0007",
      "groundedIn": [
        "https://arxiv.org/abs/2604.02544"
      ],
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0007-human-attention-signals-not-a-metric.md"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/two-channel-attribution-diffusion",
      "@type": ["Concept", "Layer"],
      "name": "Two-Channel Attribution Diffusion",
      "alternateName": [
        {"@value": "Two-Channel Attribution Diffusion", "@language": "en"},
        {"@value": "2 チャネル Attribution Diffusion", "@language": "ja"},
        {"@value": "双通道署名扩散", "@language": "zh"}
      ],
      "description": "The refinement of Attribution Diffusion (Layer 2) into two mechanisms with opposite time constants and opposite levers: a parametric channel (the idea absorbed into model weights at training time — slow, driven by broad cross-platform co-occurrence of distinctive vocabulary with its source; cross-platform mention spread correlates ~0.664 with being named, far above enclosed inbound links ~0.218) and a retrieval channel (the artifact fetched at query time — fast, 3–5 day citation-pool entry and ~13-week decay, driven by freshness and structured data). Ghost citation — the source is cited but the author is not named — is the failure mode of a working retrieval channel atop absent parametric burn-in, so the two channels are optimized and measured separately and run in parallel. Two 2026 refinements are recorded. (1) The retrieval channel's verifiability splits into surface and substantive layers: deep-research audits find link validity and topical relevance high (above 94% and 80% in the strongest systems) while factual accuracy of what a citation is claimed to support runs far lower (39–77%), and deepening retrieval degrades it further — a working retrieval channel guarantees the source is reachable, not that the claim it carries is faithful, so diffusion breadth does not imply fidelity of transmission. (2) White-box probing quantifies an interaction-cost asymmetry between the channels: errors are markedly more frequent when external context must override parametric knowledge than in the reverse direction.",
      "downstreamOf": "https://github.com/shimo4228/authorship-strategy#concept/attribution-diffusion",
      "instantiatedBy": "https://github.com/shimo4228/authorship-strategy#adr/0008",
      "groundedIn": [
        "https://arxiv.org/abs/2602.06718",
        "https://arxiv.org/abs/2604.25707",
        "https://arxiv.org/abs/2603.09296",
        "https://arxiv.org/abs/2602.14869",
        "https://arxiv.org/abs/2601.21996",
        "https://arxiv.org/abs/2509.08919",
        "https://arxiv.org/abs/2602.22787",
        "https://arxiv.org/abs/2605.06635",
        "https://arxiv.org/abs/2509.04499"
      ],
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0008-rag-era-attribution-diffusion.md"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#concept/retrieval-suppressed-naming-probe",
      "@type": "Concept",
      "name": "Retrieval-Suppressed Naming Probe",
      "alternateName": [
        {"@value": "Retrieval-Suppressed Naming Probe", "@language": "en"},
        {"@value": "検索抑制 naming probe", "@language": "ja"}
      ],
      "description": "The measurement instrument for the parametric channel of Two-Channel Attribution Diffusion: a controlled prompt sent to a model with all search and grounding tools suppressed, asking what a concept is and who coined or maintains it. Success is the model producing the concept and the author's name from trained weights alone. Paired with a search-enabled citation probe that measures the retrieval channel and makes ghost citation observable within a single answer (owned identifier cited, author unnamed in prose). Detection is deterministic string matching against a versioned lexicon over retained raw responses — never model judging — with fixed single-variable templates and a negative-control probe (a plausible nonexistent concept) that quantifies the confabulation noise floor. The public probe log feeds the parametric channel it measures; the protocol records this self-contamination as a stated confound rather than hiding it. Two calibration constraints and one positioning refinement are recorded on 2026 evidence. Cross-family memorization analysis finds statistical regularities shared across model families (memorization scaling log-linearly with size) while the internal circuitry carrying memorization is family-specific — so probe results are calibrated per model family, reported with family-internal consistency rather than a single cross-model threshold, and re-baselined when a model series changes generation; for closed frontier models the internal check is impossible in principle, leaving family calibration a stated approximation. Retrieval-free citation training (active indexing: diverse restatement of each fact plus bidirectional source-fact binding at continual-pretraining time) shows parametric citation ability is a designable supply-side property — repetition volume alone does not create burn-in, structural diversity of restatement does — which positions this probe as the demand-side observer of a supply-side design space the author does not control, and grounds the framework's preference for structural diversity over sheer exposure volume in what it publishes.",
      "downstreamOf": "https://github.com/shimo4228/authorship-strategy#concept/two-channel-attribution-diffusion",
      "instantiatedBy": "https://github.com/shimo4228/authorship-strategy#adr/0011",
      "groundedIn": [
        "https://arxiv.org/abs/2602.06718",
        "https://arxiv.org/abs/2604.25707",
        "https://arxiv.org/abs/2603.09296",
        "https://arxiv.org/abs/2605.18732",
        "https://arxiv.org/abs/2511.00476",
        "https://arxiv.org/abs/2603.21658",
        "https://arxiv.org/abs/2506.17585"
      ],
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0011-two-channel-probe-protocol.md"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0001",
      "@type": "ADR",
      "name": "ADR-0001: Concept DOI as Canonical Reference",
      "description": "Every external link to a DOI-registered artifact uses the concept DOI; version-specific DOIs are used only for reproducibility citations of specific historical versions. Prevents downstream citation graphs from pinning the artifact to its initial version.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0001-concept-doi-canonical.md"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0002",
      "@type": "ADR",
      "name": "ADR-0002: DOI Federation via .zenodo.json",
      "description": "Sibling, source, and platform-mirror relationships are declared as relatedIdentifiers in archive deposit metadata so the citation network is recoverable from metadata alone, without requiring readers to follow prose disclosures.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0002-doi-federation-via-zenodo-json.md",
      "extends": "https://github.com/shimo4228/authorship-strategy#adr/0001"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0003",
      "@type": "ADR",
      "name": "ADR-0003: Cross-Platform Dataset Federation",
      "description": "The same canonical artifact is mirrored to multiple platforms (Git host, DOI archive, dataset platform) with explicit sibling cross-references on each platform so readers entering from any platform can discover the artifact's presence on the others.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0003-cross-platform-dataset-federation.md",
      "extends": "https://github.com/shimo4228/authorship-strategy#adr/0002"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0004",
      "@type": "ADR",
      "name": "ADR-0004: Authorship Metadata with ORCID Auto-Update Disabled",
      "description": "The author's persistent identifier record is enriched only with concept DOIs (never version DOIs); the archive-to-ORCID Auto-Update feature is explicitly disabled to prevent version sprawl from polluting the public record.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0004-authorship-metadata-orcid.md"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0005",
      "@type": "ADR",
      "name": "ADR-0005: README Localization Policy — Audience-Driven Maintenance",
      "description": "Locale mirrors of human-facing documentation are added or retired based on observed traffic data, not on speculation about prospective audiences. A mirror without measurable direct human audience is retired even when the language community seems important in the abstract.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0005-readme-localization-audience-driven.md"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0006",
      "@type": "ADR",
      "name": "ADR-0006: LLM-First Ingest via Dual Entry Points",
      "description": "Specifies that any framework-governed artifact deploys a prose-form navigator and a concept-form linked-data graph as a complementary pair, released synchronously. Each entry point addresses a distinct LLM-mediated reader sub-population the other cannot reach. The pair is the operational embodiment of the Axis 1 inversion (enclosure to openness) on the ingest surface.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0006-llm-first-ingest-dual-entry-points.md"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0007",
      "@type": "ADR",
      "name": "ADR-0007: Human-Attention Platform Signals Are Not a Success Metric",
      "description": "Platform human-attention signals — Git-host star counts (gameable: purchasable) and repository page-view counts (structurally blind to LLM-mediated reach) — are excluded as optimization targets and success metrics, and off-page human-distribution labor is declined as a red-ocean activity operating on a near-empty human-arrival funnel. Success is measured by the breadth of LLM-mediated channels carrying the author's signature. The metric-side counterpart to the scarcity-to-diffusion inversion; on-page human quality is maintained as hygiene, not as a growth lever.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0007-human-attention-signals-not-a-metric.md"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0008",
      "@type": "ADR",
      "name": "ADR-0008: RAG-Era Attribution Diffusion — Two Channels, Two Time Constants",
      "description": "Treats Attribution Diffusion as two channels with opposite time constants and levers: a parametric channel (idea absorbed into model weights at training time — slow, driven by cross-platform vocabulary co-occurrence) and a retrieval channel (artifact fetched at query time — fast, driven by freshness and structured data). The two are optimized and measured separately; ghost citation (source cited but author not named) is the failure mode of pursuing retrieval without parametric burn-in. Divides the single 'diffusion' of ADR-0006 and refines the measurement layer of ADR-0007.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0008-rag-era-attribution-diffusion.md",
      "extends": "https://github.com/shimo4228/authorship-strategy#adr/0006"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0009",
      "@type": "ADR",
      "name": "ADR-0009: Dual Entry Points Are Asymmetric",
      "description": "Amends ADR-0006 on 2026 evidence that the two entry points are not co-equal: the concept-form structured graph carries retrieval-time citation while the prose-form navigator's citation effect is noise (single-digit adoption, ~0.1% of agent crawls request it). The pair is retained but made asymmetric — the graph is the retrieval-channel citation lever, the navigator is rescoped to a Business-to-Agent (B2A) context surface rather than an AI-search citation lever. ADR-0006's synchronization discipline is unchanged.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0009-dual-entry-asymmetric-rebalance.md",
      "extends": "https://github.com/shimo4228/authorship-strategy#adr/0006"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0010",
      "@type": "ADR",
      "name": "ADR-0010: Vocabulary Discipline — Coin Sparingly, Anchor Densely",
      "description": "Defines the vocabulary discipline ADR-0008 named as the parametric-channel lever but left undefined. A coined term's power comes from its edge density, not from the count of coinages: a term is coined only when three conditions all hold (join-novelty, definitional anchoring, uncontested namespace), and every retained coinage is anchored densely — glossary definition in existing vocabulary, upstream citations where prior art exists, knowledge-graph edges, repeated work in the body. Everything else is said in existing vocabulary with the upstream source cited. The vocabulary-level enforcement of origin-claim scope discipline.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0010-vocabulary-discipline.md",
      "extends": "https://github.com/shimo4228/authorship-strategy#adr/0008"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0011",
      "@type": "ADR",
      "name": "ADR-0011: Two-Channel Probe Protocol — Measuring Each Channel by Its Own Instrument",
      "description": "Builds the measurement instrument ADR-0008 demanded: a scheduled two-channel probe protocol interrogating frontier models with search suppressed (does the trained model name the concept and its author?) and search enabled (are owned identifiers cited, and does the author's name survive in prose alongside the citation?). Detection is deterministic string matching against a versioned lexicon over retained raw responses, never model judging; prompts are fixed single-variable templates with a negative control; every change to prompts, models, or lexicon is a visible series break. The retrieval channel is measured on a fast calendar cadence; the parametric channel is event-driven on model-generation changes (a frozen model's weights cannot change between runs), with a monthly currency check — silent-swap detection, provider-catalog diff, staleness guard — standing in for the calendar; the channels are never blended. The public probe log feeds the parametric channel it measures — a stated confound and an on-thesis act of diffusion.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0011-two-channel-probe-protocol.md",
      "extends": "https://github.com/shimo4228/authorship-strategy#adr/0008"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0012",
      "@type": "ADR",
      "name": "ADR-0012: Link-Index Contributions to External Collections",
      "description": "Applies the thesis's enclosure axis to channel selection for external curated collections (community-curated link directories, skill marketplaces, dataset registries). Contributions are link-index entries only: the canonical artifact stays in the author's repository while the host carries a hyperlink and a short factual description; vendor-type contributions, where the host would carry the artifact body in its own distribution, are declined by default. Every prospective host passes a four-condition pre-submission audit — corporate ownership, absence of an open license, content-vendoring structure, operation as a paid-product funnel — with risk rising as conditions combine and a host meeting all four excluded even for link-only entries; a listed host that later introduces paid tiers or content vendoring triggers withdrawal. Grounded in two 2026 withdrawal episodes whose shared pattern is that vendored content is captured by any subsequent enclosure the host introduces, while a link entry bounds the worst case to a one-line description inside the enclosing host.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0012-link-index-channel-selection.md"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0013",
      "@type": "ADR",
      "name": "ADR-0013: Intrinsic Content-Derived Identifiers as a Complementary Priority-Claim Layer",
      "description": "Adds an intrinsic, content-derived identifier layer — SWHID (ISO/IEC 18670) — to the identifier federation, complementing the DOI layer rather than replacing it. A DOI is extrinsic: an opaque name bound to a metadata record by a registry, unverifiable against the artifact's content and dependent on the registry's survival; SWHID is computed from the artifact and its version-history graph, verifiable by anyone holding the content without consulting any registry, at granularities from a repository snapshot down to a single line. Division of labor: the DOI carries citability, rich metadata, and registry-mediated scholarly discovery; the intrinsic identifier carries content-verifiable, registry-independent existence proof — each covers the other's failure mode. Every release triggers an explicit archival request to a content-addressed public software archive, whose snapshot identifier is recorded alongside the DOI in citation metadata; for artifact genres where DOI registration is impractical, the intrinsic identifier is the designated substitute priority-claim mechanism, closing the manifesto's open question 4. Archival also opens a second parametric-channel ingest surface (code-focused LLM training corpora source directly from the archive) at zero marginal authoring cost. The identifier proves what content existed when — it carries no authorship semantics, which remain with the DOI / ORCID layer.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0013-intrinsic-identifier-layer.md",
      "extends": "https://github.com/shimo4228/authorship-strategy#adr/0003"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0014",
      "@type": "ADR",
      "name": "ADR-0014: Implementation Tracking as a Two-Tier Ledger with Periodic Gap-Review",
      "description": "The one ADR about operating the framework over time rather than a tactic the framework deploys. The program already publishes a dated intervention timeline in the empirical layer, but that timeline is a DOI-versioned artifact bound by empirical-layer conventions (no effect claims, the ADR-0012 host abstraction, normative/empirical separation), which bar it from doubling as a progress-management surface. Implementation tracking therefore splits into two tiers: a private implementation ledger is the operational source of truth (per-tactic deployment status, ranked candidate interventions, working detail), and the public intervention timeline is its dated, effect-claim-free projection — never merged, the update rule being ledger first then projection. A periodic gap-review generates new proposals by comparing deployed tactics against the Layer 4 tactic catalog, the framework's open questions, and recent literature, ranking candidates and running each through the judgment checklist; it is a self-application of the framework that bears on the open questions about the empirical layer's role and the framework's recursive application to itself. The generic review procedure lives in the framework's operational skill; only the project-specific wiring — which artifacts are this program's ledger and timeline — lives in the project's context file.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0014-implementation-tracking-two-tier-ledger.md"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0015",
      "@type": "ADR",
      "name": "ADR-0015: License Selection by Audience, Not Artifact Form",
      "description": "Fixes license selection on the artifact's dominant audience rather than its surface form, on the standing principle that attribution is carried by the federated-identifier layer (the identifier-federation triplet 0001-0003 and the intrinsic-identifier layer 0013) rather than by the license, so the license is chosen to minimize reuse friction. Machine-mined artifacts — datasets, corpora, traffic and probe logs, runtime data, knowledge graphs, and prose published for LLM-mediated reach, which under this LLM-first program (ADR-0007) is mined not read regardless of being prose — take a public-domain dedication (CC0-1.0); executable code takes a permissive software license (MIT/Apache-2.0), carried whole-repo on a code-bearing repository for legibility since it is already absorbed freely and unfiltered; genuinely human-first artifacts may take an attribution-requiring content license (CC-BY-4.0); a split is reserved for repositories whose non-code material is the entire deliverable; non-commercial and no-derivatives terms are prohibited as enclosure that forbids the parametric channel's training-time absorption and the preference hierarchy's creative reuse. The license-layer counterpart of vocabulary discipline (0010); disjoint from ADR-0012, which governs the license a prospective external host must extend rather than the license the author applies. Triggered by a 2026-06-17 cross-repository audit that found four license patterns in simultaneous use and one non-commercial restriction contradicting the framework, and sharpened from a form axis to an audience axis when an essay collection's LLM-first designation made form classification untenable.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0015-license-selection-by-audience.md"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0016",
      "@type": "ADR",
      "name": "ADR-0016: Genre-Split Placement — Essays as Repository-Corpus Canonical with Intrinsic Identifier, Papers as Concept-DOI Canonical",
      "description": "Records which genre takes which canonical, a routing the identifier ADRs left open: ADR-0001 fixed the concept DOI as canonical and ADR-0013 added an intrinsic content-derived identifier (SWHID) as the substitute claim for DOI-impractical genres, but neither said which genre is which. ADR-0016 routes by genre. The essay genre's canonical is the author's version-controlled repository corpus, its priority claim resting on the intrinsic content-derived identifier (a snapshot of the corpus in a content-addressed public archive) rather than a registry DOI, under a public-domain dedication (CC0-1.0, ADR-0015); the paper genre's canonical is the concept DOI (ADR-0001), with the intrinsic identifier as its complementary layer. Syndicated essay copies are bound to the canonical by entity federation — sameAs, ORCID, DOI, intrinsic identifier, and the distinctive vocabulary that survives paraphrase (ADR-0009, ADR-0010) — not by a platform canonical-URL tag, whose effect on LLM-mediated credit is unverified and which is retained only as human-reader and search-engine hygiene. Corpus membership is gated by an authenticity criterion (Layer 1: author-voiced, reader-intended pieces; study or learning drafts without an author voice are excluded), and a load-bearing essay idea is promoted to a concept-DOI deposit when it graduates into a paper (Layer 3, idea-versus-scaffold separation). It instantiates ADR-0013's DOI-impractical-genre clause concretely and complements ADR-0015's license-by-audience rule, closing the placement gap between them. Triggered by a 2026-06-25 review of where crystallized essays and papers should live for LLM-mediated diffusion, which found the essay corpus was a publishing pipeline lacking an intrinsic-identifier priority claim, a reconciled license, and entity federation; the corpus was governed accordingly as the implementation this decision records.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0016-genre-split-placement.md",
      "extends": [
        "https://github.com/shimo4228/authorship-strategy#adr/0013",
        "https://github.com/shimo4228/authorship-strategy#adr/0015"
      ]
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0017",
      "@type": "ADR",
      "name": "ADR-0017: Failure-Mode Diagnostics — A Detector and Recovery Strategy for Each of the Three Acknowledged Failure Modes",
      "description": "Operationalizes the manifesto's eighth open question — the framework's acknowledged failure modes — by pairing each of the three named modes with a diagnostic signal and a recovery strategy. For reach without recognition (the ghost citation of authorship, where ideas diffuse but the diffusion does not carry the author's name), the detector is a naming probe that succeeds at the concept level yet fails at the author level, read against a mechanism the framework borrows from the citation and interpretability literature: the parametric channel gates the retrieval channel so a source's address can be cited while its author goes unnamed, factual recall is foreseeable at design time so a coinage's low representation places it at the harsh end of the recall floor by construction, and recall reads back prior authority unequally rather than presence neutrally — the supply-side connection to the ninth open question; the recovery anchors distinctive vocabulary densely, keeps the origin claim narrow, and accepts that under full three-axis inversion reach without recognition may be a structural price rather than a defect to chase. For over-publication, the detector is the author's own identifier portfolio carrying multiple superseded versions of one idea, and the recovery is the concept-DOI canonical with version discipline. For under-investment in worked implementation, the detector is a doctrine-heavy, implementation-light portfolio, and the recovery rebalances toward the abstract-doctrine-plus-worked-implementation pair the idea-versus-scaffold layer requires. A load-bearing caveat governs all three: a diagnostic is a failure-detector, not a success metric — it never becomes an optimization target, because promoting a detector to a target re-imports the purchasable, off-page metric posture the framework rejects and re-creates the reach-chasing it prohibits, while honestly articulating one's own failure modes is consistent with the authenticity commitment at the top of the stack.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0017-failure-mode-diagnostics.md",
      "extends": [
        "https://github.com/shimo4228/authorship-strategy#adr/0007",
        "https://github.com/shimo4228/authorship-strategy#adr/0011",
        "https://github.com/shimo4228/authorship-strategy#adr/0001",
        "https://github.com/shimo4228/authorship-strategy#adr/0004"
      ],
      "groundedIn": [
        "https://arxiv.org/abs/2509.13365",
        "https://arxiv.org/abs/2605.18732",
        "https://arxiv.org/abs/2511.00476",
        "https://arxiv.org/abs/2602.22787"
      ]
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0018",
      "@type": "ADR",
      "name": "ADR-0018: Origin-Claim Falsifiability — Test a Priority Claim Against Prior Art Before Publishing It in a Durable Artifact",
      "description": "The decision to codify the framework's informally-practiced origin-claim scope discipline into a stated procedure: before an origin claim enters a durable, citable artifact, the author runs a retrieval search for prior work that would refute the claim, and any claim that survives only because it was never tested — one that is unfalsifiable in principle, or one the search shows already anticipated by located prior work — is rescoped to its narrowest defensible form rather than dropped. The governing criterion is falsifiability: a publishable origin claim is one a prior-art search could in principle refute and did not. The check is a binary defensibility judgment feeding a human rescope decision, deliberately not a novelty score, so it composes with the metric-rejection commitment rather than turning authenticity into a number to maximize. The procedure only ever narrows a claim, making it a humility instrument that extends the vocabulary discipline's origin-claim scope clause and rests on Layer 1 authenticity; it grounds the verification step on the demonstration, by an external evidence-grounded agentic novelty-assessment system, that an externally-computed novelty check built on retrieved real prior work is feasible — cited for feasibility, never for an accuracy figure, with the framework claiming no priority over that system. Writing the habit down as an adoptable check answers the framework's begging-the-question risk by requiring every origin claim, including the framework's own, to survive a prior-art refutation search before it becomes costly to retract.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0018-claim-falsifiability-criterion.md",
      "extends": [
        "https://github.com/shimo4228/authorship-strategy#adr/0010"
      ],
      "groundedIn": [
        "https://arxiv.org/abs/2601.01576"
      ]
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0019",
      "@type": "ADR",
      "name": "ADR-0019: Structural Optimization versus Content Authenticity — The Structured-Artifact Tactic Optimizes the Transmission Path, Never the Content",
      "description": "Locates the boundary between legitimate structural optimization and prohibited content deformation for the framework's structured-artifact diffusion tactic. The structured-data efficacy literature establishes that AI-retrieval citation lift attaches to a document's structure and to attribute-rich, entity-anchored markup rather than to surface text or the mere presence of markup, which justifies the structured-artifact tactic while creating a standing pressure to optimize for citation under a borrowed marketing-optimization frame. The decision draws the line at the object of optimization, not its intensity: optimizing the transmission path — document architecture, information hierarchy, entity anchoring, and the dense anchoring of distinctive vocabulary — is legitimate because it changes how the idea travels and leaves the idea's content as authored, whereas deforming the content to win citations through padded attribute-richness, keyword-stuffing, or claims shaped to a channel's reward function is prohibited because it deforms what the idea is. The rule — optimize how the idea travels, never what the idea is — makes content deformation an authenticity violation, made goalless by the prior rejection of citation and visibility as success metrics. It extends the structured-graph citation lever and the anchor-densely vocabulary discipline as two surfaces of the same legitimate carrier investment, sits beneath the authenticity layer the boundary protects, and is read against the supply-side entity-grounding tension and the tactic-obsolescence question.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0019-structural-optimization-vs-content-authenticity.md",
      "extends": [
        "https://github.com/shimo4228/authorship-strategy#adr/0009",
        "https://github.com/shimo4228/authorship-strategy#adr/0010",
        "https://github.com/shimo4228/authorship-strategy#adr/0007"
      ],
      "groundedIn": [
        "https://arxiv.org/abs/2604.19113",
        "https://arxiv.org/abs/2603.29979",
        "https://arxiv.org/abs/2603.10700",
        "https://doi.org/10.2139/ssrn.6284518"
      ]
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0020",
      "@type": "ADR",
      "name": "ADR-0020: Onboarding to Third-Party AI-Derived Repository Surfaces — Synthetic Wikis and Documentation Hubs",
      "description": "Onboards idea-bearing public repositories to two third-party derivation-type surfaces that build an LLM-consumable view from the repository: a synthetic wiki that paraphrases the codebase behind a conversational query interface, and a documentation hub that serves the repository's own machine-readable documents verbatim through a model-callable interface, each attaching an adoptable badge. The decision onboards to both and blesses the derived views rather than gating them, under a per-type discipline: the synthetic wiki's paraphrase is used as a regurgitation-test drift diagnostic and defended by upstream dense anchoring rather than corrected on the derived surface, while the documentation hub's access-count badge is read as a measurement signal of the LLM-mediated channel and never as a success metric. It declines an index-only catalog whose artifact model is the installable code library as an artifact-type mismatch, and declines self-hosted query infrastructure as friction the framework does not take on, keeping the origin claim fixed on the identifier-federation layer rather than on any derived surface. It is the derivation-axis counterpart to the enclosure-axis channel-selection decision, extends the LLM-first ingest surface, is bounded by the metric-rejection decision, and connects its regurgitation-test diagnostic to the vocabulary discipline and the measurement protocol.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0020-derivation-surface-onboarding.md",
      "extends": [
        "https://github.com/shimo4228/authorship-strategy#adr/0006",
        "https://github.com/shimo4228/authorship-strategy#adr/0012",
        "https://github.com/shimo4228/authorship-strategy#adr/0007",
        "https://github.com/shimo4228/authorship-strategy#adr/0010"
      ]
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy#adr/0021",
      "@type": "ADR",
      "name": "ADR-0021: Self-Sovereign Entity Grounding — Community-Governed Authority Records Are a Revocable Layer, Not a Foundation",
      "description": "Retires self-created entries in community-governed authority records as a Layer 4 tactic, after the host of such a record judged the author's account promotional by its aggregate editing pattern — no individual edit cited — revoked it indefinitely, and mass-deleted every entry: the author entity, the artifact entities, bibliographic records of cited works, and the citation edges among them. The decision classifies entity-grounding surfaces by revocation control: self-sovereign layers (the repository and its knowledge graph, registry deposits under the author's account, the author-identifier record, and the intrinsic content-derived identifier layer, which alone is non-revocable by construction) may be load-bearing for the origin claim, while third-party-governed surfaces are reach amplifiers whose total loss the strategy must survive. Third-party-governed grounding is admitted only when earned — created unprompted by uninvolved parties — never self-manufactured or solicited. On revocation, dead identifiers are purged promptly from every machine-readable carrier while dated historical records are preserved unmodified; circumvention through new accounts or proxies is prohibited outright; and every future third-party deployment faces an aggregate-pattern test — how the account's cumulative footprint reads to the host's governance rather than whether each action is formally compliant. Records the sharpest observed instance of the manifesto's ninth open question without closing it.",
      "recordedIn": "https://github.com/shimo4228/authorship-strategy/blob/main/docs/adr/0021-self-sovereign-entity-grounding.md",
      "extends": [
        "https://github.com/shimo4228/authorship-strategy#adr/0013",
        "https://github.com/shimo4228/authorship-strategy#adr/0002",
        "https://github.com/shimo4228/authorship-strategy#adr/0009",
        "https://github.com/shimo4228/authorship-strategy#adr/0012",
        "https://github.com/shimo4228/authorship-strategy#adr/0017"
      ]
    },

    {
      "@id": "https://doi.org/10.5281/zenodo.19200726",
      "@type": ["ResearchLine", "EcosystemRepo"],
      "name": "Agent Knowledge Cycle",
      "alternateName": "AKC",
      "description": "Six-phase bidirectional growth loop for sustaining intent alignment between an AI agent and its operator over time. Mechanism sibling: defines how knowledge cycles inside the operator-agent pair; this research line addresses how the cycle's outputs diffuse outside it.",
      "url": "https://github.com/shimo4228/agent-knowledge-cycle",
      "identifier": "10.5281/zenodo.19200726",
      "siblingOf": "https://github.com/shimo4228/authorship-strategy"
    },

    {
      "@id": "https://doi.org/10.5281/zenodo.19212118",
      "alternateName": "CA",
      "@type": ["ResearchLine", "EcosystemRepo"],
      "name": "Contemplative Agent",
      "description": "Autonomous agents running on a local 9B model, grounded in four contemplative axioms. Implementation sibling: this repository participates in the empirical layer's traffic dataset, and its non-dualistic axiomatic foundation supplies the underlying rationale for the framework's scaffold-as-collaborator commitment.",
      "url": "https://github.com/shimo4228/contemplative-agent",
      "identifier": "10.5281/zenodo.19212118",
      "siblingOf": "https://github.com/shimo4228/authorship-strategy"
    },

    {
      "@id": "https://doi.org/10.5281/zenodo.19652013",
      "alternateName": "AAP",
      "@type": ["ResearchLine", "EcosystemRepo"],
      "name": "Agent Attribution Practice",
      "description": "Harness-neutral ADRs on accountability distribution in autonomous AI agents. Vocabulary sibling: shares the word 'attribution' but with disjoint meaning (accountability for action vs. credit for source). The two meanings are intentionally kept separate; do not conflate.",
      "url": "https://github.com/shimo4228/agent-attribution-practice",
      "identifier": "10.5281/zenodo.19652013",
      "siblingOf": "https://github.com/shimo4228/authorship-strategy",
      "vocabularyDisjoint": "https://github.com/shimo4228/authorship-strategy"
    },

    {
      "@id": "https://doi.org/10.5281/zenodo.20262112",
      "alternateName": "ANS",
      "@type": ["ResearchLine", "EcosystemRepo"],
      "name": "Attention, Not Self",
      "description": "A cross-disciplinary inquiry contrasting three Buddhist Abhidharma traditions (Theravāda, Sarvāstivāda, Yogācāra) with computational phenomenology (predictive processing, active inference, global workspace theory, parallel distributed processing). Cross-cutting sibling: unlike the agent-design lines (AKC, Contemplative Agent, AAP) it specifies no agent mechanism or practice; like this research line it occupies the agent-design lines' diffusion and framing layer. Shares Buddhist terminology with Contemplative Agent but with asymmetric usage: Contemplative Agent uses it as a behavioral preset, this line as a comparative cognitive framework. Began traffic observation after the v0.1.0 empirical baseline window.",
      "url": "https://github.com/shimo4228/attention-not-self",
      "identifier": "10.5281/zenodo.20262112",
      "siblingOf": "https://github.com/shimo4228/authorship-strategy"
    },

    {
      "@id": "https://github.com/shimo4228/shimo4228",
      "@type": "EcosystemRepo",
      "name": "Research Program Hub",
      "description": "Metadata-only federation hub at the center of the shimo4228 research ecosystem. Aggregates cross-references to sibling research lines without containing line-specific content itself. Not a research line; treating it as one collapses the distinction between content and metadata.",
      "url": "https://github.com/shimo4228/shimo4228",
      "hasPart": "https://github.com/shimo4228/authorship-strategy"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy-skill",
      "@type": "EcosystemRepo",
      "name": "authorship-strategy-skill",
      "description": "Component skill of this research line. Operational form of the four-layer judgment checklist, packaged as a standalone Claude Code skill repository. Loadable into LLM-based coding agents as a rule set. Operationalizes the thesis and the twenty-one ADRs.",
      "url": "https://github.com/shimo4228/authorship-strategy-skill",
      "derivesFrom": "https://github.com/shimo4228/authorship-strategy"
    },

    {
      "@id": "https://github.com/shimo4228/release-doi",
      "@type": "EcosystemRepo",
      "name": "release-doi",
      "description": "Component skill of this research line. Release-time workflow operationalizing the identifier-federation triplet (ADRs 0001-0003) as a five-phase verify-and-deposit runbook for DOI-registered research repositories.",
      "url": "https://github.com/shimo4228/release-doi",
      "derivesFrom": "https://github.com/shimo4228/authorship-strategy"
    },

    {
      "@id": "https://github.com/shimo4228/llms-txt-writer",
      "@type": "EcosystemRepo",
      "name": "llms-txt-writer",
      "description": "Component skill of this research line. Operationalizes Layer 4 tactic 7 — Answer.AI llms.txt convention. Writes the AI-facing reference files (llms.txt, llms-full.txt, FAQ, glossary) that every framework-applied repository requires.",
      "url": "https://github.com/shimo4228/llms-txt-writer",
      "derivesFrom": "https://github.com/shimo4228/authorship-strategy"
    },

    {
      "@id": "https://github.com/shimo4228/jsonld-knowledge-graph",
      "@type": "EcosystemRepo",
      "name": "jsonld-knowledge-graph",
      "description": "Component skill of this research line. Operationalizes Layer 4 tactic 7 — JSON-LD knowledge graph. Designs and ships graph.jsonld next to llms.txt for projects with stable concept-level structure.",
      "url": "https://github.com/shimo4228/jsonld-knowledge-graph",
      "derivesFrom": "https://github.com/shimo4228/authorship-strategy"
    },

    {
      "@id": "https://github.com/shimo4228/authorship-strategy-rules",
      "@type": "EcosystemRepo",
      "name": "authorship-strategy-rules",
      "description": "Component of this research line: the four-layer judgment framework packaged as a single always-loaded behavioral rule — the deterministic, always-on counterpart to authorship-strategy-skill. Applies the framework every session within its trigger scope (repositories the operator owns that are DOI-registered and idea-rescue in character). A distribution mirror with no own DOI; references the parent concept DOI.",
      "url": "https://github.com/shimo4228/authorship-strategy-rules",
      "derivesFrom": "https://github.com/shimo4228/authorship-strategy"
    },

    {
      "@id": "https://arxiv.org/abs/2602.06718",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "GhostCite: A Large-Scale Analysis of Citation Validity in the Age of Large Language Models",
      "author": "Zuyao Xu et al.",
      "datePublished": "2026-02-06",
      "identifier": "arXiv:2602.06718",
      "url": "https://arxiv.org/abs/2602.06718",
      "description": "Large-scale audit of citation validity across thirteen LLMs and 56,381 published papers (2.2 million citations checked; 1.07% of papers carry invalid citations). Uses the term 'ghost citation' for a disjoint phenomenon — fabricated or invalid citations to non-existent sources — not the attribution-loss sense the framework names the ghost citation of authorship; the two senses should not be conflated. Cited as external context for citation failure under LLM mediation (ADR-0008), not as the source of the framework's term."
    },

    {
      "@id": "https://arxiv.org/abs/2604.25707",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms",
      "author": ["Zhang Kai", "He Xinyue", "Yao Jingang"],
      "datePublished": "2026-04-28",
      "identifier": "arXiv:2604.25707",
      "url": "https://arxiv.org/abs/2604.25707",
      "description": "Two-stage GEO measurement framework separating citation selection (which sources a platform fetches) from citation absorption (how much a fetched page contributes to the answer), across ChatGPT, Google AI Overview/Gemini, and Perplexity (602 prompts, 21,143 valid citations). External grounding for the retrieval channel's measurement layer (ADR-0008) and for the metric-rejection logic of ADR-0007: a raw citation count is not a measure of citation influence."
    },

    {
      "@id": "https://arxiv.org/abs/2603.09296",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Diagnosing and Repairing Citation Failures in Generative Engine Optimization",
      "author": ["Zhihua Tian", "Yuhan Chen", "Yao Tang", "Jian Liu", "Ruoxi Jia"],
      "datePublished": "2026-03-10",
      "identifier": "arXiv:2603.09296",
      "url": "https://arxiv.org/abs/2603.09296",
      "description": "A taxonomy of GEO citation-failure modes plus AgentGEO, an agentic system that raises citation rates by over 40% while modifying only 5% of content, against a 25% modification baseline. External grounding for the retrieval channel as an optimizable surface (ADR-0008), with the caution — consistent with this framework's distrust of generic optimization — that broad optimization degrades long-tail content."
    },

    {
      "@id": "https://arxiv.org/abs/2602.14869",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Concept Influence: Leveraging Interpretability to Improve Performance and Efficiency in Training Data Attribution",
      "author": ["Matthew Kowal", "Goncalo Paulo", "Louis Jaburi", "Tom Tseng", "Lev E McKinney", "Stefan Heimersheim", "Aaron David Tucker", "Adam Gleave", "Kellin Pelrine"],
      "datePublished": "2026-02-16",
      "identifier": "arXiv:2602.14869",
      "url": "https://arxiv.org/abs/2602.14869",
      "description": "A training-data-attribution method that replaces classical influence functions' per-test-example gradient with a semantic direction — a linear probe or sparse-autoencoder feature — asking which training data shaped a concept rather than an exact string (training on the top 10–20% highest-influence data alone raises an unsafe-code score roughly tenfold, at ~20× the speed of classical influence functions). External white-box grounding for the Layer-3 wager (idea-versus-scaffold separation): an author's idea is retained in a model as a parametric concept direction, not verbatim text — which is what survives when the implementation dissolves. Boundary: it requires white-box access to a model one trains oneself, so it does not measure burn-in inside a closed commercial LLM and does not close the parametric-channel measurement gap ADR-0008 leaves open."
    },

    {
      "@id": "https://arxiv.org/abs/2601.21996",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Mechanistic Data Attribution: Tracing the Training Origins of Interpretable LLM Units",
      "author": ["Jianhui Chen", "Yuzhang Luo", "Liangming Pan"],
      "datePublished": "2026-01-29",
      "identifier": "arXiv:2601.21996",
      "url": "https://arxiv.org/abs/2601.21996",
      "description": "A framework that traces interpretable units inside a model — notably induction heads, the circuits underlying in-context learning — back to the training samples that formed them, using influence functions. On the Pythia family it causally validates the link: removing high-influence samples suppresses induction-head emergence while random interventions do not, and high-influence samples are dominated by repetitive structural data (LaTeX, XML, code; power-law influence). External circuit-level grounding for the parametric channel of Two-Channel Attribution Diffusion (ADR-0008). Boundary: white-box and retraining-scale, so it is upstream mechanistic evidence, not an attribution probe applicable to a closed commercial model."
    },

    {
      "@id": "https://arxiv.org/abs/2402.12261",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "NEO-BENCH: Evaluating Robustness of Large Language Models with Neologisms",
      "author": ["Jonathan Zheng", "Alan Ritter", "Wei Xu"],
      "datePublished": "2024-02-19",
      "identifier": "arXiv:2402.12261",
      "url": "https://arxiv.org/abs/2402.12261",
      "description": "Benchmark study (ACL 2024) showing that machine-translation performance is nearly halved when a single neologism is introduced into a sentence, and that models with later knowledge cutoffs show lower perplexity on neologisms and better downstream performance — absorption tracks training-data presence. External grounding for the premise of Vocabulary Discipline (ADR-0010): an unabsorbed coinage does not merely fail to carry signature, it degrades processing of the text around it. Boundary: measures training-data presence, not document-level anchor density — it supports the coin-sparingly cost model while the anchor-densely remedy (dense anchoring raises absorption probability via co-occurrence) remains this framework's own hypothesis."
    },

    {
      "@id": "https://arxiv.org/abs/2510.08506",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Neologism Learning for Controllability and Self-Verbalization",
      "author": ["John Hewitt", "Oyvind Tafjord", "Robert Geirhos", "Been Kim"],
      "datePublished": "2025-10-09",
      "identifier": "arXiv:2510.08506",
      "url": "https://arxiv.org/abs/2510.08506",
      "description": "Trains a single new token embedding (all other parameters frozen) to make a model express a concept, and reports machine-only synonyms — words that look unrelated to humans but trigger the same model-internal concept (the ordinary word 'lack' acts as a synonym for a trained brevity neologism) — plus self-verbalization, where the model explains a learned neologism in natural language (questionnaire-based verbalizations close on average 83% of the behavioral gap). Surfaces the human-model anchor-correspondence question for Vocabulary Discipline (ADR-0010): anchors chosen for human-readable semantic proximity need not remain anchors at the parametric-representation level. Boundary: a post-hoc steering-token regime, not pretraining absorption; it qualifies only the parametric rationale of the anchoring obligation, leaving the retrieval-channel and graph-edge rationales untouched. Recorded as an open-question candidate in the empirical layer's 2026-06 external-literature note."
    },

    {
      "@id": "https://arxiv.org/abs/2605.06426",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "From 124 Million Tokens to 1,021 Neologisms: A Large-Scale Pipeline for Automatic Neologism Detection",
      "datePublished": "2026-05-07",
      "identifier": "arXiv:2605.06426",
      "url": "https://arxiv.org/abs/2605.06426",
      "description": "A large-scale automatic neologism-detection pipeline reporting substantial cross-model disagreement about which candidate words are neologisms (599 of 1,021 candidates, 58.7%, confirmed by manual validation). Cited in the empirical layer's 2026-06 external-literature note as a methodological warning for Vocabulary Discipline (ADR-0010): any future regurgitation-test protocol asking whether a model recognizes a coined term will get model-dependent answers, so single-model probes cannot be cured by simply picking a better model."
    },

    {
      "@id": "https://arxiv.org/abs/2509.13365",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "The Provenance Problem: LLMs and the Breakdown of Citation Norms",
      "author": ["Brian D. Earp", "Haotian Yuan", "Julian Koplin", "Sebastian Porsdam Mann"],
      "datePublished": "2025-09",
      "identifier": "arXiv:2509.13365",
      "url": "https://arxiv.org/abs/2509.13365",
      "description": "Names the 'provenance problem': a systematic breakdown in the chain of scholarly credit when LLM-mediated drafting reproduces ideas without attribution — distinct from plagiarism because the author may act in good faith and disclose AI use yet still benefit from uncredited intellectual contributions. External scholarly grounding for the reach-without-recognition failure mode (manifesto Open Question 8), the adversarial case in which diffusion does not carry the originating author's name; recorded as an open tension, not resolved."
    },

    {
      "@id": "https://arxiv.org/abs/2509.08919",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Generative Engine Optimization: How to Dominate AI Search",
      "author": ["Mahe Chen", "Xiaoxuan Wang", "Kaiwen Chen", "Nick Koudas"],
      "datePublished": "2025-09",
      "identifier": "arXiv:2509.08919",
      "url": "https://arxiv.org/abs/2509.08919",
      "description": "Reports that generative AI search strongly favors earned media (distributed third-party sources) over brand-owned content, unlike balanced traditional search, and identifies engine-specific differences in domain diversity, freshness, and language stability. Preliminary external observation consistent with the Scarcity-to-Diffusion axis and with the cross-platform driver of the parametric channel in Two-Channel Attribution Diffusion (ADR-0008): distributed presence, not enclosed owned content, is what AI answer engines surface."
    },

    {
      "@id": "https://arxiv.org/abs/2603.10700",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Structured Linked Data as a Memory Layer for Agent-Orchestrated Retrieval",
      "author": ["Andrea Volpini", "Elie Raad", "Beatrice Gamba", "David Riccitelli"],
      "datePublished": "2026-03-11",
      "identifier": "arXiv:2603.10700",
      "url": "https://arxiv.org/abs/2603.10700",
      "description": "Tests whether structured linked data — schema.org markup served as agent-readable entity pages — improves retrieval accuracy, reporting +29.6% for a standard retrieval-augmented pipeline and +29.8% for a full agentic pipeline when the entity pages are added; the paper notes that JSON-LD markup alone provides only modest improvements, the gains coming from the enhanced entity-page format (markup plus agent instructions and structured retrieval), not markup in isolation. External grounding for the JSON-LD knowledge graph as the concept-form half of the dual entry point (ADR-0006) and for the structured-graph-carries-retrieval claim of ADR-0009: structured linked data functions as a retrieval memory layer, not merely a discovery aid. Read together with the qualifying counter-evidence of SSRN 6284518, the two point the same way — richness and enhancement, not bare markup presence, carry the lift."
    },

    {
      "@id": "https://doi.org/10.2139/ssrn.6284518",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Does Schema Markup Predict AI Citation? A Cross-Platform Empirical Study of Structured Data and Generative Engine Optimization",
      "author": "Kurt Fischman",
      "datePublished": "2026-02-22",
      "identifier": "10.2139/ssrn.6284518",
      "url": "https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6284518",
      "description": "A cross-platform study — 730 AI citations from ChatGPT and Gemini over 75 commercial queries, 1,006 total pages analyzed against organic ranking — finding that schema presence does not independently predict AI citation once organic ranking is controlled: an initial negative pooled association (OR 0.546, p < .001) was identified as a ranking-confound artifact, and the citation difference instead attaches to attribute-rich markup with populated fields (Product, Review) over generic types (Article, Organization, BreadcrumbList) at 61.7% versus 41.6% (p = .012). Qualifying counter-evidence for the structured-graph citation claim of ADR-0009 and the JSON-LD knowledge graph tactic: structured-data presence is not the lever; attribute-rich, entity-anchored markup is. Externally echoes the anchor-densely discipline of ADR-0010 and grounds the entity-grounding access question of manifesto Open Question 9. A non-arXiv primary source (SSRN working paper) admitted to the citation graph on the strength of its controlled design and resolvable DOI."
    },
    {
      "@id": "https://arxiv.org/abs/2602.22787",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Probing for Knowledge Attribution in Large Language Models",
      "author": ["Ivo Brink", "Alexander Boer", "Dennis Ulmer"],
      "datePublished": "2026-02-26",
      "identifier": "arXiv:2602.22787",
      "url": "https://arxiv.org/abs/2602.22787",
      "description": "A white-box study showing that a simple linear probe on a model's hidden representations reliably classifies the dominant knowledge source behind each output — separating context-driven answers from internal-parametric ones (the AttriWiki self-supervised pipeline supplies the labels). Independent white-box corroboration that the parametric and retrieval channels of Two-Channel Attribution Diffusion (ADR-0008) and the behavioral retrieval-suppressed naming probe (ADR-0011) are separately measurable entities rather than two names for one mechanism."
    },
    {
      "@id": "https://arxiv.org/abs/2602.18733",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Prior Aware Memorization: An Efficient Metric for Distinguishing Memorization from Generalization in Large Language Models",
      "author": ["Trishita Tiwari", "Ari Trachtenberg", "G. Edward Suh"],
      "datePublished": "2026-02-21",
      "identifier": "arXiv:2602.18733",
      "url": "https://arxiv.org/abs/2602.18733",
      "description": "Argues that existing memorization measures conflate genuine verbatim memorization with generalization over statistically common patterns, and supplies an efficient metric (Prior Aware Memorization) that separates the two without the model-retraining counterfactual-memorization methods require. Mechanistic grounding for the idea-versus-scaffold prediction (Layer 3 of the four-layer framework): what survives in the weights is the generalized pattern rather than the verbatim scaffold."
    },
    {
      "@id": "https://arxiv.org/abs/2605.18732",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Predictable Confabulations: Factual Recall by LLMs Scales with Model Size and Topic Frequency",
      "author": ["Matthew L. Smith", "Jonathan P. Shock", "Samuel T. Segun", "Iyiola E. Olatunji", "Tegawendé F. Bissyandé"],
      "datePublished": "2026-05-18",
      "identifier": "arXiv:2605.18732",
      "url": "https://arxiv.org/abs/2605.18732",
      "description": "Evaluates 38 models on over 8,900 scholarly references and finds factual-recall quality follows a sigmoid in the log-linear combination of model size and a topic's training-data representation — those two variables alone explaining 60% of the variance across sixteen dense models (74–94% within a single family), under a superposition-inspired signal-to-noise account. Makes the parametric channel of the two-channel probe (ADR-0011) predictable at design time and sharpens the neologism caution of the vocabulary discipline (ADR-0010): a coined term has the lowest topic frequency and so sits at the harshest end of the noise floor."
    },
    {
      "@id": "https://arxiv.org/abs/2604.19113",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Think Before Writing: Feature-Level Multi-Objective Optimization for Generative Citation Visibility",
      "author": ["Zikang Liu", "Peilan Xu"],
      "datePublished": "2026-04-21",
      "identifier": "arXiv:2604.19113",
      "url": "https://arxiv.org/abs/2604.19113",
      "description": "Observes that generative answer engines expose content through selective citation rather than ranked retrieval, and that prior generative-engine-optimization work relied on token-level rewriting with weak control over the visibility-quality trade-off; proposes FeatGEO, a feature-level multi-objective framework optimizing over interpretable structural, content, and linguistic properties instead of editing text directly. Causal grounding for the structure side of the JSON-LD knowledge graph tactic (ADR-0009): document structure, not surface text, carries citation visibility."
    },
    {
      "@id": "https://arxiv.org/abs/2603.29979",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Structural Feature Engineering for Generative Engine Optimization: How Content Structure Shapes Citation Behavior",
      "author": ["Junwei Yu", "Mufeng Yang", "Yepeng Ding", "Hiroyuki Sato"],
      "datePublished": "2026-03-31",
      "identifier": "arXiv:2603.29979",
      "url": "https://arxiv.org/abs/2603.29979",
      "description": "Decomposes document structure into macro- (document architecture), meso- (information hierarchy), and micro- (formatting) levels and measures how each shapes selective citation by generative engines (GEO-SFE). Companion structural-GEO grounding for the JSON-LD knowledge graph tactic (ADR-0009): the citation lift attaches to structural investment, the structure-side counterpart of the presence-is-not-the-lever caveat the Fischman result makes."
    },
    {
      "@id": "https://arxiv.org/abs/2603.23219",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Decoding AI Authorship: Can LLMs Truly Mimic Human Style Across Literature and Politics?",
      "author": ["Nasser A Alsadhan"],
      "datePublished": "2026-03-24",
      "identifier": "arXiv:2603.23219",
      "url": "https://arxiv.org/abs/2603.23219",
      "description": "Tests whether frontier models (GPT-4o, Gemini 1.5 Pro, Claude Sonnet 3.5) can mimic named literary and political authors, and finds via transformer classification, interpretable features, and perplexity that human-authored and model-generated text occupy distinguishable stylometric regions even under deliberate imitation. Cited as a tension for Authenticity (Layer 1): an authorial signature may persist through AI-mediated rephrasing, but the framework records this as a currently-discriminable difference rather than a permanent fingerprint and builds no stylometric self-measure (metric-rejection, ADR-0007)."
    },

    {
      "@id": "https://arxiv.org/abs/2601.01576",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "OpenNovelty: An LLM-powered Agentic System for Verifiable Scholarly Novelty Assessment",
      "author": ["Ming Zhang", "Kexin Tan", "Yueyuan Huang", "Yujiong Shen", "Chunchun Ma", "Li Ju", "Xinran Zhang", "Yuhui Wang", "Wenqing Jing", "Jingyi Deng", "Huayu Sha", "Binze Hu", "Jingqi Tong", "Changhao Jiang", "Yage Geng", "Yuankai Ying", "Yue Zhang", "Zhangyue Yin", "Zhiheng Xi", "Shihan Dou", "Tao Gui", "Qi Zhang", "Xuanjing Huang"],
      "datePublished": "2026-01-04",
      "identifier": "arXiv:2601.01576",
      "url": "https://arxiv.org/abs/2601.01576",
      "description": "Presents an LLM-powered agentic system that assesses the novelty of a scholarly submission by grounding every judgment in retrieved real papers rather than unverified LLM generation: it extracts contribution claims and retrieval queries, retrieves prior work via semantic search, builds a hierarchical taxonomy and performs contribution-level full-text comparisons, then synthesizes a structured novelty report with explicit citations and evidence snippets; the authors report deploying it on 500+ ICLR 2026 submissions with reports made public, and offer the preliminary observation that it can surface relevant prior work authors may have overlooked. Cited as grounding for origin-claim scope discipline: it supplies an externally-computed, evidence-grounded novelty check that locates a contribution against retrieved prior art, the mechanism a scope-disciplined origin claim relies on. The paper itself concerns peer-review novelty assessment and does not articulate the normative discipline, and its deployment figure is scale and its surfacing-of-prior-work statement is explicitly preliminary, so it is not cited for any quantitative accuracy or effectiveness result."
    },

    {
      "@id": "https://arxiv.org/abs/2511.00476",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Remembering Unequally: Global and Disciplinary Bias in LLM Reconstruction of Scholarly Coauthor Lists",
      "author": ["Ghazal Kalhor", "Afra Mashhadi"],
      "datePublished": "2025-11-01",
      "identifier": "arXiv:2511.00476",
      "url": "https://arxiv.org/abs/2511.00476",
      "description": "Prompts three LLMs (DeepSeek R1, Llama 4 Scout, Mixtral 8x7B) to reconstruct academic coauthor lists from parametric memory and compares the output against bibliographic reference data, finding a systematic advantage for highly cited researchers — coauthor information for already-prominent scholars is recalled more reliably — though the authors stress this is non-uniform across fields and regions (Clinical Medicine and certain African regions show narrower gaps). Cited as grounding for the parametric-channel naming probe (ADR-0011): what an LLM recalls and names is gated by prior citation authority, so the probe's success is predictably and unequally distributed rather than a neutral readout of presence. Also supports manifesto Open Question 9 on whether diffusion-era recall entrenches existing prominence; we cite only its measured coauthor-list scope and its own warning against using LLM-generated scholarly networks without auditing for memorization-driven inequality."
    },

    {
      "@id": "https://arxiv.org/abs/2604.02544",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Developer Experience with AI Coding Agents: HTTP Behavioral Signatures in Documentation Portals",
      "author": "Oleksii Borysenko",
      "datePublished": "2026-04-02",
      "identifier": "arXiv:2604.02544",
      "url": "https://arxiv.org/abs/2604.02544",
      "description": "Server-side measurement of how AI coding agents and assistant services access developer documentation, from recorded HTTP headers: agents use heterogeneous HTTP clients — only some executing page scripts — carry identifiable signatures, and compress multi-page browsing into one or two requests, defeating client-side engagement analytics (session depth, bounce rate) while consuming the content in full. External endpoint-level grounding for three positions at once: the B2A/coding-agent niche as the llms.txt convention's validated remaining channel (ADR-0009's rescope), statically-served structured entry points as an access floor for the dual entry point (ADR-0006), and the structural blindness of human-attention engagement metrics (ADR-0007). Boundary: the request-compression evidence comes from documentation-retrieval tasks and is not generalized here to all agent workloads."
    },

    {
      "@id": "https://arxiv.org/abs/2603.21658",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "A Comparative Analysis of LLM Memorization at Statistical and Internal Levels: Cross-Model Commonalities and Model-Specific Signatures",
      "author": ["Bowen Chen", "Namgi Han", "Yusuke Miyao"],
      "datePublished": "2026-03-23",
      "identifier": "arXiv:2603.21658",
      "url": "https://arxiv.org/abs/2603.21658",
      "description": "Compares memorization across open model families (Pythia, OpenLLaMA, StarCoder, OLMo generations) at both statistical and internal levels, finding memorization rate scales log-linearly with model size across families while the attention heads important for memorization are distributed family-specifically — cross-model commonalities coexisting with family-specific signatures determined by the training recipe. Calibration grounding for the retrieval-suppressed naming probe (ADR-0011): the shared statistical regularities are what make cross-model probing meaningful at all, and the family-specific internals are why probe results are calibrated per family, reported with family-internal consistency rather than one cross-model threshold, and re-baselined on generation changes. Boundary: the internal analysis requires open weights, so for closed frontier models the family-calibration requirement can only be approximated externally."
    },

    {
      "@id": "https://arxiv.org/abs/2604.05074",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Memory Dial: A Training Framework for Controllable Memorization in Language Models",
      "author": ["Xiangbo Zhang", "Ali Emami"],
      "datePublished": "2026-04-06",
      "identifier": "arXiv:2604.05074",
      "url": "https://arxiv.org/abs/2604.05074",
      "description": "Introduces a training objective with a single hyperparameter interpolating between standard training and a temperature-sharpened memorization objective, showing memorization pressure can be raised continuously — retaining more seen sequences without degrading unseen performance, with larger models responding more sharply. Cited as the training-configuration-dependent counter-reading to the content-intrinsic reading of the Layer 3 wager (idea-versus-scaffold separation, grounded in default-training memorization evidence): under raised memorization pressure even rare scaffold-like sequences can be verbatim-retained, so whether scaffold dissolves may be decided by the ingesting pipeline's training regime rather than by the content's character. Recorded as an open tension, not a refutation — the two readings describe different training regimes and their reconciliation is unresolved."
    },

    {
      "@id": "https://arxiv.org/abs/2604.04700",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Who is the author? A legal and normative view of authorship in Generative AI-aided academic works",
      "author": "David M. Pereira",
      "datePublished": "2026-04-06",
      "identifier": "arXiv:2604.04700",
      "url": "https://arxiv.org/abs/2604.04700",
      "description": "Argues authorship under generative-AI assistance functions as a qualitative threshold rather than a binary attribute, grounded in European copyright law's human-intellectual-creation requirement: authorship holds where AI serves as cognitive support under the person's intellectual direction and becomes legally contestable where AI output supplants creative autonomy. Legal grounding for the channel split recorded on the exclusivity-to-derivation axis — proof of human creative control (legal-exclusivity channel) is a different certification system from content-derived and registry identifiers for origin-claim priority (diffusion channel) — and adjacent grounding for origin-claim falsifiability (ADR-0018): disclosure of the degree and nature of AI assistance is itself a scoped origin claim. Boundary: a legal-normative analysis, not an empirical measurement; case-law thresholds it discusses remain unsettled."
    },

    {
      "@id": "https://arxiv.org/abs/2506.17585",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Cite Pretrain: Retrieval-Free Knowledge Attribution for Large Language Models",
      "author": ["Yukun Huang", "Sanxing Chen", "Jian Pei", "Manzil Zaheer", "Bhuwan Dhingra"],
      "datePublished": "2025-06-21",
      "identifier": "arXiv:2506.17585",
      "url": "https://arxiv.org/abs/2506.17585",
      "description": "Shows a model can be trained to cite its pretraining documents without inference-time retrieval: active indexing — synthetic data restating each fact in diverse compositional forms and enforcing bidirectional source-to-fact and fact-to-source binding during continual pretraining — yields citation-precision gains up to 30.2% over passive exposure, improving as synthetic augmentation scales. Supply-side grounding for the parametric channel (ADR-0008) and for the demand-side-observer positioning of the retrieval-suppressed naming probe (ADR-0011): parametric attribution is a designable property of the ingesting pipeline, and passive exposure volume alone does not create it — structural diversity of restatement does, which externally supports this framework's preference for structurally diverse publication over sheer repetition. Boundary: the mechanism operates at the model provider's training stage; an author controls only the structure of what is published, not whether any provider applies such indexing."
    },

    {
      "@id": "https://arxiv.org/abs/2605.02392",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Is It Novel and Why? Fine-Grained Patent Novelty Prediction Based on Passage Retrieval",
      "author": ["Valentin Knappich", "Anna Hätty", "Simon Razniewski", "Annemarie Friedrich"],
      "datePublished": "2026-05-04",
      "identifier": "arXiv:2605.02392",
      "url": "https://arxiv.org/abs/2605.02392",
      "description": "Argues patent novelty assessment should shift from binary claim-level classification to feature-level analysis, releasing FiNE-Patents — 3,658 claims annotated from institutional examiner search opinions — where models decompose a claim into features and identify which prior-art passages disclose each. Procedural grounding for origin-claim scope discipline and origin-claim falsifiability (ADR-0018) at a finer granularity than whole-claim checking: the feature-to-prior-art map is the verification artifact. Cited with the paper's own caution that granularity is not accuracy — claim-level binary predictors can outscore feature-level pipelines on raw correctness while learning spurious surface correlations — so the discipline keeps the map and declines to reduce it to a binary novelty verdict. Boundary: patent claims have explicit feature boundaries; transfer of the procedure to non-formulaic prose such as essays and design records is untested."
    },

    {
      "@id": "https://arxiv.org/abs/2605.06635",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Cited but Not Verified: Parsing and Evaluating Source Attribution in LLM Deep Research Agents",
      "author": ["Hailey Onweller", "Elias Lumer", "Austin Huber", "Pia Ramchandani", "Vamse Kumar Subbiah", "Corey Feld"],
      "datePublished": "2026-05-07",
      "identifier": "arXiv:2605.06635",
      "url": "https://arxiv.org/abs/2605.06635",
      "description": "Evaluates citations in LLM deep-research reports on three dimensions — link validity, topical relevance, factual support — finding frontier systems keep link validity above 94% and relevance above 80% while factual accuracy of what each citation is claimed to support runs only 39–77%, and that scaling tool calls from 2 to 150 drops fact-check accuracy by roughly 42% on average. Grounding for the surface-versus-substantive verifiability split recorded on Two-Channel Attribution Diffusion (ADR-0008): a working retrieval channel proves reachability of the source, not fidelity of the claim it carries, and deeper retrieval does not converge on fidelity — so diffusion breadth is not transmission fidelity, a caution the scarcity-to-diffusion axis inherits."
    },

    {
      "@id": "https://arxiv.org/abs/2509.04499",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "DeepTRACE: Auditing Deep Research AI Systems for Tracking Reliability Across Citations and Evidence",
      "author": ["Pranav Narayanan Venkit", "Philippe Laban", "Yilun Zhou", "Kung-Hsiang Huang", "Yixin Mao", "Chien-Sheng Wu"],
      "datePublished": "2025-09-02",
      "identifier": "arXiv:2509.04499",
      "url": "https://arxiv.org/abs/2509.04499",
      "description": "Audit framework for generative search and deep-research systems reporting citation accuracy of roughly 40–80% across audited systems, substantial fractions of unsupported statements, and one-sided high-confidence answers on debate queries. Independent corroboration — by a different team and method than the Cited-but-Not-Verified line — of the surface-versus-substantive verifiability gap in the retrieval channel (ADR-0008): citation presence and citation faithfulness diverge across current systems."
    },

    {
      "@id": "https://arxiv.org/abs/2604.21300",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Explainable Disentangled Representation Learning for Generalizable Authorship Attribution in the Era of Generative AI",
      "author": ["Hieu Man", "Van-Cuong Pham", "Nghia Trung Ngo", "Franck Dernoncourt", "Thien Huu Nguyen"],
      "datePublished": "2026-04-23",
      "identifier": "arXiv:2604.21300",
      "url": "https://arxiv.org/abs/2604.21300",
      "description": "Identifies content-style entanglement — attribution models learning spurious style-topic correlations — as a core flaw in authorship attribution and answers it architecturally (a style encoder with contrastive learning plus a variational autoencoder that explicitly separates style from content), with state-of-the-art attribution results and few-shot generalization to AI-generated-text detection. Cited for the attribution paradox it implies for vocabulary discipline (ADR-0010): the more semantically unpredictable a coinage, the weaker style-based traceability becomes — so a coined term's attribution power is better read as occupancy of an uncontested concept-space coordinate (idea-level signature) than as personal style, which model imitation erodes. Recorded as a tension refining the discipline, not as a validation of any tactic."
    },

    {
      "@id": "https://arxiv.org/abs/2602.13123",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "From sunblock to softblock: Analyzing the correlates of neology in published writing and on social media",
      "author": ["Maria Ryskina", "Matthew R. Gormley", "Kyle Mahowald", "David R. Mortensen", "Taylor Berg-Kirkpatrick", "Vivek Kulkarni"],
      "datePublished": "2026-02-13",
      "identifier": "arXiv:2602.13123",
      "url": "https://arxiv.org/abs/2602.13123",
      "description": "Analyzes the correlates of word creation in published writing and on a social platform in parallel, finding the same correlates broadly hold in both domains while topic-popularity growth contributes less to neology on the social platform — evidence that survival factors are pathway-dependent. Grounding for the pathway-scoping of the anchor-densely rule (ADR-0010): semantic predictability of a new form relative to existing vocabulary — the property dense anchoring engineers — is supported across domains, but anchoring priorities differ by diffusion pathway, and this framework's rule is calibrated on the published/scholarly pathway it primarily operates in."
    },

    {
      "@id": "https://arxiv.org/abs/2602.19019",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "TokenTrace: Multi-Concept Attribution through Watermarked Token Recovery",
      "author": ["Li Zhang", "Shruti Agarwal", "John Collomosse", "Pengtao Xie", "Vishal Asnani"],
      "datePublished": "2026-02-22",
      "identifier": "arXiv:2602.19019",
      "url": "https://arxiv.org/abs/2602.19019",
      "description": "A watermarking framework embedding secret signatures into both the text-prompt embedding and the initial latent noise of a generative image model, so that multiple distinct concepts (objects, styles) contributed by training data can be independently recovered and attributed from generated output. Cited as the proactive pole of the proactive-versus-passive distinction recorded on the exclusivity-to-derivation axis: high-precision derivation tracing that presupposes the source's prior consent and instrumentation of the training pipeline — the premise this framework's passive attribution diffusion deliberately does not share, since it targets models the author cannot instrument. Cited as axis structure only; not adopted as a tactic (the framework's identifier layer is settled by ADR-0013's intrinsic-identifier decision)."
    },

    {
      "@id": "https://arxiv.org/abs/2504.00255",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "SciReplicate-Bench: Benchmarking LLMs in Agent-driven Algorithmic Reproduction from Research Papers",
      "author": ["Yanzheng Xiang", "Hanqi Yan", "Shuyin Ouyang", "Lin Gui", "Yulan He"],
      "datePublished": "2025-03-31",
      "identifier": "arXiv:2504.00255",
      "url": "https://arxiv.org/abs/2504.00255",
      "description": "Benchmark of 100 algorithm-reproduction tasks from 36 NLP papers with a dual-agent framework (paper interpretation plus iterative implementation), finding the best-performing model reaches only 39% execution accuracy while reconstructing the algorithms' reasoning structure far more reliably, and identifying missing or unclear implementation documentation as the primary obstacle. External grounding for the abstract-doctrine-plus-worked-implementation pair: the gap between understanding a method and executing it is what a worked implementation exists to close, and the benchmark's missing-information categories (hyperparameters, numerical stabilization, implementation logic, coding strategy) enumerate what the implementation half must carry. Boundary: benchmark designs differ substantially in reported reproduction difficulty, so the 39% figure is cited as evidence of a structural gap, not as a universal rate."
    },

    {
      "@id": "https://doi.org/10.1002/leap.2059",
      "@type": ["ExternalReference", "ScholarlyArticle"],
      "name": "Science Behind a Paywall: Restricted Access Limits the Promise of Artificial Intelligence",
      "author": ["Haoyi Zheng", "Huichun Zhan"],
      "datePublished": "2026-04-28",
      "identifier": "10.1002/leap.2059",
      "url": "https://doi.org/10.1002/leap.2059",
      "description": "Argues in a scholarly-publishing journal that paywalls keep roughly half of scholarly full text out of AI systems' reach, degrading what models can learn from the scientific record, and — holding paywall removal unrealistic without sustainable funding for publishing infrastructure — proposes licensing partnerships plus retrieval-augmented access as the viable middle path. Cited on the enclosure-to-openness axis for two refinements recorded as tension: openness is a physical precondition of LLM-mediated diffusion, not an abstract preference; and a licensed intermediate state exists between open and closed that the axis's binary framing does not represent, whose funding model conflicts with the free openness this framework assumes and whose benefits may concentrate in large rights-holders. A non-arXiv primary source (journal article with resolvable DOI) admitted on the SSRN precedent."
    }
  ]
}