互通性不是二元判斷：為何 AI 代理需要關係強度治理永續證據

Evidence Infrastructure Analysis 014 觀察的不是「框架能不能互相對照」這個表面問題，而是 AI 代理在跨框架移動永續證據時，能否保持每一種關係原有的制度意義。國際站原文以六份受控 Direct Mapping Guide 為分析基礎，涵蓋 GRI、ESRS、TNFD、COSO、GHG Protocol Scope 3 與聯合國永續發展目標。六份指南都從同一組 128 個 canonical MME task positions 出發，卻不把相同證據在不同制度中的用途壓縮成單一「有對應／無對應」答案。 核心判斷是：互通性不是二元狀態。AI 代理若只收到 A 對應 B 的訊號，可能把候選關係當成適用性、把支撐關係當成證明、把高關聯當成充分證據，或把一個框架的路由結果誤寫成另一個框架的合規結論。可靠的機器執行需要同時讀取關係類型、關係強度、啟動條件、證據狀態、主張邊界、方法版本與仍須由授權人員作成的治理判斷。 本篇把這套能力稱為 Relationship-Governed Interoperability，中文可理解為「關係治理型互通性」。它不是六個框架共同採納的新名詞，也不是任何發布機構對 EMJ.LIFE 方法的認可。它是一個分析構念，用來說明 AI 代理何時可以 Route、何時必須 Hold、何時需要 Escalate，以及何時應保留 No Route。

Evidence Infrastructure Analysis 014 does not ask only whether sustainability frameworks can be cross-referenced. It asks whether an AI Agent can move evidence across frameworks without changing the institutional meaning of each relationship. The canonical edition draws on six controlled Direct Mapping Guides covering GRI, ESRS, TNFD, COSO, the GHG Protocol Scope 3 Standard and the United Nations Sustainable Development Goals. All six begin from the same 128 canonical MME task positions, but they do not compress the use of common evidence into a binary mapped or unmapped result. The central proposition is that interoperability is not binary. If an Agent receives only an A-to-B signal, it may treat candidate relevance as applicability, support as proof, a strong route as sufficient evidence, or a route in one framework as a compliance conclusion in another. Machine execution requires relationship type, relationship strength, activation conditions, evidence state, claims boundaries, authority and method version, together with the governance decisions that remain reserved for authorised people. This edition describes that capability as Relationship-Governed Interoperability. It is an independent EMJ.LIFE analytical construct, not terminology adopted by the six framework institutions and not evidence of their approval. The purpose is to govern when an Agent may Route, when it must Hold, when it must Escalate and when No Route is the correct result.

多數傳統 crosswalk 把互通性簡化成「A 是否可映射至 B」。對熟悉制度邊界的人類分析者而言，這有時足以啟動下一步，因為人可以自行補上適用範圍、重大性、證據充分性、控制有效性或確信範圍的限制。但 AI 代理不應被要求從一個布林值自行重建這些差異。 六份 DOI 指南因此把關係拆成多個治理維度。Relationship Type 說明是直接、支撐、情境、條件式或控制相關的關係；Routing Relevance 說明是否值得導向下一程序；Route State 說明路由已啟動、待定或不應預設；Evidence Sufficiency 與 Assurance Readiness 分別回答證據是否足以支撐既定用途，以及是否具備進入確信程序的條件。涉及特定框架時，還需另外保存 Materiality、Control、Inventory 或 Contribution 狀態。 關係強度可標示為 HIGH、MEDIUM、CONDITIONAL、LOW 與 NO-DEFAULT-ROUTE，但這些標籤不是模型信心機率。HIGH 不代表揭露已滿足，MEDIUM 不是百分之五十的可信度，CONDITIONAL 不等於證據薄弱，NO-DEFAULT-ROUTE 也不表示系統故障。它們描述的是一項制度關係可以被如何使用，而不是 AI 對答案有多確定。 因此必須維持一條基本分界：Relationship Strength 不等於 Decision Confidence。前者屬於制度路由的治理屬性；後者若存在，則屬於模型或決策程序的另一種評估。兩者混用，會讓可重用的證據關係在自動化過程中膨脹成未經授權的結論。

Many crosswalks reduce interoperability to a question: does A map to B? A knowledgeable human may understand that the relationship does not automatically establish applicability, materiality, compliance, control effectiveness, inventory completeness, contribution or assurance. An AI Agent should not be expected to reconstruct those distinctions from a Boolean flag. The six DOI publications therefore separate multiple governance dimensions. Relationship Type identifies whether a route is direct, supporting, contextual, conditional or control-related. Routing Relevance identifies whether the evidence should enter a next process. Route State distinguishes active, pending, held and unavailable routes. Evidence Sufficiency asks whether evidence is adequate for its intended role. Assurance Readiness asks whether the evidence and controls are ready for an assurance procedure. Framework-specific states may separately address materiality, control, inventory or contribution. Relationship strength may be described as HIGH, MEDIUM, CONDITIONAL, LOW and NO-DEFAULT-ROUTE. These labels are not model confidence probabilities. HIGH does not mean a disclosure is satisfied. MEDIUM does not mean fifty per cent confidence. CONDITIONAL does not mean weak evidence. NO-DEFAULT-ROUTE does not mean the system failed. They describe how an institutional relationship may be used. The boundary must remain explicit: Relationship Strength is not Decision Confidence. The first is a governance attribute of a route. The second, if used, belongs to a separate model or decision assessment. Combining them allows reusable evidence relationships to expand into conclusions that the original source never authorised.

人類看到「Activity A 指向 TNFD」時，通常知道這只表示可能相關，仍需判斷實體背景、自然相關依賴與衝擊、重大性、適用揭露及 LEAP 程序。AI 代理若只收到 TNFD=true，卻可能直接檢索、排序、合併並把結果送入下一工作流，最後生成看似合理但制度上越界的結論。 同樣的錯誤會在六個框架中以不同形式出現。GRI 的候選路由不能取代組織對 impact materiality 的判斷；ESRS 的 PENDING-DMA 不能被改寫成揭露適用；TNFD 的關聯不能證明 LEAP 已完成；COSO 的 CONTROL-SUPPORT 不能證明控制設計、建置或運作有效；Scope 3 的類別路由不能證明盤查邊界完整或排放數量正確；SDG Target 關係也不能證明貢獻、成果或目標達成。 風險不在於映射本身，而在於映射變成可執行關係後，代理可以沿著關係連續推論。每一步都可能看似合理，累積後卻超過原始證據的權限。例如「Scope 3 相關」被擴張為「類別適用」，再擴張為「盤查完整」，最後變成「排放結論」。EIA-014 把這種風險稱為 Relationship Inflation，也就是關係膨脹。 真正的治理問題因而不是能否製作更大的 crosswalk，而是 AI 在採取行動之前必須知道什麼。至少要知道關係為何存在、強度屬於哪一級、哪些事實會啟動它、證據目前處於何種狀態、哪些主張被禁止，以及哪個節點仍需要授權決策。

A human analyst who sees Activity A routed toward TNFD will usually understand that entity context, nature-related dependencies and impacts, materiality, an applicable disclosure and a LEAP assessment remain to be considered. An Agent receiving only TNFD=true may retrieve, rank, combine and execute the relationship in another workflow, producing an institutionally invalid conclusion from a technically correct mapping. The same failure takes different forms across the six environments. A GRI route does not determine impact materiality. An ESRS route marked PENDING-DMA does not establish disclosure applicability. A TNFD relationship does not demonstrate that LEAP has been completed. COSO CONTROL-SUPPORT does not prove that a control is designed, implemented or operating effectively. A Scope 3 route does not establish category applicability, inventory completeness or emissions quantity. An SDG Target relationship does not demonstrate contribution, outcome or achievement. The risk becomes greater when machine-readable relationships propagate. Scope 3 relevance can become category applicability, then inventory coverage and finally an emissions conclusion. Each step may appear plausible while the final claim exceeds the authority of the first relationship. EIA-014 calls this Relationship Inflation. The governing question is therefore not whether a larger mapping can be built. It is what the Agent must know before it is allowed to act. At minimum, it needs the reason for the relationship, its governed strength, the facts that activate it, the current evidence state, prohibited claims and the point at which an authorised human decision is still required.

從二元映射轉向受治理關係，可以把機器可執行路徑寫成：Evidence Object → Relationship Type → Relationship Strength → Activation Condition → Evidence State → Claims Boundary → Governance Action。每一層回答不同問題，不能互相代替。 Canonical Evidence Identity 讓同一營運證據在進入多個框架路由時仍保持穩定身分；Typed Relationships 讓代理辨識直接、支撐、情境、條件式與控制相關關係；Relationship Strength 記錄制度上的路由相關性；Activation Conditions 限定只有在實體、期間、地點、方法或重大性事實成立後，條件式路由才可啟動；Independent Evidence States 則把路由價值與證據充分性、確信準備度分開。 Claims Boundary 必須以機器可讀方式指出禁止推論。例如「可導向 GRI 主題評估」不等於「該主題具重大性」；「可支撐 COSO 控制程序」不等於「控制有效」；「與 SDG Target 有關」不等於「產生可衡量貢獻」。Authority and Version Control 還需把關係連回適用方法、發布狀態與版本，讓後續使用者能判斷當時哪一套規則有效。 代理的正確流程應是 Evidence → Relationship → Strength → Governance Check → Route / Hold / Escalate / No Route，而不是 Find Mapping → Generate Answer。當關係不存在或條件不足時，NO-DEFAULT-ROUTE 是積極的治理指令：不要以最近似概念填補空白，應停止、保留負面結果、等待實體特定事實，或把判斷升級給具權限的人員。 落地時，每一筆關係紀錄至少應含 evidence_id、relationship_type、strength、activation_condition、evidence_state、claims_boundary、authority、method_version、effective_date、review_owner 與 decision_state。任何狀態改變都應新增版本，而不是覆寫先前判斷。只有如此，系統才能在日後重建代理為何路由、為何停止，以及哪一項人為決策改變了結果。

Moving from binary mappings to governed relationships changes the executable path to: Evidence Object → Relationship Type → Relationship Strength → Activation Condition → Evidence State → Claims Boundary → Governance Action. Each layer answers a different question and cannot be substituted for another. Canonical Evidence Identity keeps the same operational evidence identifiable across several framework routes. Typed Relationships distinguish direct, supporting, contextual, conditional and control-related uses. Relationship Strength records institutional routing relevance. Activation Conditions prevent a conditional route from becoming active before entity, period, location, method or materiality facts are present. Independent Evidence States keep routing relevance separate from Evidence Sufficiency and Assurance Readiness. Claims Boundaries should be machine-readable. A route to GRI topic assessment is not a materiality conclusion. A COSO support route is not control effectiveness. An SDG relationship is not measurable contribution. Authority and Version Control must connect the route to its governing methodology, release status and effective version so a later user can reconstruct which rule applied at the time. The correct reasoning path is Evidence → Relationship → Strength → Governance Check → Route / Hold / Escalate / No Route, not Find Mapping → Generate Answer. NO-DEFAULT-ROUTE is a positive governance instruction. The Agent should not fill an institutional gap with the nearest concept. It may need to stop, preserve a negative result, wait for entity-specific facts or escalate the decision. An implementable relationship record should contain at least evidence_id, relationship_type, strength, activation_condition, evidence_state, claims_boundary, authority, method_version, effective_date, review_owner and decision_state. A state change creates a new version rather than overwriting the earlier decision. That record makes it possible to reconstruct why an Agent routed, why it stopped and which authorised decision changed the outcome.

本次證據鏈分成三層，避免以來源數量製造虛假獨立性。第一層是 sustainabilitynewsnetwork.net 的 EIA-014 正式原文，支持系列身分、原始英文內容、作者、發布日期與分析邊界；它是同步來源，不計入外部證據。第二層是六個直接 DOI 連結，分別對應 DMG01 至 DMG06。它們是 EIA-014 的 publication basis 與 source data，證明六份受控方法出版物的持久識別、版本、發布日期、128 個任務位置及各框架的限制；由於發布者同為 EMJ LIFE HOLDINGS PTE. LTD.，不把它們當成獨立外部背書。 第三層由 17 筆獨立官方或標準制定機構來源構成，超過 15 筆門檻。GRI Universal Standards、IFRS S1、IFRS S2、SASB 國際適用性更新、GRI 與 IFRS Foundation 的溫室氣體互通性資源，以及 IFRS Foundation 與 EFRAG 的互通性指南，支持不同報導架構擁有不同目的、重大性視角與使用者需求。EUR-Lex 的 ESRS 法源及 EFRAG 的 materiality 與 value-chain implementation guidance，支持 ESRS 適用、雙重重大性與價值鏈資料不能由技術 mapping 取代。 TNFD Recommendations 與 LEAP guidance 支持自然相關辨識、評估與揭露是獨立程序；COSO ICSR 與 COSO/WBCSD ERM guidance 支持控制與風險治理不能由關聯性推定；GHG Protocol Scope 3 Standard 支持價值鏈排放盤查的類別、方法與邊界要求；聯合國 2030 Agenda 與全球指標框架分別界定 Goals、Targets 與 Indicators 的制度層次；W3C PROV-O 則支持跨系統保存 entity、activity、agent 與 provenance 關係。 所有外部來源逐筆保存標題、機構、角色、出版日期、查核日期與支撐範圍。來源能支撐框架邊界與技術背景，但不表示這些機構審查、批准或認可 Relationship-Governed Interoperability、Relationship Inflation，或 EMJ.LIFE 的 128 個 MME task positions。

The evidence chain is deliberately separated into three layers so that source volume does not create false independence. The first layer is the canonical EIA-014 publication on sustainabilitynewsnetwork.net. It supports the series identity, original English analysis, author, publication date and analytical boundary. It is the synchronisation source and is not counted as external evidence. The second layer contains the six direct DOI links for DMG01 through DMG06. They are the publication basis and source data for the analysis. They support persistent identity, version, publication date, the common 128 MME task positions and the stated limitations for each framework. Because all six are issued by EMJ LIFE HOLDINGS PTE. LTD., they are not represented as independent external validation. The third layer contains seventeen independent official or standard-setting sources, exceeding the required fifteen-source threshold. GRI Universal Standards, IFRS S1, IFRS S2, the SASB international-applicability update, the joint GRI and IFRS Foundation GHG interoperability resource, and the IFRS Foundation and EFRAG interoperability guidance support the distinct purposes, users and materiality perspectives of reporting systems. The ESRS legal text on EUR-Lex and EFRAG materiality and value-chain guidance support the boundaries that a technical mapping cannot replace. TNFD Recommendations and LEAP guidance support nature-related identification, assessment and disclosure as separate processes. COSO ICSR and COSO/WBCSD ERM guidance support the separation between a relationship and control or risk conclusions. The GHG Protocol Scope 3 Standard supports category, method and inventory-boundary requirements. The UN 2030 Agenda and Global Indicator Framework distinguish Goals, Targets and Indicators. W3C PROV-O supports preservation of entities, activities, agents and provenance across systems. Every external record includes title, institution, source role, publication date, access date and claims supported. These records support framework and technical boundaries. They do not imply that any issuing institution reviewed or endorsed Relationship-Governed Interoperability, Relationship Inflation or the 128 MME task architecture.

對企業而言，關係治理型互通性改變的不只是資料模型，也改變責任分工。永續、財務、風險、內控、資訊與確信團隊不能把 mapping table 當成完成的判斷。資料團隊負責保存穩定身分、版本與轉換歷程；框架負責人辨識關係類型與啟動條件；管理階層或治理單位負責重大性、適用性與主張授權；內控與內稽測試關係規則是否按設計運作；外部確信則只對約定範圍形成結論。 對技術供應商而言，搜尋相關性、向量相似度或生成模型信心都不能直接替代制度關係強度。系統可以建議候選路由，但需要把候選、已啟動、受限、待決、無預設路由等狀態明確分開。介面也應讓使用者看到觸發條件與禁止主張，而不是只顯示一個「已匹配」標記。 對準則制定者與監理者而言，interoperability 不只涉及共同欄位或一致 taxonomy。當兩個框架面向不同使用者、決策、重大性或責任結構時，同一證據可在上游重用，下游治理判斷仍需重新執行。EIA-014 的原則是：Reuse the evidence. Re-perform the governance decision. 因此，同一筆證據對某框架可能是 HIGH、對另一框架是 MEDIUM、對第三個框架是 CONDITIONAL，對第四個框架則是 NO-DEFAULT-ROUTE。這不是不一致，而是制度問題不同。成熟的互通性應保留差異、揭露轉換並限制推論距離，而不是追求看似完整的等價矩陣。

For enterprises, Relationship-Governed Interoperability changes both the data model and the allocation of responsibility. Sustainability, finance, risk, internal control, technology and assurance functions cannot treat a mapping table as a completed decision. Data teams preserve identity, version and transformation history. Framework owners define relationship types and activation conditions. Management or governance bodies authorise materiality, applicability and claims. Internal control and audit test whether relationship rules operate as designed. External assurance reaches conclusions only within an agreed scope. For technology providers, retrieval relevance, vector similarity and model confidence cannot substitute for institutional relationship strength. A system may suggest candidate routes, but candidate, active, conditional, held, pending and no-default-route states must remain distinct. Interfaces should expose activation conditions and prohibited claims instead of displaying only a matched indicator. For standard setters and regulators, interoperability extends beyond shared fields or aligned taxonomies. When two frameworks serve different users, decisions, materiality perspectives or accountability structures, common evidence may be reused upstream while the downstream governance decision must be performed again. The practical principle is: Reuse the evidence. Re-perform the governance decision. The same evidence can therefore have a HIGH relationship to one framework, MEDIUM to another, CONDITIONAL to a third and NO-DEFAULT-ROUTE to a fourth. That is not inconsistency. It reflects different institutional questions. Mature interoperability preserves differences, records transformations and limits inference distance instead of forcing apparent equivalence.

對台灣產業而言，六個國際框架會透過不同路徑進入企業與供應鏈。上市櫃公司可能同時準備 IFRS 永續揭露、客戶要求的 GRI 或 ESRS 資料、自然相關評估、內控治理與 Scope 3 盤查；電子、半導體、製造、金融及出口供應鏈還會面對集團總部、國際客戶、金融機構與確信服務提供者的多重資料要求。把這些要求合併成一張「框架對照表」會降低重複蒐集，卻也可能掩蓋重大性、實體邊界、控制狀態、排放類別與貢獻主張的差異。 台灣企業可在揭露前證據基礎設施（Pre-Disclosure Evidence Infrastructure）中，先為工廠、設備、供應商問卷、採購交易、排放因子、計算模型、核准與確信程序建立穩定 evidence identity。每一個跨框架路由都應保存 relationship type、strength、activation condition、claims boundary、方法版本與責任人。中文與英文揭露應指向同一受治理主張，避免翻譯版本形成另一套未受控結論。 供應鏈資料尤其不能因「可對應 Scope 3、ESRS 或 TNFD」就被視為已驗證。企業應區分供應商陳述、估算、原始憑證、平台轉換、內部覆核與第三方確信，並讓確信服務能追溯每一段轉換。當條件不足時，系統應保留 Hold 或 Escalate，而不是為了完成報表自動補出最接近答案。 對台灣的董事會、審計委員會與確信服務市場，重要問題不只是採用哪一個框架，而是誰有權把候選關係升級為可披露主張、哪些控制驗證了狀態轉換、以及新證據出現後舊結論是否仍有效。這是 SNN.TW 的台灣產業傳導解讀，不代表台灣主管機關新增統一規定，也不構成報導、法律、盤查或確信意見。

For Taiwan industries, the six international systems can enter the enterprise and supply chain through different routes. Listed companies may prepare IFRS sustainability disclosures while responding to customer requests based on GRI or ESRS, performing nature-related assessment, governing internal controls and compiling Scope 3 inventories. Electronics, semiconductor, manufacturing, finance and export supply chains may also receive overlapping requests from group headquarters, global customers, financial institutions and assurance providers. A single framework crosswalk may reduce repeated collection while concealing differences in materiality, entity boundary, control state, emissions category and contribution claims. Pre-Disclosure Evidence Infrastructure allows Taiwan enterprises to assign stable evidence identities to factories, equipment, supplier questionnaires, procurement transactions, emission factors, calculation models, approvals and assurance procedures before reporting begins. Each cross-framework route can retain relationship type, strength, activation condition, claims boundary, method version and accountable owner. Chinese and English disclosures should point to the same governed claim instead of creating a second uncontrolled conclusion through translation. Supply-chain information must not become verified merely because it can be routed to Scope 3, ESRS or TNFD. Enterprises should distinguish supplier representation, estimation, original record, platform transformation, internal review and independent assurance, while enabling assurance services to trace each transformation. When conditions are missing, the system should preserve Hold or Escalate instead of generating the nearest answer to complete a report. For Taiwan boards, audit committees and the assurance-services market, the important questions are who may promote a candidate relationship into a disclosure claim, which controls validated the state transition and whether new evidence affects the continuing validity of an earlier conclusion. This is an SNN.TW Taiwan industry transmission analysis. It is not a new requirement issued by a Taiwan authority and does not constitute reporting, legal, inventory or assurance advice.

下一階段的重點不是再增加無限多的對照欄位，而是驗證關係治理能否在真實工作流中保持穩定。企業可先選取跨 GRI、ESRS、TNFD、COSO、Scope 3 與 SDGs 的高重用證據，測試同一 evidence object 在不同路由下是否保留身分、方法、期間、實體與權限，並確認代理能在 NO-DEFAULT-ROUTE 時停止。 實務上可建立四項最小測試。第一，讓獨立覆核者從最終揭露回到原始證據與每一段關係。第二，改變方法版本或重大性決定，確認系統只更新受影響路由。第三，移除啟動條件，確認代理把狀態改為 Hold，而非繼續生成結論。第四，測試人員能否辨識關係強度、模型信心與證據充分性是三個不同欄位。 六份 DOI 指南提供一套受控研究基礎，17 筆獨立外部來源提供框架與技術邊界，但兩者都不能替代個別企業的適用性、重大性、控制、盤查、貢獻或確信判斷。未來真正需要觀察的是：AI 代理能否重用證據，同時不改變關係的制度意義。

The next stage is not an unlimited expansion of crosswalk fields. It is testing whether relationship governance remains stable in real workflows. Enterprises can select high-reuse evidence across GRI, ESRS, TNFD, COSO, Scope 3 and the SDGs, then test whether the same evidence object retains its identity, method, period, entity and authority under each route. They should also confirm that the Agent stops when a relationship has NO-DEFAULT-ROUTE. Four minimum tests are useful. First, an independent reviewer should travel from a final disclosure to source evidence and every relationship transition. Second, a method-version or materiality change should update only affected routes. Third, removing an activation condition should move the Agent to Hold rather than allowing continued inference. Fourth, users should be able to distinguish relationship strength, model confidence and evidence sufficiency as separate fields. The six DOI publications provide a controlled research basis and the seventeen independent external records provide framework and technical boundaries. Neither layer replaces entity-specific decisions on applicability, materiality, controls, inventories, contribution or assurance. The continuing question is whether an AI Agent can reuse the evidence without changing what the relationship means.