Jobin, Ienca & Vayena (2019) analyzed 84 AI ethics documents globally in Nature Machine Intelligence and found convergence on five core principles: transparency, justice and fairness, non-maleficence, responsibility, and privacy. However, they found "substantial divergence on interpretation, importance, domain applicability, and implementation." Floridi et al. (2018) synthesized these into a coherent ethical framework (AI4People). The convergence is on what matters. The divergence is on how to operationalize it. This divergence is not a disagreement about values — it is a consequence of missing infrastructure. Without a shared substrate for implementing governance, each organization invents its own approach.

Mittelstadt (2019) demonstrated in Nature Machine Intelligence that AI ethics initiatives implicitly follow a medical ethics paradigm — but medicine has three things AI development lacks: (1) common aims and fiduciary duties binding practitioners, (2) centuries of professional history establishing norms, and (3) proven methods to translate abstract principles into operational practice (clinical guidelines, ethics review boards, malpractice liability). Without these, Mittelstadt concluded, principled approaches are "unlikely to succeed." This is the sharpest diagnosis of the principles-practice gap: the failure is not in the principles but in the absence of translation infrastructure.

Hagendorff (2020) evaluated 22 AI ethics guidelines in Minds and Machines. He found that most guidelines omit entire governance domains (democratic control, ecological networks, care ethics) and that the translation from social aspiration to operational specification is "non-trivial" — meaning no one has solved it. The gap between principle and practice is not a temporary implementation delay but a structural deficit: the translation mechanism is missing.

The EU AI Act (2024) mandates risk management, data governance, technical documentation, record-keeping, transparency, and human oversight (Articles 9-15). NIST AI RMF (2023) specifies GOVERN/MAP/MEASURE/MANAGE. ISO/IEC 42001 (2023) provides the first certifiable AI management system standard. OECD Principles (2019, updated 2024), adopted by 47 nations, establish five recommendations. IEEE P7000 series requires "discoverable decision basis" and "unambiguous rationale for all decisions." Singapore's Model AI Governance Framework (updated January 2026) includes the first global framework for agentic AI governance. China's AI Safety Governance Framework (2024) requires algorithm filing from 1,400+ algorithms. Despite different legal traditions, all six converge on transparency, accountability, fairness, and human oversight — and all specify requirements without specifying the infrastructure to implement them. The regulatory landscape tells organizations WHAT to do. No framework provides the structural substrate for HOW to do it at the infrastructure level.

Rakova, Yang, Cramer & Chowdhury (2021) documented in ACM CSCW through interviews with practitioners at major technology companies three critical organizational findings: (1) responsible AI practitioners lack organizational authority — they can identify problems but cannot mandate solutions; (2) responsible AI is organizationally disconnected — ethics teams, engineering teams, and product teams lack a shared governance substrate; (3) responsible AI metrics are absent — organizations cannot measure governance effectiveness because they lack infrastructure to capture governance activities in queryable form. The failure is structural: organizations lack the structures to translate their own stated values into operational practice.

Selbst, Boyd, Friedler, Venkatasubramanian & Vertesi (2019) identified five traps at FAT* 2019: the Framing Trap (defining fairness computationally rather than socially), the Portability Trap (assuming solutions transfer across contexts), the Formalism Trap (reducing social concepts to mathematical constraints), the Ripple Effect Trap (ignoring how interventions reshape dynamics), and the Solutionism Trap (assuming technology alone solves social problems). Each trap occurs because organizations lack the infrastructure to maintain contextual richness through technical implementation. The Framing Trap occurs without infrastructure for integrating social and technical perspectives. The Portability Trap occurs when governance infrastructure is treated as context-independent. The Formalism Trap occurs without epistemic infrastructure that distinguishes what an organization declared from what an external authority established from what a system derived.

Diakopoulos (2016) established algorithmic accountability reporting in Communications of the ACM — systematic investigation of algorithms by sampling along key dimensions. Raji et al. (2020) proposed the SMACTR framework at FAccT (Scoping → Mapping → Artifact Collection → Testing → Reflection) — one of the few accountability frameworks specifying an operational process. But SMACTR requires organizational capacity for documentation collection, testing protocols, and structured reflection. Metcalf et al. (2021) identified the "assessor's regress" at FAccT — justifying an assessment requires another assessment, forming an endless chain. Without structural foundations (shared constructs, agreed invariants, common epistemic classification), assessment quality depends on assessor judgment, which must itself be assessed. The regress terminates only when structural governance primitives define what must be assessed and what counts as compliance.

Busuioc (2021) documented in Public Administration Review that AI creates automation bias — humans over-relying on automated recommendations — producing accountability gaps where responsibility diffuses across human-algorithm chains. She found that "the twin foundations of bureaucratic legitimacy — expertise and accountability — are being simultaneously diminished" by AI adoption. Her recommended solution: models with built-in audit trails of decision-making. The authority relationship between humans and AI is the critical governance challenge — when AI recommends and humans rubber-stamp, the governance question of who exercised authority becomes unanswerable without structural lineage.

Raji & Buolamwini (2019) demonstrated at AIES that the Gender Shades audit — identifying performance disparities in commercial facial recognition — drove all three audited companies (IBM, Microsoft, Megvii) to release improved API versions within seven months, reducing error rates by 17.7–30.4% for darker-skinned females. This demonstrates that governance infrastructure is often built reactively under external accountability pressure. The infrastructure for proactive, continuous accountability — detecting disparities before external audits find them — requires deviation measurement infrastructure that most organizations lack.

Amodei et al. (2016) defined five concrete AI safety problems: avoiding negative side effects, avoiding reward hacking, scalable oversight, safe exploration, and robustness to distributional shift. Each requires infrastructure: monitoring systems, reward specification processes, oversight mechanisms, safety evaluation protocols. Russell (2019) proposed that AI systems should maintain deference to humans as a core alignment mechanism — requiring preference learning, deference protocols, and uncertainty quantification. Christiano et al. (2017) developed RLHF at NeurIPS 2017 — now foundational alignment infrastructure for modern language models. Bai et al. (2022) proposed Constitutional AI — training harmless AI using written principles (constitutions) as governance infrastructure embedded in training objectives. The convergence: safety is a systems engineering problem, governance must be architectural (not advisory), and organizations using AI systems inherit alignment infrastructure they neither understand nor control.

Lipton (2018) showed in Communications of the ACM that different communities mean different things by "interpretability": transparency, post-hoc explanation, simulatability, decomposability. Without definitional infrastructure, organizations cannot implement interpretability requirements. Rudin (2019) argued in Nature Machine Intelligence that organizations should use inherently interpretable models rather than explaining black boxes — an infrastructure argument requiring different algorithms, pipelines, and incentive structures. Doshi-Velez & Kim (2017) proposed three evaluation approaches (application-grounded, human-grounded, functionally-grounded), each requiring different organizational infrastructure. Gunning's (2021) DARPA XAI retrospective demonstrated that explanation is a sociotechnical system, not a technical feature. Steyvers & Kumar (2024) identified three challenges for AI-assisted decision-making in Perspectives on Psychological Science: complementarity recognition, mental model accuracy, and interaction design — each requiring infrastructure for decision routing, model transparency, and cognitive load management.

S1 found the same pattern across provenance, design rationale, institutional memory, process mining, AI accountability, and audit compliance: the requirement is established, the consequences of absence are documented, but no infrastructure provides it as a by-product of operation. S2 confirms this pattern in the AI governance domain specifically. The 84+ ethics documents (F1), six regulatory frameworks (F4), and dozens of standards all specify requirements. The organizational failure documented by Rakova et al. (F5) and the translation failure documented by Mittelstadt (F2) and Hagendorff (F3) are consequences of the same structural gap. The convergence across S1 and S2 strengthens the meta-pattern: it holds for data provenance AND AI governance — two very different domains arriving at the same boundary.

Barocas & Selbst (2016) established in the California Law Review that algorithms inherit the prejudices of prior decision-makers through flawed training data — data mining finds correlations reflecting historic prejudice. Kaminski (2019) showed in the Berkeley Technology Law Journal that GDPR's algorithmic accountability regime is broader than commonly understood, creating requirements for data access, impact assessments, and human review. Wachter, Mittelstadt & Russell (2018) proposed counterfactual explanations in the Harvard Journal of Law and Technology — requiring computational infrastructure for explanation generation. Madaio et al. (2020) co-designed AI fairness checklists with 48 practitioners at CHI (Best Paper Award), finding that even checklists require cross-functional coordination, trained facilitators, and organizational commitment. The pattern: fairness is not a mathematical property of a model but a sociotechnical property requiring infrastructure for data quality assessment, bias detection, historical auditing, remediation tracking, and epistemic classification (distinguishing what was organizationally declared from what was externally validated from what was system-derived).