Situated Epistemic Infrastructures (SEI)

• Situated Epistemic Infrastructures (SEI) are complex socio-technical systems that integrate AI-driven relevance judgments with human and contextual mediation.
• They combine formal models, graph-based representations, and empirical methods to trace the distributed nature of authority and knowledge across scales.
• SEI frameworks enable adaptive governance and transparency by shifting epistemic agency from individuals to infrastructural processes, fostering equitable knowledge practices.

Situated Epistemic Infrastructures (SEI) designate socio-technical arrangements in which knowledge production, validation, circulation, and action are conditioned by embedded infrastructures rather than by isolated tools or bounded expert communities. Across the literature, SEI are described as context-specific deployments of cognitive infrastructures, as situation-specific instantiations of knowledge infrastructures, and as a diagnostic framework for tracing how authority is mediated across institutional, computational, and temporal arrangements under post-coherence conditions (Riva, 19 Jun 2025, Wittenborg, 6 Mar 2026, Kelly, 7 Aug 2025). In this formulation, AI systems do not merely assist reasoning on demand; they become ambient layers of epistemic mediation that automate relevance judgments, reshape attention and memory, and redistribute epistemic agency across humans, platforms, models, and organizational routines (Riva, 19 Jun 2025).

1. Genealogy and conceptual foundations

A central lineage of SEI begins with the shift from “AI as tool” to “AI as cognitive infrastructure.” In this account, infrastructural properties include embeddedness in socio-technical arrangements, transparency during normal operation, visibility upon breakdown, and the role of standards and conventions in shaping what counts as appropriate or true. Riva’s extension identifies distinctive AI characteristics: anticipatory personalization, adaptive invisibility, automation of relevance judgment, and a shift in the locus of epistemic agency from humans to non-human systems (Riva, 19 Jun 2025). SEI name the situated form of these processes: context-specific deployments that perform semantic transport, exercise anticipatory personalization, and enact adaptive invisibility.

A second lineage derives from knowledge infrastructure theory. Wittenborg reconstructs SEI from the concept of a knowledge infrastructure as “a robust network of people, artifacts, and institutions that generate, share, and maintain specific knowledge about the human and natural worlds,” and formalizes this as $KI=(P,A,I,R)$. In that reconstruction, a situated epistemic infrastructure is the domain-specific instantiation $SEI_d=(KI_d,C_d)$, where $C_d$ denotes contextual constraints and affordances such as legal, social, technical, and incentive structures (Wittenborg, 6 Mar 2026). This moves SEI away from a platform-centric description toward a graph of persons, artifacts, institutions, and their relations.

A third lineage is explicitly diagnostic and post-coherence. “Situated Epistemic Infrastructures: A Diagnostic Framework for Post-Coherence Knowledge” defines SEI as a framework for analyzing how knowledge becomes authoritative through the interaction of infrastructural, institutional, and temporal forces. It treats epistemic systems as hybrid assemblages of human actors, computational processes, and material-symbolic arrangements, and foregrounds coordination over classification (Kelly, 7 Aug 2025). This is a direct departure from coherence-based models that assume stable disciplinary domains, shared paradigms, and classification as the primary ordering device.

Taken together, these lineages indicate that SEI are not reducible to recommender systems, datasets, or interfaces in isolation. This suggests that the concept is best understood as a meso-level analytic for infrastructures that organize what can be known, who can know it, and under what temporal and institutional conditions that knowledge acquires authority.

2. Core mechanisms and formal models

Riva’s formulation makes the automation of relevance judgment explicit. The relevance score assigned to item $i$ for user $u$ at time $t$ is modeled as

$$r_t(i \mid u) = \sigma \left( w_s \cdot S(i) + w_p \cdot P(u,i,t) + w_c \cdot C(t) \right)$$

where $S(i)$ is the semantic embedding of the item, $P(u,i,t)$ captures personalization features, $C(t)$ encodes situational context, and $SEI_d=(KI_d,C_d)$0 are learned weights. The system then promotes the top-$SEI_d=(KI_d,C_d)$1 items by descending $SEI_d=(KI_d,C_d)$2, thereby automating judgment of what is “most relevant” (Riva, 19 Jun 2025). The same account formalizes epistemic agency through the allocation

$SEI_d=(KI_d,C_d)$3

where $SEI_d=(KI_d,C_d)$4 is the count or intensity of manual relevance selections and $SEI_d=(KI_d,C_d)$5 the count of algorithmic selections. As SEI deepen, $SEI_d=(KI_d,C_d)$6, representing near-complete outsourcing of epistemic agency (Riva, 19 Jun 2025).

The post-coherence framework introduces a different formalization, centered on credibility mediation across three arrangements: institutional $SEI_d=(KI_d,C_d)$7, computational $SEI_d=(KI_d,C_d)$8, and temporal $SEI_d=(KI_d,C_d)$9. Its conceptual discussion proposes

$C_d$0

with weights for situational relevance, institutional legitimacy metrics, algorithmic trustworthiness, and a temporal alignment factor (Kelly, 7 Aug 2025). The purpose of this model is not to reduce epistemic authority to a single scalar in practice, but to formalize how credibility is co-produced by bureaucratic routines, computational opacity, and polytemporal rhythms.

Wittenborg’s reconstruction supplies a structural formalism. Domain-specific SEI are represented as typed, labeled graphs:

$C_d$1

where vertices are persons, artifacts, and institutions; edges are collaboration, curation, governance, and related links; $C_d$2 types vertices; and $C_d$3 labels relations (Wittenborg, 6 Mar 2026). This graph-theoretic treatment complements the ranking and credibility models by emphasizing relational robustness, connectivity, and FAIRness.

These models describe different observables of the same broad phenomenon. One focuses on algorithmic selection, another on mediation of authority, and another on infrastructure topology. A plausible implication is that SEI research gains analytical traction when these levels are treated as complementary rather than competing abstractions.

3. Scales and instantiated forms

SEI are theorized and implemented across multiple scales. At the individual scale, Riva’s narrative of “Sarah” depicts cognitive dependency and habituation: an AI-curated digest, autopredicted needs, and pre-tagged documents gradually alter attention span, reasoning rhythms, and self-image as manual relevance judgments decline and $C_d$4 rises (Riva, 19 Jun 2025). At the collective scale, the “Mayor Chen” scenario depicts democratic deliberation transformed by emotionally charged amplification, algorithmic echo chambers, and reordered collective memory. At the societal scale, the “Commissioner Martinez” scenario frames SEI as policy objects, with affluent regions receiving full SEI suites while others rely on legacy systems, raising issues of transparency, equity, and cognitive justice (Riva, 19 Jun 2025).

In education, SEI are described as socio-material networks through which knowledge is created, validated, and shared. Chen analyzes MagicSchool AI and Brisk Teaching through three dimensions: affordances for skilled epistemic actions, support for epistemic sensitivity, and implications for long-term habit formation. Across both cases, the systems are said to undermine skilled epistemic actions, erode epistemic sensitivity, and promote passive habits, with efficiency-first interaction patterns risking the atrophy of lesson-planning expertise and feedback acuity (Chen, 9 Apr 2025). This use of SEI shifts analysis from performance convenience to infrastructural effects on professional judgment.

In AI alignment, Arzberger et al. operationalize a micro-scale SEI through reflexive annotating. Their probe invites annotators to articulate how positionality shapes subjective judgments, producing situated metadata in the form of highlighted text spans, identity tags, and rationales. The study involved 30 Prolific crowd workers and 5 follow-up interviews, and it reports that reflexive annotating elicits intersectional reasoning, positional humility, and viewpoint change while also surfacing affective burden (Arzberger et al., 25 Jan 2026). Here, SEI consist not in large-scale ranking systems but in the design of provenance-rich annotation interfaces.

In situated collaboration, TRACE implements an SEI for live common-ground tracking. It maintains three proposition banks—QBank, EBank, and FBank—and updates them from speech, gesture, gaze, and object manipulation. Epistemic stance is represented through four move types: STATEMENT, ACCEPT, DOUBT, and NONE (VanderHoeven et al., 12 Mar 2025). TRACE therefore instantiates SEI as a real-time infrastructure for maintaining shared epistemic state.

In organizational AI, OIDA argues that the relevant substrate is not retrieval fidelity but epistemic fidelity. Its typed Knowledge Objects distinguish decisions, hypotheses, contested claims, and open questions, while signed contradiction edges and class-specific decay maintain their importance over time. The primitive QUESTION-as-modeled-ignorance gives unknowns inverse decay so that unresolved issues become more urgent rather than fading from view (Bottino et al., 13 Apr 2026). This is a direct organizational instantiation of SEI as computable commitment, contradiction, and ignorance.

Science communication and aerospace engineering provide domain-specific KI/SEI cases. Wittenborg describes wiki- and knowledge-graph-based digital libraries, human-in-the-loop extraction workflows, stakeholder-driven interfaces, and graph-native representations in Wikibase, OWL, RDF, and ORKG (Wittenborg, 6 Mar 2026). Across these settings, SEI consistently denote infrastructures in which epistemic mediation is embedded into routine workflows rather than appended as an external audit layer.

4. Methods, workflows, and empirical visibility

A recurring methodological problem in SEI research is that infrastructural influence is difficult to detect precisely because it is embedded and adaptive. Riva’s proposed solution is the “infrastructure breakdown methodology”: place participants in a neutral digital environment where SEI operate in the background, habituate them for $C_d$5 hours, then disable ranking modules by setting $C_d$6 and observe changes in performance, strategy, and attention. The suggested observables are $C_d$7 Performance, $C_d$8 Strategy, and $C_d$9 Attention, including longer search paths and more revisits in eye-tracking (Riva, 19 Jun 2025). Comparative quasi-experiments then exploit platform updates or policy changes as natural breakdowns and track shifts in $i$0, reasoning quality, information diversity, and collective network structures (Riva, 19 Jun 2025).

Wittenborg’s workflow research exemplifies a different empirical strategy: tool-supported reconstruction of underdeveloped knowledge infrastructures. The SWARM-SLR framework and the ExtracTable module implement a linear pipeline from planning, search and retrieval, and NLP-based extraction through validation, knowledge-graph ingestion, and synthesis. Jupyter notebooks orchestrate modules $i$1–$i$2, with reference-manager APIs, spaCy, Transformers, JSON-Schema with a human GUI, and RDFLib plus a SPARQL endpoint (Wittenborg, 6 Mar 2026). Reported evaluations include validation of 65 SLR requirements via two online surveys of more than 40 researchers, a SUS score of 84.17 (“A+”) for ExtracTable, and a reduction from 4 hours–2 weeks to approximately 24 minutes per review cycle. Further results include 88% of SciCom Wiki testers reporting “substantial improvement” in findability, an OWL knowledge graph with more than 700 processes/software/data formats in aerospace, and 75% of respondents finding veracity scores “helpful” in computational fact-checking (Wittenborg, 6 Mar 2026).

TRACE contributes an online multimodal methodology. It operates at 5–6 frames per second on a 12-core Intel i7-12700H laptop with 16 GB RAM and an NVIDIA RTX 3070 Ti, using Faster–Whisper, openSMILE, Faster R-CNN ResNet50-FPN, and gesture/gaze grounding to update common-ground banks (VanderHoeven et al., 12 Mar 2025). Evaluation uses the Sørensen–Dice Coefficient over QBank, EBank, FBank, and FUE, with live TRACE achieving FUE-DSC scores in the 0.25–0.46 range and QBank-DSC up to 0.74 (VanderHoeven et al., 12 Mar 2025). The substitution study additionally identifies ASR as the largest source of downstream error.

OIDA introduces a composite evaluation methodology, the Epistemic Quality Score:

$i$3

where the components are Epistemic Classification Accuracy, Contextual Precision, Contextual Recall, Epistemic Coherence, and Decision Enablement (Bottino et al., 13 Apr 2026). In a controlled comparison over $i$4 response pairs, the OIDA RAG condition achieved EQS $i$5 versus $i$6 for the full-context baseline, with a $i$7 token budget difference identified as the primary confound; the decisive equal-token ablation was pre-registered and not yet run (Bottino et al., 13 Apr 2026). This explicitly frames SEI evaluation as an epistemic, not merely retrieval, problem.

5. Governance, actionability, and fairness

SEI theory is closely tied to governance. Riva’s framework proposes transparency standards that expose $i$8 to users or auditors, cognitive sovereignty interventions such as user-controlled sliders on personalization intensity, and regulatory paradigms that govern SEI as common-pool resources or public utilities with equity metrics based on $i$9 distributions (Riva, 19 Jun 2025). The post-coherence framework similarly argues for technologies of humility, anticipatory governance, disclosure of algorithmic provenance, prompt-sharing, calibrated review timelines, provenance logging, and interfaces that support user interrogation of algorithmic reasoning (Kelly, 7 Aug 2025). These recommendations treat epistemic stewardship as an infrastructural design problem rather than an after-the-fact compliance exercise.

Vogt’s Action Unit framework extends SEI from mediation to execution. An Action Unit is a compound semantic object with typed parts for input, plan, applicability conditions, objective, and output:

$u$0

Three AU types are distinguished—epistemic, transformational, and intervention—and conditional AUs implement graph-native decision support through the form “IF Cond(Context) THEN execute PlanSpec” (Vogt, 2 May 2026). The associated TripleA Principle—Actionability, Applicability, and Auditability—extends FAIR and CLEAR by requiring explicit operational structure, contextual validity checks, and provenance of execution (Vogt, 2 May 2026). In this view, SEI are not only infrastructures of knowing but infrastructures that bind representation, procedure, and goal.

The fairness literature pushes the governance question further by arguing that algorithmic systems function as epistemic infrastructures yet are usually audited only for predictive fairness. Quaresmini et al. formalize epistemic injustice as a deficit

$u$1

between ideal and realized conditions for features such as credibility, uptake, and epistemic agency (Quaresmini et al., 24 Apr 2026). They translate canonical distributive indices—Jain, Gini, Hoover, Generalized Entropy, Atkinson, Dissimilarity, Palma, and S80/S20—into an epistemic context and show in a recommender-mediated opinion-dynamics simulation that targeted or random boosts can increase resource inequality and reshape epistemic environments even when conventional fairness constraints would not diagnose the harm (Quaresmini et al., 24 Apr 2026). SEI governance therefore includes not only transparency and provenance, but also longitudinal auditing of epistemic standing and interpretative access.

6. Misconceptions, tensions, and open questions

A persistent misconception is that SEI are simply better retrieval systems or more capable tools. Several strands of the literature reject this. Chen argues that generative AI in education should be understood not as isolated utilities but as epistemic infrastructure that mediates teaching and learning (Chen, 9 Apr 2025). OIDA argues that the ceiling on organizational AI is epistemic fidelity, not retrieval fidelity, because semantically relevant content may still fail to distinguish decisions from abandoned hypotheses or questions from settled facts (Bottino et al., 13 Apr 2026). The post-coherence framework likewise shifts analysis from classification to coordination across institutional, computational, and temporal arrangements (Kelly, 7 Aug 2025).

A second tension concerns reflexivity and burden. Arzberger et al. report that reflexive annotating produces richer situated metadata, but also emotional exposure, strategic withdrawal, and discomfort with revealing intimate identity facets (Arzberger et al., 25 Jan 2026). This indicates that SEI designed to preserve standpoint and provenance may also intensify affective and cognitive labor. The tension is not incidental; it is part of the infrastructural politics of whose knowledge is solicited, how, and at what personal cost.

A third issue is the difficulty of empirical adjudication. OIDA’s strongest organizational result is explicitly limited by token-budget confounding, with the decisive ablation not yet run (Bottino et al., 13 Apr 2026). TRACE demonstrates real-time common-ground tracking but with modest end-to-end accuracy trade-offs relative to offline conditions (VanderHoeven et al., 12 Mar 2025). Riva’s broader claims about adaptive invisibility and preconscious reshaping are methodologically challenging precisely because infrastructural mediation is designed to remain below awareness (Riva, 19 Jun 2025). These are not refutations of SEI, but they constrain what can presently be established.

Open directions are already well specified in the literature. Wittenborg identifies expansion of FAIR ground-truth knowledge graphs, automated provenance capture, dynamic governance models for federated SEIs, incentive mechanisms for community curation, and cross-domain SEI mashups as future work (Wittenborg, 6 Mar 2026). Quaresmini et al. call for intersectional and multi-group extensions, adaptive SEI in which real-time indices trigger corrective interventions, and validation of proxy choices against lived experiences of epistemic injustice (Quaresmini et al., 24 Apr 2026). Vogt’s framework implies a further shift toward post-FAIR infrastructures in which knowledge graphs become native decision-support systems through conditional, auditable Action Units (Vogt, 2 May 2026).

Across these formulations, SEI mark a transition in the analysis of knowledge systems: from repositories to infrastructures, from coherence to coordination, from retrieval to epistemic fidelity, and from passive representation to situated, auditable mediation of judgment and action.