Principled Abstention and Semantic Governance: Architecting Uncertainty in Embedded Knowledge Systems

Full report

On this page

1. Executive Summary

The prevailing paradigm in applied semantic resolution and embedded vector search relies heavily on forced nearest-neighbor matching. In these systems, every input expression—regardless of its ambiguity, novelty, or linguistic origin—is mathematically compelled to resolve to the nearest available vector in the latent space. While this approach maximizes query-response coverage, it fundamentally compromises semantic integrity. The architecture proposed in this report operates on a singular, uncompromising core principle: a semantic resolver should be able to say, "I do not have enough evidence to assign a stable Concept," rather than forcing every expression into the nearest vector neighbor. This report provides an exhaustive investigation into the operationalization of principled abstention within embedded semantics. By integrating selective classification, open-set recognition (OSR), and conformal prediction, semantic systems can formally quantify uncertainty and reject out-of-distribution or ambiguous inputs. Furthermore, semantic uncertainty is not exclusively algorithmic; it is deeply intertwined with ontology governance, human-in-the-loop (HITL) triage, cultural context, and adversarial threat models. The analysis synthesizes precedents from the Unicode Consortium, Wikidata, SNOMED CT, and established ISO standards to propose a comprehensive framework for semantic authority. It delineates the boundaries of who can propose, approve, deprecate, or merge concepts, and establishes the cryptographic and operational audit trails required to maintain trust in safety-critical domains. The resulting architecture ensures that semantic systems prioritize accuracy and provenance over forced, brittle resolution, ultimately safeguarding against semantic drift, data poisoning, and linguistic bias. By avoiding claims of perfect semantic certainty, knowledge graphs can evolve from fragile taxonomies into resilient, multi-perspectival semantic registries.

2. Why Semantic Systems Need Abstention

The core vulnerability of modern dense retrieval and embedding-based classification is the phenomenon of softmax overconfidence and metric space contiguity1. In a standard closed-set environment, models partition the entire vector space among known classes. Consequently, when an out-of-distribution (OOD) sample, a novel concept, or a highly ambiguous expression is projected into this space, the system assigns it to the nearest centroid, often with artificially high confidence3. This forced resolution creates a cascade of silent failures. First, it causes semantic conflation, where distinct but previously unmodeled concepts are merged into existing nodes, destroying the granularity of the knowledge graph5. Second, it eliminates epistemic humility; the system cannot distinguish between a highly confident match and a "best available but fundamentally incorrect" match1. Finally, this structural rigidity creates a vulnerability to adversarial perturbation, wherein attackers can exploit forced resolution by crafting inputs that mathematically align with a target vector despite lacking semantic equivalence8. Abstention—the capacity to reject an input or defer it to human review—is a fundamental requirement for reliable artificial intelligence. Implementing a reject option transforms the semantic space from a closed universe into an open-world model10. By defining rejection thresholds, semantic systems can trade raw coverage for guaranteed risk control, ensuring that when a concept is resolved, it possesses a mathematically bounded guarantee of accuracy12. Without this abstention layer, knowledge graphs inevitably drift toward hallucination and structural degradation.

3. Types of Uncertainty

To architect a system capable of abstention, it is necessary to categorize the distinct sources of uncertainty that trigger a rejection. Uncertainty in embedded semantics is not monolithic; it spans algorithmic, ontological, and sociolinguistic dimensions.

These four pillars of uncertainty demand specialized algorithmic treatments. Treating epistemic uncertainty with the tools meant for aleatoric uncertainty leads to over-engineering, while ignoring ontological uncertainty guarantees the eventual obsolescence of the registry.

4. Model Uncertainty

Quantifying model uncertainty requires moving beyond standard uncalibrated outputs to employ rigorous statistical frameworks that bound open-space risk. Contemporary research identifies several primary mechanisms for embedding-based uncertainty quantification.

4.1 Selective Classification

Selective classification, frequently referred to as reject-option classification, allows a model to abstain from prediction when its confidence falls below a dynamically calibrated threshold1. The primary objective is to optimize the tradeoff between coverage (the fraction of inputs the system attempts to resolve) and risk (the error rate of the resolved inputs)12. Recent theoretical approaches frame optimal selection functions through the lens of the Neyman–Pearson lemma, treating rejection as a likelihood ratio test23. For deep neural networks, selective classification can be applied via post-hoc scoring functions or by training models with an integrated reject option, such as SelectiveNet, which explicitly trains the network to optimize the coverage-risk curve10.

4.2 Open-Set Recognition (OSR) and OpenMax

Traditional classifiers operate under a closed-set assumption, which is fatal in open-ended semantic environments. Open-Set Recognition (OSR) addresses the scenario where unseen classes appear at inference time14. The foundational algorithmic approach to OSR in deep learning is the OpenMax framework11. OpenMax rejects the assumption that probabilities must sum to one over known classes. Instead, it utilizes Extreme Value Theory (EVT) to model the distribution of activation vectors in the penultimate layer of a neural network29. For each known class, OpenMax calculates a Mean Activation Vector (MAV) and fits a Weibull distribution to the distances of the training samples from this central point4. During inference, if an input vector's distance to the nearest MAV falls into the extreme tail of the Weibull distribution, the system mathematically revises the activation scores and explicitly allocates probability mass to an "unknown" class32. This process mathematically bounds open-space risk and prevents the overconfident assignment of novel concepts to existing vectors4.

4.3 Conformal Prediction

Conformal prediction provides a distribution-free paradigm for uncertainty quantification that guarantees non-asymptotic coverage without requiring strict distributional assumptions about the underlying data7. Instead of outputting a single point-estimate concept, a conformal predictor outputs a prediction set that is mathematically guaranteed to contain the true concept with a user-specified probability (e.g., a 95% confidence bound)36. In embedded semantics, Selective Conformal Risk Control (SCRC) unifies selective classification and conformal prediction to create a highly efficient uncertainty architecture12. SCRC operates as a two-stage procedure. The first stage, selection control, determines whether to accept the sample based on a pre-defined confidence function37. The second stage, risk control, constructs calibrated conformal sets for the accepted samples12. Methodologies like SCRC-T (transductive) compute thresholds symmetrically over calibration and test data to maintain strict exchangeability guarantees, while SCRC-I (inductive) reuses calibration thresholds to maximize computational efficiency12. By adjusting these conformal thresholds over calibration data, semantic systems can ensure rigorous error bounds even under severe distribution shifts39.

4.4 Mahalanobis Distance for OOD Detection

Standard geometric measurements like Euclidean distance and cosine similarity frequently fail to capture the complex information geometry of latent feature spaces. The Mahalanobis distance improves out-of-distribution detection by explicitly incorporating the covariance matrix of the training distribution41. By measuring the distance of a test sample from a class centroid relative to the variance in that specific direction, Mahalanobis-based detectors effectively identify inputs that reside in low-density regions of the latent space, triggering a principled abstention43. Advanced implementations, such as Mahalanobis Residual Scoring (MaRS) and dynamic covariance calibration, refine this by dynamically updating the prior covariance matrix using real-time input features, preserving essential data characteristics while heavily penalizing anomalous vectors44.

5. Registry Uncertainty

Registry uncertainty addresses the structural integrity of the knowledge graph itself. An algorithmic model may successfully map an expression to a concept, but if the concept's definition has degraded over time, the resolution remains invalid.

5.1 Concept Shift and Semantic Drift

Semantic drift occurs when the attributes, boundaries, or real-world referents of a concept change organically over time17. In medical terminologies, a disease definition might shift due to new diagnostic criteria16. In embedding spaces, this phenomenon manifests as a growing geometric divergence between historical embeddings and contemporary contextual usage5. To detect and mitigate semantic drift, modern registries must employ topological data analysis and persistent homology to track the evolution of embedding spaces over conversational time48. Drift is quantified using distance metrics—such as Kullback-Leibler (KL) divergence and k-nearest neighbor (k-NN) overlap—applied to chronological text corpora49. When the calculated semantic drift of a cluster exceeds a defined structural threshold, the resolver must abstain from mapping new expressions to the historical concept. The concept must be locked until human governance intervenes to split, deprecate, or supersede the node6.

5.2 Granularity and Null-Valued Objects

Registries frequently suffer from varying levels of granularity. An input expression might be highly specific, while the registry only contains a broad, generalized parent category. Resolving the specific expression to the broad category introduces an unacceptable loss of precision and risks downstream logical errors. The semantic system must abstain from a direct exact match and instead output a relationship indicating approximation, or create a null-valued candidate object pending human review51. The Simple Knowledge Organization System (SKOS) standard accommodates this through the utilization of skos:broadMatch or skos:closeMatch, ensuring that the lack of exact granularity is explicitly communicated to the querying client16.

6. Translation Uncertainty

Multilingual semantic resolution introduces severe mapping challenges. Embedded semantic systems often utilize cross-lingual vector spaces to align concepts across languages, but these spaces are prone to structural and linguistic biases.

6.1 Cross-Lingual Alignment Failures

When aligning high-resource languages with low-resource languages, severe geometric distortions occur in the latent space. The primary vulnerability is the hubness problem, a phenomenon in high-dimensional spaces where a small number of vectors (hubs) frequently appear as the k-nearest neighbors for a vastly disproportionate number of queries53. When a system attempts to resolve a low-resource expression, it is frequently and erroneously mapped to a high-resource hub vector rather than its true semantic equivalent55. To prevent this, resolvers must apply cross-domain regularization and local ranking alignment to penalize hub formations56. If the local neighborhood density indicates a hub collision, the system must recognize the geometric instability and abstain from the translation.

6.2 Lexical Gaps and Translation Equivalence

Semantic resolution across languages is rarely a one-to-one mapping. Concepts deeply embedded in one language's cultural matrix may have no direct equivalent in another. In formal terminology systems, forcing a translation creates dangerous false equivalencies. The system must utilize ontologies like SKOS to explicitly label relationships16. If an exact equivalent does not exist, the resolver must state that a stable, exact concept cannot be assigned, opting instead to return an approximate bounding concept or deferring to a human localization team.

7. Cultural Uncertainty

Semantic registries are not objective mirrors of a singular reality; they encode the cultural, political, and historical biases of their creators. Handling culturally contested concepts requires a departure from absolute taxonomy toward pluralistic modeling.

7.1 Essentially Contested Concepts

Philosopher W.B. Gallie introduced the theory of "essentially contested concepts"—ideas whose proper use inevitably involves endless, unresolvable disputes (examples include "justice", "human dignity", and "democracy")57. Disambiguation cannot resolve these conflicts because the disagreement is not lexical; it is deeply ontological and normative59. A semantic resolver must not force essentially contested concepts into a single consensus vector, as doing so enforces a hegemonic viewpoint. Instead, the architecture must support multi-perspectival modeling. When an expression maps to a contested concept, the system should abstain from declaring a singular truth. The appropriate response is to return a cluster of context-dependent definitions, preserving the epistemological flexibility of the knowledge network and allowing the downstream application to select the appropriate localized context61.

7.2 Cultural Disagreement and Deprecation Precedents

The Unicode Consortium provides a primary case study in cultural uncertainty and governance fatigue. Historically, the Unicode Emoji Subcommittee accepted proposals for geographic, cultural, and identity-based flags. However, deciding which regions, identities, or geopolitical movements warranted a flag forced the Consortium into acting as an arbiter of global geopolitical legitimacy63. In response, the Consortium enacted a sweeping deprecation policy, refusing to accept new flag emoji proposals. They cited the transient nature of these symbols and the inherent exclusivity they create—arguing that the inclusion of one identity inherently invites demands for all others, which is unscalable65. This precedent demonstrates that when a semantic domain becomes overwhelmingly contested or politically toxic, the most responsible governance action is structural abstention: freezing the ontology branch and refusing to encode further concepts in that specific category.

8. Conflicting Evidence

In distributed knowledge graphs and federated semantic systems, a resolver will frequently encounter conflicting claims regarding the properties, definitions, or relationships of a specific concept.

8.1 Reification and RDF-star

Traditional Resource Description Framework (RDF) triples, formatted as Subject \-\> Predicate \-\> Object, are fundamentally insufficient for modeling conflicting evidence because they assert facts as absolute, context-free truths68. To handle conflicting evidence responsibly, registries must utilize RDF-star (RDF 1.2), which introduces quoted triples and the concept of reification70. RDF-star allows the system to make statements about statements. For example, if Source A and Source B provide conflicting medical classifications for a single symptom, the system does not overwrite one with the other. Instead, both are preserved as quoted triples, appended with rich metadata regarding their source, timestamp, and calculated confidence score72. This allows the resolver to return both possibilities, heavily weighted by the querying application's trust parameters, rather than collapsing the data into a false consensus.

8.2 Provenance Modeling with PROV-O

To resolve conflicting evidence safely, the system must trace the exact lineage of the data. The W3C PROV-O ontology provides a standardized framework for this requirement, utilizing core classes (prov:Entity, prov:Activity, prov:Agent) and connective properties (prov:wasDerivedFrom, prov:wasAttributedTo, prov:wasGeneratedBy)74. When the resolver identifies conflicting assertions, it evaluates the prov:Agent and prov:wasDerivedFrom pathways. If the provenance data reveals equal authoritative weight for conflicting claims, the algorithmic resolver is strictly bound to abstain and escalate the conflict to the human governance body for manual review77.

9. Human Review

When a semantic system mathematically abstains, it must seamlessly interface with a Human-in-the-Loop (HITL) triage system to ensure that the unmapped concept is eventually resolved or formally deprecated.

9.1 Active Learning and Query-by-Committee

Routing every abstained query to human reviewers is economically unscalable and induces severe reviewer fatigue79. Systems must employ active learning strategies to prioritize inputs that maximize expected model change.

• Uncertainty Sampling: The system prioritizes routing samples where the model's conformal prediction sets are excessively large, indicating high boundary confusion and maximum potential for algorithmic learning upon resolution80.
• Query-by-Committee (QBC): The system deploys a diverse ensemble of embedding models. It routes expressions to humans only when there is maximum disagreement (high statistical variance) among the committee members regarding the nearest centroid81.

9.2 Asymmetric Cost Matrices

Deployment of human review requires cost-controlled abstention frameworks83. By integrating Mondrian conformal prediction with an asymmetric cost matrix, systems parameterize the exact cost of a false positive (a forced incorrect resolution), a false negative (a missed resolution), and the operational cost of human review83. The system algorithmically determines the break-even threshold: abstaining and deferring to a human only when the expected cost of an automated misclassification exceeds the economic cost of routing the query to a reviewer84.

10. Governance Models

The rules governing how concepts are added, modified, or deprecated are as critical as the algorithms resolving them. A review of established governance models highlights diverse approaches to maintaining semantic authority.

An embedded semantic system should adopt a hybrid approach: leveraging the structured lifecycle management of ISO 11179, the immutable versioning and safety protocols of SNOMED CT, and the transparent proposal mechanisms of Wikidata.

11. Registry Roles

Authority boundaries must be explicitly defined through Role-Based Access Control (RBAC) integrated directly into the registry's cryptographic graph layer100.

This separation of powers ensures that automated agents and general users can enrich the vocabulary without possessing the destructive permissions necessary to compromise the core ontology.

12. Proposal and Approval Workflows

The lifecycle of a concept modification must follow a strict cryptographic and procedural workflow to maintain trust and prevent chaotic database expansions.

12.1 The Proposal Phase

Drawing heavily from Wikidata's "Property Proposal" system, every new concept or structural change must be submitted as a structured dossier85. This dossier must mandate:

• The proposed label, an exhaustive description, and all known aliases.
• The proposed hierarchical position utilizing SKOS relationships (e.g., skos:broader, skos:related).
• Evidence of real-world usage or scientific consensus in the form of verifiable provenance URLs.
• Impact analysis: An automated query demonstrating exactly how this new concept affects existing embedding clusters and nearest-neighbor calculations.

12.2 Splitting and Merging

Semantic concepts frequently require restructuring as human knowledge evolves103.

• Splitting: When an overloaded concept is divided into narrower concepts, the original concept must not be deleted. It is marked as deprecated or ambiguous, and formal relationships are drawn to the new, narrower concepts to preserve legacy data links.
• Merging: When two concepts are deemed identical, one is retained as the primary node, and the other is inactivated. The inactivated concept must be entirely cleared of conflicting sub-data and redirected using a definitive relationship (e.g., SAME AS)104.

12.3 Automated Sanity Checks

Before human review occurs, the system must perform automated topological checks. If a proposed change introduces cycles in the ontology (e.g., A is broader than B, B is broader than C, C is broader than A), or creates massive semantic drift among its child nodes, the proposal is algorithmically blocked and returned to the Submitter106.

13. Versioning

A core tenet of semantic governance is that knowledge is immutable over time; only our understanding of it changes. Following the critical precedents of medical ontologies like SNOMED CT and LOINC, concepts must never be physically deleted from the database108. Deletion destroys the auditability of historical records, making it impossible to understand why a clinical or financial decision was made three years prior. Instead, concepts undergo state transitions. When a concept is inactivated, it is appended with a Component Inactivation Reference Set (CIRS) detailing the precise reason for inactivation (e.g., "Outdated", "Ambiguous", "Duplicate")95. The system must mandate the use of Historical Association Reference Sets (HARS) to map inactive concepts to active ones94.

• REPLACED BY: Provides a 1:1 mapping to a newly authored concept95.
• SAME AS: Resolves duplication by pointing to the surviving node113.
• MOVED TO: Indicates structural namespace changes95.
• POSSIBLY EQUIVALENT TO: Used when a highly ambiguous concept is inactivated, pointing to potential narrower replacements112.

From a database engineering perspective, this requires a temporal graph architecture where every transaction is an immutable, cryptographically signed commit. Updates create new commits rather than destructive writes, allowing full, flawless reconstruction of the ontology at any prior timestamp100.

14. Appeals/Disputes

When Domain Reviewers disagree on the validity of a concept, or when external users contest a classification, a formal, transparent dispute resolution mechanism is strictly necessary.

• Public Discourse: Following the Wikidata model, disputes must occur in public, threaded forums appended directly to the concept in question (commonly known as Requests for Comment, or RfC)115.
• Escalation to ArbCom: Intractable disputes are escalated to an Arbitration Committee (ArbCom)87. The ArbCom does not rule on absolute semantic truth, but rather on whether procedural guidelines (e.g., sourcing, neutral point of view, lack of conflict of interest) were rigorously followed during the proposal phase118.
• State During Dispute: While a concept is under active dispute, its status is temporarily flagged as contested. The embedded resolver will penalize the confidence score of any input matching this concept, forcing the system into an abstention state until the dispute is resolved51.

15. Safety-Critical Concepts

In domains such as medicine, aviation, or legal compliance, incorrect semantic resolution carries severe liability and direct safety risks7. A medical system utilizing SNOMED CT or LOINC cannot afford a "close enough" nearest-neighbor match; misclassifying a medication dosage unit or a highly specific diagnostic code can be fatal120. For safety-critical branches of the ontology, the system must enforce draconian resolution protocols:

1. Zero-Tolerance Abstention: The conformal prediction risk target must be set exceptionally close to zero7.

2. Prohibition of Approximate Matching: The resolver is forbidden from utilizing cross-lingual hubs or relaxed vector thresholds. It must require an exact lexical match or a verified, human-approved synonymous match122.

3. Strict Provenance Constraints: Queries resolving to safety-critical concepts must validate the prov:wasDerivedFrom chain in real-time. If the concept relies on unverified external assertions or weak evidence, the system must abstain77.

16. Abuse Threat Model

As semantic systems become deeply embedded in search infrastructure and Retrieval-Augmented Generation (RAG) pipelines, they become high-value targets for manipulation. The threat model encompasses both sociotechnical abuse and algorithmic attacks.

16.1 Malicious Registry Edits

Attackers may attempt to alter ontology metadata to manipulate downstream applications. This includes subtly changing the definition of a political entity or altering the hierarchy of a pharmaceutical product to influence SEO or generative AI outputs88. Mitigation requires strict RBAC, mandatory multi-signature approvals for critical nodes, and automated semantic drift detection that flags massive, sudden structural deviations before they are committed107.

16.2 Adversarial Expressions and Evasion

Attackers can craft inputs specifically designed to bypass moderation filters or force a specific, malicious resolution124. By adding imperceptible textual perturbations or utilizing Unicode character substitutions (e.g., homoglyphs), attackers can shift the input vector just enough to evade detection while maintaining human readability56. Systems must sanitize inputs, strip zero-width characters, and apply robust dense retrieval alignment techniques to neutralize these perturbations before they ever enter the vector space126.

17. Poisoning

Data poisoning is arguably the most critical vulnerability in vector databases and dense retrieval systems. In a corpus poisoning attack, an adversary injects maliciously crafted documents or concepts into the knowledge base128. Because ingestion pipelines often lack semantic verification, these adversarial vectors are geometrically identical to legitimate entries130. When a user submits a query, the injected vector acts as an artificial nearest-neighbor, hijacking the retrieval process (seen in methodologies like HotFlip attacks or targeted clean-label poisoning)9. Standard Euclidean thresholds fail entirely because the poisoned vector is explicitly designed to sit exactly on the targeted query path131. Defenses must employ Mahalanobis Residual Scoring (MaRS) and dynamic covariance calibration to detect structural anomalies in the embedding distribution at ingestion time42. Any ingestion candidate that distorts the local information geometry beyond a statistical threshold must be quarantined44.

18. Model Bias

Embedded semantics inexorably encode the biases present in their training corpora. This leads to the over-representation of dominant cultural norms and the systemic marginalization of minority contexts. A concept like "marriage" or "democracy" will inherently embed the legal and cultural frameworks of the Anglosphere if the training data is predominantly English62. To counter model bias, the governance framework must decouple the underlying concept from its lexical representation. The semantic graph must allow for diverse, region-specific representations of the same universal identifier without enforcing a singular normative definition. Furthermore, continuous auditing must be performed to monitor whether the selective classification threshold disproportionately forces abstentions on demographics, dialects, or vernaculars that fall outside the dominant training distribution83.

19. Low-Resource Languages

Low-resource languages suffer from compounding systemic bias in semantic resolution53. Because training data is exceedingly sparse, the embeddings for low-resource languages are poorly clustered, leading to severe alignment failures in cross-lingual vector spaces. When applying zero-shot entity linking, the lack of robust synonym coverage for low-resource entities forces the resolver to guess, often mapping the query to incorrect but high-density hubs54. To address this, the system must track the linguistic provenance of every input. If the input is detected as a low-resource language and the vector match relies heavily on cross-lingual projection, the system must apply a significantly stricter Mahalanobis distance threshold. If it fails to meet this threshold, the system must abstain rather than execute a flawed, hallucinated zero-shot link.

20. Audit Trails

Without comprehensive auditability, semantic abstention and resolution are indistinguishable from uninterpretable black-box magic. A high-trust semantic system must maintain an unbroken, verifiable chain of custody for every concept and every resolution decision ever made. Utilizing architectures akin to FlureeDB, the registry should operate as a cryptographically signed graph100. Every node addition, split, merge, or deprecation must be recorded as a cryptographically signed transaction101. The W3C PROV-O standard must be strictly enforced. A concept must carry metadata defining wasAttributedTo (the user or agent who proposed it), wasGeneratedBy (the specific model version that created the embedding), and invalidatedAtTime (for inactive concepts)76. Unlike traditional databases where data is hard-deleted, tracking invalidations via tombstones ensures that past queries can be re-run against historical states of the graph, providing irrefutable, cryptographically sound proof of why a system resolved (or abstained) on a specific date77.

21. Recommended Governance Principles

To maintain absolute integrity, the semantic registry should adhere to the following core governance principles:

1. Abstention over Assumption: The system shall always prioritize abstention over low-confidence resolution. Supplying no concept is vastly preferable to supplying a false concept that poisons downstream logic.

2. Immutable History: Concepts and relationships are never physically deleted. They transition through strict status lifecycles and are retained cryptographically forever to preserve auditability.

3. Explicit Lineage: Every assertion, translation, and rendering must be traceable to a specific agent, model, or source via formal provenance ontologies (PROV-O and RDF-star).

4. Decentralized Curation, Centralized Enforcement: Proposals should be open to the broad community or automated agents to foster growth, but structural alterations (merges and splits) must be strictly governed by authorized domain experts.

5. Multi-Perspectival Reality: The ontology must allow conflicting, heavily sourced claims to coexist without forcing a premature consensus, particularly for culturally contested concepts.

22. Recommended Confidence Language

To transparently communicate the state of a semantic resolution, the site, user interface, and API must utilize a standardized, rigorously defined vocabulary. The platform must avoid all claims of "perfect certainty."

23. Site Transparency Pages

Public trust demands that the mechanics and shortcomings of the semantic system are highly visible. The platform must host dynamic transparency pages detailing:

• System Performance Metrics: Real-time dashboards (e.g., Grafana) displaying query volume, abstention rates, and measured error rates across domains136.
• Model Versioning: Explicit logs of which embedding models are currently active, their training cutoff dates, and their thoroughly documented biases.
• Rejection Thresholds: Public documentation of the current conformal prediction risk targets and Mahalanobis distance thresholds.
• Governance Action Log: A live, immutable feed of recent concept merges, splits, and deprecations, mirroring Wikipedia's recent changes functionality.

24. Public Documentation

External developers and system integrators require crystal clear rules of engagement. Public documentation must include:

• API Contracts: Clear definitions of how the system outputs Abstained or Ambiguous responses via HTTP status codes and strict JSON schemas137.
• Querying Historical States: Comprehensive instructions on how to utilize Expression Constraint Language (ECL) or SPARQL-star to query the ontology at previous points in time72.
• Adversarial Guidelines: Explicit warnings regarding the system's defenses against HotFlip and data poisoning, clarifying that adversarial inputs will result in immediate API rate-limiting or permanent bans.

25. FAQs

Q: Why did my query return "Abstained" instead of the closest match? A: Our system utilizes strict conformal risk control. If the mathematical probability of an incorrect match exceeds our safety threshold, the system abstains. We do not force inputs into the nearest vector neighbor if the confidence is too low, as this corrupts downstream data integrity. Q: Can I retrieve concepts that were deleted last year? A: Concepts are never deleted. They are marked as Deprecated or Superseded. You can query historical states using our temporal graph API to retrieve the exact concept state as it existed at any prior timestamp, ensuring full auditability. Q: How do you handle words that have opposite meanings in different contexts? A: We do not collapse conflicting meanings into a single node. If context is missing, the system will return an Ambiguous response with multiple candidate paths. If sufficient context is provided, the system traverses the appropriate sub-graph to return the exact match. Q: Who decides when a concept is superseded? A: Structural changes are governed by appointed Domain Reviewers and the central Governance Board. However, any user can propose a change or flag a concept for review using our transparent, public proposal workflow.

26. Diagrams

26.1 Proposal and Resolution Lifecycle

26.2 System Architecture Flow

27. Open Questions

While this framework establishes robust, scientifically grounded boundaries for semantic systems, several critical research frontiers remain unresolved:

• Dynamic Conformal Thresholding in Real-Time: Can systems dynamically adjust conformal risk targets on a per-query basis depending on the detected linguistic origin (e.g., automatically tightening thresholds to prevent hallucinations in low-resource languages)?
• LLM Hallucination vs. Semantic Drift: As LLMs are increasingly used to generate candidate relationships in automated knowledge graph construction, how can systems mathematically distinguish between an LLM hallucinating a non-existent connection versus accurately identifying a legitimate, emerging semantic drift?
• Scaling RDF-star: While RDF-star is mathematically elegant for handling conflicting claims via reification and quoted triples, the immense computational overhead of querying millions of quoted triples in real-time requires significant advancements in graph database optimization.

28. Annotated Sources

The foundational methodologies supporting this architecture draw heavily upon several distinct domains of computer science, linguistics, and database management. To operationalize uncertainty, the framework relies on the foundational literature of Selective Classification and Reject Options, which establishes the absolute necessity of the reject option in deep neural networks to trade coverage for risk, viewing optimal rejection through the Neyman-Pearson lemma1. This is heavily augmented by Open-Set Recognition (OSR) and OpenMax research, which defines the severe limits of closed-set Softmax and introduces Extreme Value Theory (EVT) and Weibull distributions to explicitly model the "unknown" and bound open-space risk4. Furthermore, the statistical backbone of the system's abstention mechanics is derived from Conformal Prediction literature, which establishes distribution-free uncertainty quantification, allowing systems to output prediction sets with finite-sample statistical guarantees—most notably via Selective Conformal Risk Control (SCRC)7. Mahalanobis Distance research further dictates how out-of-distribution detection must utilize covariance matrices and information geometry to identify outliers before they corrupt the resolver41. In addressing the temporal decay of knowledge, Semantic Drift studies provide the mathematical basis for tracking the evolution of concepts over time in embedding spaces, utilizing topological data analysis and similarity metrics to detect divergence5. The threat models dictating the architecture's security posture are derived from Poisoning & Adversarial Attacks research, which demonstrates severe vulnerabilities in dense retrieval and vector indexing (such as HotFlip and backdoor attacks) and establishes robust geometric defense mechanisms8. Sociolinguistic handling within the graph is informed by Low-Resource Languages & Hubness research, detailing geometric distortions in cross-lingual spaces and zero-shot alignment failures53. Additionally, the philosophical handling of unresolvable terms is rooted in W.B. Gallie’s framework of Essentially Contested Concepts, proving that disambiguation cannot resolve normative semantic disputes57, while Semantic Uncertainty methodologies dictate how to calculate entropy over meaning rather than exact lexical sequences in generative retrieval15. Finally, the procedural and structural governance mechanisms are synthesized from real-world precedents. Active Learning & HITL strategies dictate how to optimize the economic cost of routing abstentions to human reviewers79. SNOMED CT Governance provides the structural precedents for medical ontologies, emphasizing non-deletion, Component Inactivation, Historical Association Reference Sets, and ECL querying93, underscoring the necessity of strict code stability in Safety-Critical Terminology109. Wikidata Governance provides insights into community-driven ontology management, handling of empty merged items, and dispute resolution via ArbCom and RfCs51. The Unicode Consortium reveals the structural constraints of the Character Encoding Stability Policy and the geopolitical fallout resulting in the deprecation of flag emojis63. The underlying data models themselves are governed by ISO 11179, ISO 1087 & SKOS, defining terminology work and W3C ontology structures16, while PROV-O, RDF-star & Cryptographic Graphs ensure immutable data provenance, reification, quoted triples, and cryptographically signed knowledge graphs68.

Works cited

1. Out-of-Distribution Detection: A Task-Oriented Survey of Recent Advances \- arXiv, https://arxiv.org/html/2409.11884v4

2. Deep Neural Network Benchmarks for Selective Classification \- arXiv, https://arxiv.org/html/2401.12708v2

3. Post-hoc Selective Classification for Reliable Synthetic Image Detection \- arXiv, https://arxiv.org/html/2605.08574v1

4. arXiv:2207.10287v1 \[cs.CV\] 21 Jul 2022, https://arxiv.org/pdf/2207.10287

5. Survey in Characterization of Semantic Change \- arXiv, https://arxiv.org/html/2402.19088v3

6. 1 Introduction \- arXiv, https://arxiv.org/html/2511.04926v1

7. A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification \- arXiv, https://arxiv.org/html/2107.07511v6

8. Be Careful about Poisoned Word Embeddings: Exploring the Vulnerability of the Embedding Layers in NLP Models | Request PDF \- ResearchGate, https://www.researchgate.net/publication/352365171\_Be\_Careful\_about\_Poisoned\_Word\_Embeddings\_Exploring\_the\_Vulnerability\_of\_the\_Embedding\_Layers\_in\_NLP\_Models

9. Towards Secure Retrieval-Augmented Generation: A Comprehensive Review of Threats, Defenses and Benchmarks \- arXiv, https://arxiv.org/html/2603.21654v1

10. arXiv:2209.04944v2 \[cs.CV\] 15 Feb 2023, https://arxiv.org/pdf/2209.04944

11. MetaMax: Improved Open-Set Deep Neural Networks via Weibull Calibration \- arXiv, https://arxiv.org/html/2211.10872v2

12. Selective Conformal Risk Control \- arXiv, https://arxiv.org/pdf/2512.12844

13. Conformal Prediction as Bayesian Quadrature \- arXiv, https://arxiv.org/html/2502.13228v2

14. Category Discovery: An Open-World Perspective \- arXiv, https://arxiv.org/html/2509.22542v1

15. SUGAR: Leveraging Contextual Confidence for Smarter Retrieval \- arXiv, https://arxiv.org/html/2501.04899v1

16. TBX and 'Lemon': What perspectives in terminology? \- Oxford Academic, https://academic.oup.com/dsh/article/38/Supplement\_1/i61/7203991

17. Monitoring Term Drift Based on Semantic Consistency in an Evolving Vector Field \- arXiv, https://arxiv.org/pdf/1502.01753

18. Achieving Semantic Consistency: Contextualized Word Representations for Political Text Analysis \- arXiv, https://arxiv.org/pdf/2412.04505

19. \[2501.04899\] SUGAR: Leveraging Contextual Confidence for Smarter Retrieval \- arXiv, https://arxiv.org/abs/2501.04899

20. LLMs Should Express Uncertainty Explicitly \- arXiv, https://arxiv.org/html/2604.05306v2

21. The Anatomy of Uncertainty in LLMs \- arXiv, https://arxiv.org/html/2603.24967v2

22. LLMs Uncertainty Quantification via Adaptive Conformal Semantic Entropy \- arXiv, https://arxiv.org/html/2605.04295v1

23. Know When to Abstain: Optimal Selective Classification with Likelihood Ratios \- arXiv, https://arxiv.org/html/2505.15008v1

24. Know When to Abstain: Optimal Selective Classification with Likelihood Ratios \- arXiv, https://arxiv.org/html/2505.15008v2

25. SelectiveNet: A Deep Neural Network with an Integrated Reject Option \- arXiv, https://arxiv.org/pdf/1901.09192

26. Selective Classification via One-Sided Prediction, https://arxiv.org/pdf/2010.07853

27. OSSAR: Towards Open-Set Surgical Activity Recognition in Robot-assisted Surgery \- arXiv, https://arxiv.org/pdf/2402.06985

28. Learning Placeholders for Open-Set Recognition \- arXiv, https://arxiv.org/html/2103.15086v1

29. Generative OpenMax for Multi-Class Open Set Classification \- arXiv, https://arxiv.org/pdf/1707.07418

30. Performance Analysis of Out-of-Distribution Detection on Various Trained Neural Networks \- arXiv, https://arxiv.org/pdf/2103.15580

31. COSTARR: Consolidated Open Set Technique with Attenuation for Robust Recognition, https://arxiv.org/html/2508.01087v1

32. Multi-Domain Supervised Contrastive Learning for UAV Radio-Frequency Open-Set Recognition \- arXiv, https://arxiv.org/html/2508.12689v2

33. Generalized Out-of-Distribution Detection: A Survey \- arXiv, https://arxiv.org/html/2110.11334v1

34. VAEMax: Open-Set Intrusion Detection based on OpenMax and Variational Autoencoder \- arXiv, https://arxiv.org/pdf/2403.04193

35. \[2107.07511\] A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification \- arXiv, https://arxiv.org/abs/2107.07511

36. Distribution-Free Uncertainty Quantification A short introduction to Conformal Prediction \- CP tutorial ·, https://conformalpredictionintro.github.io/assets/files/cptuto.pdf

37. Selective Conformal Risk Control \- arXiv, https://arxiv.org/html/2512.12844v1

38. (PDF) Selective Conformal Risk Control \- ResearchGate, https://www.researchgate.net/publication/398721023\_Selective\_Conformal\_Risk\_Control

39. arXiv:2205.14310v3 \[cs.LG\] 1 Dec 2022, https://arxiv.org/pdf/2205.14310

40. Calibrated Selective Classification \- arXiv, https://arxiv.org/html/2208.12084v2

41. Unsupervised Out-of-Distribution Dialect Detection with Mahalanobis Distance \- arXiv, https://arxiv.org/html/2308.04886v1

42. Improving Out-of-Distribution Detection via Dynamic Covariance Calibration \- arXiv, https://arxiv.org/pdf/2506.09399

43. arXiv:2305.13849v3 \[cs.CV\] 29 Sep 2023, https://arxiv.org/pdf/2305.13849

44. Improving Out-of-Distribution Detection via Dynamic Covariance Calibration \- arXiv, https://arxiv.org/html/2506.09399v1

45. MaRS: Robust Out-of-Distribution Detection via Mahalanobis Residual Scoring \- arXiv, https://arxiv.org/html/2606.22649v1

46. A Geometry-Based View of Mahalanobis OOD Detection \- arXiv, https://arxiv.org/html/2510.15202v3

47. MLHOps: Machine Learning for Healthcare Operations \- arXiv, https://arxiv.org/pdf/2305.02474

48. DHoTT: A Temporal Extension of Homotopy Type Theory for ... \- arXiv, https://arxiv.org/html/2506.09671

49. Evaluating the Stability of Embedding-based Word Similarities | Request PDF, https://www.researchgate.net/publication/329554345\_Evaluating\_the\_Stability\_of\_Embedding-based\_Word\_Similarities

50. Mitigating Semantic Collapsing Problem in Generative Personalization with Test-time Embedding Adjustment \- arXiv, https://arxiv.org/html/2506.22685v3

51. On assessing weaker logical status claims in Wikidata cultural heritage records, https://www.semantic-web-journal.net/content/assessing-weaker-logical-status-claims-wikidata-cultural-heritage-records

52. Mobi User Manual \- GitHub Pages, https://inovexcorp.github.io/mobi-docs/latest/index.html

53. MEXA: Multilingual Evaluation of English-Centric LLMs via ... \- arXiv, https://arxiv.org/html/2410.05873v2

54. Reveal the Unknown: Out-of-Knowledge-Base Mention Discovery with Entity Linking \- arXiv, https://arxiv.org/pdf/2302.07189

55. Hubness, Not Anisotropy, Drives Cross-Lingual Retrieval Asymmetry in Multilingual Embedding Models \- arXiv, https://arxiv.org/pdf/2605.26575

56. Robust Information Retrieval \- arXiv, https://arxiv.org/html/2406.08891v1

57. Difference in Political Argumentation Theory \- Scholarship @ UWindsor, https://uwindsor.scholaris.ca/bitstreams/bce38447-d09e-4a7b-9591-a194abcb85c0/download

58. The Chimerical Concept of Original Public Meaning \- Virginia Law Review, https://virginialawreview.org/articles/the-chimerical-concept-of-original-public-meaning/

59. On the explanatory value of the concept-conception distinction \- Milano-Bicocca, https://boa.unimib.it/retrieve/handle/10281/53305/80328/lalumera-concept.conception-10.14.pdf

60. The concept of vulnerability in law \- OpenEdition Journals, https://journals.openedition.org/revus/12019

61. TOWARDS A NEW GRAMMAR OF REASONING FOR ARTIFICIAL LEGAL INTELLIGENCE AND MECELLEM AS ITS SEMANTIC PROTOCOL \- arXiv, https://arxiv.org/pdf/2608.04011

62. Ambiguity and Engagement1 | American Journal of Sociology: Vol 124, No 3, https://www.journals.uchicago.edu/doi/10.1086/701298

63. New Flag Emoji Proposals No Longer Accepted, Unicode Says \- CNET, https://www.cnet.com/culture/internet/new-flag-emoji-proposals-no-longer-accepted-unicode-says/

64. Unicode will no longer accept proposals for new flag emoji \- Quartz, https://qz.com/2148267/the-reason-why-unicode-wont-add-new-flag-emoji

65. Our Response To Unicode's Announcement About Flags \- More Pride Emojis, https://moreprideemojis.com/response.html

66. Propose A New Emoji, But No More Flags \- Emojipedia Blog, https://blog.emojipedia.org/propose-a-new-emoji-but-no-more-flags/

67. The Past and Future of Flag Emoji \- The Unicode Blog, http://blog.unicode.org/2022/03/the-past-and-future-of-flag-emoji.html

68. Graph Reification | TrustGraph, https://trustgraph.ai/guides/key-concepts/graph-reification/

69. Ending the RDF vs. Property Graph debate with RDF\* | AI Business, https://aibusiness.com/data/ending-the-rdf-vs-property-graph-debate-with-rdf-

70. Declarative generation of RDF-star graphs from heterogeneous data \- Lirias, https://lirias.kuleuven.be/retrieve/e0e35748-290c-427d-b95a-4d5f2a869e59

71. TrieDF: Efficient In-memory Indexing for Metadata-augmented RDF \- CEUR-WS.org, https://ceur-ws.org/Vol-3225/paper4.pdf

72. \[Citation needed\]: provenance with RDF-star \- metaphacts Blog, https://blog.metaphacts.com/citation-needed-provenance-with-rdf-star

73. Ora Lassila and Adrian Gschwend: RDF 1.2 Working Group Update – Episode 50, https://knowledgegraphinsights.com/rdf12/

74. ProvCaRe Semantic Provenance Knowledgebase: Evaluating Scientific Reproducibility of Research Studies \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC5977728/

75. PROV-O: The PROV Ontology \- W3C, https://www.w3.org/TR/prov-o/

76. PROVENANCE & e-Science, http://sharp-sys.nl/provman/images/PROV%20overview.pdf

77. Provenance: close PROV-O compliance gaps and high-stakes trust blockers \#825 \- GitHub, https://github.com/semantica-agi/semantica/issues/825

78. PROV-O OWL \- W3C, https://www.w3.org/ns/prov-o.owl

79. Human-in-the-Loop Artificial Intelligence: A Systematic Review of Concepts, Methods, and Applications \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC13114286/

80. Active Learning Machine Learning: Boost AI Efficiency \- Group 107, https://group107.com/blog/active-learning-machine-learning/

81. From Insight to Intelligence: \- IGI Global, https://www.igi-global.com/viewtitle.aspx?TitleId=391185\&isxn=9798337344607

82. Harnessing Disagreement: Detecting Correlated Agreement Blindness in Multi-Agent Triage \- arXiv, https://arxiv.org/pdf/2607.19899

83. Cost-Sensitive Conformal Prediction and Human-in-the-Loop Abstention for Imbalanced High-Stakes Decision Support: A Multi-Domain Benchmark \- arXiv, https://arxiv.org/html/2607.27143

84. Active Boundary Discovery and Conservative Certification for Deployable Reinforcement Learning in Safety-Critical Con \- AWS, https://ijcai-preprints.s3.us-west-1.amazonaws.com/2026/AI4T22.pdf

85. Wikidata:Project chat/Archive/2021/01, https://www.wikidata.org/wiki/Wikidata:Project\_chat/Archive/2021/01

86. An Analysis of Links in Wikidata, https://wikidataworkshop.github.io/2022/papers/Wikidata\_Workshop\_2022\_paper\_1166.pdf

87. Wikipedia:Arbitration/Requests/Enforcement, https://en.wikipedia.org/wiki/Wikipedia:AEQUICK

88. Requests for comment/Do something about azwiki \- Meta-Wiki, https://meta.wikimedia.org/wiki/Requests\_for\_comment/Do\_something\_about\_azwiki

89. UAX \#44: Unicode Character Database, https://www.unicode.org/reports/tr44/

90. RFC 3718: A Summary of Unicode Consortium Procedures, Policies, Stability, and Public Access, https://www.rfc-editor.org/info/rfc3718/

91. UTR \#33 \- Conformance Model \- Unicode, http://www.unicode.org/reports/tr33/tr33-4.html

92. Accumulated Feedback on PRI \#350 \- Public Review Issues, http://www.unicode.org/review/pri350/

93. SNOMED International Release Management \- SNOMED Spaces, https://conf.spaces.snomed.org/wiki/spaces/RMT/pages/131992659/SNOMED+CT+Managed+Service+-+US+Edition+Release+Notes+-+September+2025

94. Maintenance and Management | Practical Guides SNOMED CT Reference Set Guide, https://docs.snomed.org/snomed-ct-practical-guides/snomed-ct-reference-set-guide/3-requirements-and-use-cases/3.2-use-cases/3.2.6-maintenance-and-management

95. Representing SNOMED CT Concept Evolutions using Process Profiles., https://ceur-ws.org/Vol-1660/womocoe-paper1.pdf

96. Maintaining Postcoordinated Expressions \- SNOMED CT Document Library, https://docs.snomed.org/snomed-ct-practical-guides/snomed-ct-postcoordination-guide/snomed-ct-expressions/maintaining-postcoordinated-expressions

97. METADATA REGISTRY, ISO/IEC 11179 \- UNT Digital Library, https://digital.library.unt.edu/ark:/67531/metadc926399/m2/1/high\_res\_d/973862.pdf

98. ISO/IEC 11179-1:2023 Framework Overview | PDF | Metadata | Information Science \- Scribd, https://www.scribd.com/document/788210638/ISO-IEC-11179-1-2023-Information-technology-Metadata-registries-MDR-Part-1-Framework

99. ISO/IEC 11179-6 \- iTeh Standards, https://cdn.standards.iteh.ai/samples/35348/eebbe699fb694712819e708ac9f3a9f1/ISO-IEC-11179-6-2005.pdf

100. FlureeDB — The Verifiable Knowledge Graph Database | Fluree, https://flur.ee/platform/core

101. Fluree Launches Verifiable Knowledge Graph Database for Agentic AI, https://flur.ee/blog/fluree-launches-verifiable-knowledge-graph-database-for-agentic-ai

102. Wikidata:Property proposal/Organization, https://www.wikidata.org/wiki/Wikidata:Property\_proposal/Organization

103. Wikidata:Project chat/Archive/2022/08, https://www.wikidata.org/wiki/Wikidata:Project\_chat/Archive/2022/08

104. Wikidata:Project chat/Archive/2022/03, https://www.wikidata.org/wiki/Wikidata:Project\_chat/Archive/2022/03

105. Wikidata:Project chat/Archive/2022/12, https://www.wikidata.org/wiki/Wikidata:Project\_chat/Archive/2022/12

106. Simple Knowledge Organization System (SKOS)† \- IMR Press, https://storage.imrpress.com/imr/journal/KO/article/504647/1752844795941.pdf

107. Diagnosing and Mitigating Semantic Inconsistencies in Wikidata's Classification Hierarchy \- arXiv, https://arxiv.org/html/2511.04926v2

108. History Supplements | Specifications SNOMED CT Expression Constraint Language, https://docs.snomed.org/snomed-ct-specifications/snomed-ct-expression-constraint-language/behaviour-specification-with-examples/6.11-history-supplements

109. The Hitchhiker's Guide to NAHLN Messaging Don't Panic \- Veterinary Extension of SNOMED CT, https://vtsl.vetmed.vt.edu/vs/Documents/HitchhikersGuideToNAHLN\_May2021.pdf

110. Analyzing SNOMED CT's Historical Data: Pitfalls and Possibilities \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC5333315/

111. Historical Association Reference Sets \- SNOMED CT Document Library, https://docs.snomed.org/snomed-ct-specifications/snomed-ct-release-file-specification/reference-set-release-file-specification/5.2-reference-set-types/5.2.1-content-reference-sets/5.2.1.4-association-reference-set/5.2.5.1-historical-association-reference-sets

112. Data analytics using SNOMED CT \- Health informatics and information technology, https://wardle.org/snomed/2021/07/26/making-valuesets.html

113. Practical Guide to Reference Sets \- NRCeS, https://www.nrces.in/download/files/pdf/doc\_ReferenceSetPracticalGuide\_Current-en-US\_INT\_20170516.pdf

114. Fluree Launches Verifiable Knowledge Graph Database for Agentic AI \- GlobeNewswire, https://www.globenewswire.com/news-release/2026/06/23/3316053/0/en/fluree-launches-verifiable-knowledge-graph-database-for-agentic-ai.html

115. Wikidata:Project chat/Archive/2019/07, https://www.wikidata.org/wiki/Wikidata:Project\_chat/Archive/2019/07

116. Wikidata:Project chat, https://www.wikidata.org/wiki/Wikidata:Project\_chat

117. Wikidata:Project chat/Archive/2017/11, https://www.wikidata.org/wiki/Wikidata:Project\_chat/Archive/2017/11

118. Wikidata:Requests for comment/Privacy and Living People, https://www.wikidata.org/wiki/Wikidata:Requests\_for\_comment/Privacy\_and\_Living\_People

119. Computation and Language 2025 \- arXiv, https://www.arxiv.org/list/cs.CL/2025?skip=15125\&show=2000

120. A Framework for Comprehensive Health Terminology Systems in the United States \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC61331/

121. Artificial Intelligence in Medicine: Technical Basis and Clinical Applications 9780128212592, 0128212594 \- DOKUMEN.PUB, https://dokumen.pub/artificial-intelligence-in-medicine-technical-basis-and-clinical-applications-9780128212592-0128212594.html

122. Identification and classification of ion-channels across the tree of life provide functional insights into understudied CALHM channels, https://elifesciences.org/reviewed-preprints/106134

123. (PDF) Robust Information Retrieval \- ResearchGate, https://www.researchgate.net/publication/381404355\_Robust\_Information\_Retrieval

124. FlippedRAG: Black-Box Opinion Manipulation Adversarial Attacks to Retrieval-Augmented Generation Models \- arXiv, https://arxiv.org/html/2501.02968v3

125. Semantic Stealth: Adversarial Text Attacks on NLP using Several Methods \- TechRxiv, https://www.techrxiv.org/doi/pdf/10.36227/techrxiv.171392778.89337138/v1

126. Robust Neural Information Retrieval: An Adversarial and Out-of-distribution Perspective, https://arxiv.org/html/2407.06992v1

127. liyongkang123/Awesome-Robustness-Adversarial-Attacks-and-Defenses-for-Information-Retrieval \- GitHub, https://github.com/liyongkang123/Awesome-Robustness-Adversarial-Attacks-and-Defenses-for-Information-Retrieval

128. Robust Neural Information Retrieval: An Adversarial and Out-of-distribution Perspective \- arXiv, https://arxiv.org/pdf/2407.06992

129. Robust Neural Information Retrieval: An Adversarial and Out-of-distribution Perspective, https://arxiv.org/html/2407.06992v2

130. Coverage Is Not Containment: A Fundamental Limit of Admission-Time Defenses Against Coordinated Poisoning of Vector Retrieval \- ResearchGate, https://www.researchgate.net/publication/412347762\_Coverage\_Is\_Not\_Containment\_A\_Fundamental\_Limit\_of\_Admission-Time\_Defenses\_Against\_Coordinated\_Poisoning\_of\_Vector\_Retrieval

131. A Fundamental Limit of Admission-Time Defenses Against Coordinated Poisoning of Vector Retrieval \- arXiv, https://arxiv.org/pdf/2608.16044

132. Understanding and Mitigating the Threat of Vec2Text to Dense Retrieval Systems, https://www.researchgate.net/publication/386550232\_Understanding\_and\_Mitigating\_the\_Threat\_of\_Vec2Text\_to\_Dense\_Retrieval\_Systems

133. 18 Talk and Thought \- Oxford Academic, https://academic.oup.com/book/36673/chapter/321706635

134. Hierarchical Group-Conditional Conformal Risk Control for Selective Prediction in Language Models \- arXiv, https://arxiv.org/html/2607.24562v1

135. Cross-lingual word embeddings for low-resource and ... \- UNB Scholar, https://unbscholar.lib.unb.ca/bitstreams/b63b6561-e5a9-42da-825a-795b214b44d0/download

136. Wikidata:Project chat/Archive/2019/05, https://www.wikidata.org/wiki/Wikidata:Project\_chat/Archive/2019/05

137. GitHub \- scrypster/muninndb: The cognitive database. A new class of data storage. Not a vector store, not a graph DB, not a RAG wrapper. Ebbinghaus decay, Hebbian learning, and Bayesian confidence are engine-native primitives. Memories evolve on their own. MCP-native. Single binary., https://github.com/scrypster/muninndb

138. Valid Selection among Conformal Sets \- OpenReview, https://openreview.net/pdf?id=S1Gxt8c8pC

139. LOINC, a Universal Standard for Identifying Laboratory Observations: A 5-Year Update, https://www.researchgate.net/publication/10842689\_LOINC\_a\_Universal\_Standard\_for\_Identifying\_Laboratory\_Observations\_A\_5-Year\_Update

140. The Building and Application of a Semantic e-Research Society, https://www.southampton.ac.uk/\~drn1e09/phd/Thesis.pdf

141. Documentation \- nomenclature.info, https://page.nomenclature.info/documentation.app?lang=en

142. SKOS Core Guide \- W3C, https://www.w3.org/TR/2005/WD-swbp-skos-core-guide-20051102/

143. Traxia: A Framework for Verifiable, Agent-Native Scientific Publishing \- arXiv, https://arxiv.org/html/2606.08256v1