Relational-Norms Framework Overview

Updated 1 March 2026

The Relational-Norms Framework is a multidisciplinary model that defines and interrelates social, ethical, and epistemic norms for complex sociotechnical systems.
It operationalizes abstract normative concepts into measurable metrics across operational, epistemic, and normative domains for coherent system evaluation.
It employs formal methods such as norm nets and optimization to ensure conflict-free, auditable frameworks that support effective AI governance and human–AI interaction.

A relational-norms framework formalizes how social, ethical, and epistemic norms are structured, operationalized, and interrelated—conceptually and computationally—for use in reasoning about, designing, or governing complex systems (AI, institutional, multi-agent, or conversational). In cutting-edge research, the term denotes multidisciplinary architectures and explicit formalisms that capture (1) semantic relationships among normative concepts, (2) interdependencies with operational and epistemic metrics, and (3) methods to encode, infer, or optimize over sets of norms, including their application to AI governance, regulation, and modeling human–AI interaction.

1. The Relational Problem and Multidomain Conceptual Architecture

Straub et al. identify the “relational problem” as a foundational ontological challenge for institutional AI: semantic ambiguity, lack of explicit relations among key concepts, and divergent terminologies across ML, HCI, law, social science, and policy impede coherent governance and deployment (Straub et al., 2023). This yields “conceptual isolation,” preventing unified audits, metric design, or policy interventions. The relational-norms framework directly addresses this through a three-domain horizontal architecture:

Operational Domain: system-centric, quantifiable properties—accuracy, efficiency, reliability, robustness.
Epistemic Domain: knowledge-centric, human–machine interface properties—interpretability, explainability, transparency, reproducibility.
Normative Domain: moral, legal, and rights-based entitlements—fairness, justice, equality, welfare, accountability.

Each metric or governance tool in one domain is interdependent with, or operationalizes, concepts in the others: e.g., a fairness (Normative) audit presupposes operational metrics (classification rates) and epistemic intelligibility to domain experts.

Diagrammatic Overview

	Micro (single app)	Meso (family)	Macro (system class)
Operational	accuracy, efficiency, benchmarks	...	...
Epistemic	interpretability, model cards	...	...
Normative	fairness, justice, impact assessment	...	...

Analyses must scan this 3×N design space at varying institutional scales.

2. Formal Representations: Modeling Relationships Among Norms

The explicit modeling of relationships among norms is foundational for both regulation and multi-agent system design.

Norm Net Formalism

Governatori and colleagues (Lopez-Sanchez et al., 2017) define a norm net as:

$\mathsf{NN} = \langle N, R \rangle$

where

$N$ : set of norms
$R = \{R_g,\,R_x,\,R_s\}$ $R = {R_{g}, R_{x}, R_{s}}$ :
- $R_g$ = generalisation relation (transitive, irreflexive, anti-symmetric); $(n_i,n_j)\in R_g$ means $n_i$ is more general than $n_j$
- $R_x$ = exclusivity relation (mutually exclusive, irreflexive, symmetric, intransitive)
- $R_s$ = substitutability relation (mutually substitutable, irreflexive, symmetric, transitive)

A sound subset (norm system) must be conflict-free ( $R_x$ ), non-redundant ( $R_g$ ), and not contain substitutable duplicates ( $R_s$ ).

Norm representation power $r(n_i)$ , cost $c(n_i)$ , and moral-value support $\operatorname{val}(n_i)$ parameterize each norm. Linear and multi-objective optimization (often as MIP) selects optimal, non-conflicting systems under constraints.

3. Operationalization in Institutional and Technical Contexts

The framework translates high-level principles into actionable metrics, standards, and mechanisms across domains (Straub et al., 2023, Lopez-Sanchez et al., 2017):

Operationalization of Normative Concepts: Fairness, justice, and equality must be instantiated as audit criteria (e.g., demographic parity, equality of opportunity), mapped to key performance indicators, and checked for institutional compliance.
Algorithmic Impact Assessment: Structured procedures document and forecast harms/benefits, including distributional and rights-based analyses.
Human-in-the-Loop Overrides: Procedural protocols specify exactly when and how human agents must intercede in automated decision processes.

Example: In a public-sector recommender,

Operational: Click-through rates, latency, and error rates are quantified.
Epistemic: Interpretability is evaluated via user- and documentation-focused methods.
Normative: AIAs and governance protocols are invoked to protect equality and welfare.

Relational-norms frameworks are employed in diverse domains:

Multi-agent social systems: Higher-order network models embed relational norms into both dyadic and triadic interactions, allowing the emergence and maintenance of cooperation via explicit norm and strategy update rules that transcend simple payoff imitation. Prosocial and antisocial norms (e.g., KJD, SIC, AIC, EBH) coevolve with link-level directed strategies, and information privacy modulates the diffusion of cooperative norms (Ma et al., 2024).
Human behavioral modeling: Bayesian Inference of Social Norms formalizes norms as latent, shared behavioral constraints and employs joint inference over agent-level desires and actions. Observers infer the governing norm structure from observed compliance and enforcement behaviors among agents (Tan et al., 2019).
Institutional and economic macrostructures: Three-stage models posit that individual-level reciprocal dynamics, via sufficient memory and cost–benefit heuristics, probabilistically aggregate into population-level norm stabilization, which scaffolds the emergence of formal institutional constraints (Diau, 13 May 2025). Agents remain cognitively minimal throughout.

Recent work formalizes relational norms as mappings from relationship type and function to prescriptive, proscriptive, or neutral evaluations. For human–AI systems, this mandates that AI behaviors be role-conditioned to respect the distinct normative functions (care, transaction, hierarchy, mating) expected in each relationship class (e.g., teacher–student, parent–child, employer–employee). Norm “profiles” assign signed strength to each cooperative function per relationship, informing behavioral policy shaping, guardrail setting, and role-specific transparency for AI agents (Earp et al., 17 Feb 2025). The absence of consciousness, immunity to fatigue, and lack of self-interest in AI agents requires adaptation and sometimes redefinition of familiar human–human norms to avert normative confusion, misplaced trust, or hazardous dependencies.

In communication, relational-norms architectures use relationship-conditioned functions Φ(r, m) to predict and enforce contextually appropriate utterances. Models operationalize appropriateness as P(A=1 | r, m), train on relationally diverse message–role datasets, and support downstream tasks such as condescension and politeness detection (Jurgens et al., 2023). Normative violations in dialog can be detected, annotated, and corrected with respect to interlocutor attributes, social roles, and shared situational ground (Wu et al., 2024).

6. Outstanding Challenges and Research Questions

Several open questions guide the future evolution of relational-norm frameworks (Straub et al., 2023):

Precision of domain definitions: How strictly should the operational, epistemic, and normative domains be defined to avoid conflation and category error?
Membership and prioritization: Which concepts or metrics belong to which domains, and how should trade-offs be weighed?
Domain focus and meta-structures: Is an imbalance in effort across domains acceptable, or should a meta-level unification be sought?
Cross-domain feedback: How do discoveries or shifts in one domain propagate definitional or methodological changes in others?
Universality and local contextuality: Which methods and metrics generalize, and which are inherently context-dependent?
Quantitative measurement: How can the relationships between concepts and methods across domains be quantified and compared?

The relational-norms framework is discipline-agnostic and is ultimately justified by its capacity to provide rigorous, checklist-based coverage of all relevant evaluative perspectives for complex sociotechnical systems. Its central value lies in enabling explicit, auditable, and operational linkage between measurable properties, interpretability, and value-laden institutional goals (Straub et al., 2023, Lopez-Sanchez et al., 2017, Earp et al., 17 Feb 2025, Ma et al., 2024, Diau, 13 May 2025, Tan et al., 2019).