Thick Models of Value

• Thick models of value are formal representations that structure and justify normative reasoning by embedding explicit values, norms, and attentional policies.
• They incorporate graph-based justificatory structures to distinguish enduring values from transient preferences, enhancing robust decision-making.
• TMVs are applied in AI alignment, economic valuation, and institutional regulation to support principled, context-sensitive normative operations.

Thick models of value (TMVs) refer to a class of formal representations that explicitly structure how values and norms are modeled, justified, and operationalized within agents, institutions, or computational systems. Unlike preference-based or unstructured text approaches, TMVs embed justificatory relations, deontic constraints, attentional policies, and normative reasoning procedures that support both individual and collective forms of value reasoning. This structuring enables systems to distinguish enduring values from transient preferences, facilitate normative judgment in novel domains, and model the social embedding of choice, with implications for economic, institutional, and AI alignment domains (Edelman et al., 3 Dec 2025, Alicea, 2014).

1. Fundamental Definition and Distinctions

A thick model of value is any representation of values and norms that (1) takes a direct stance on the ontology of values, (2) embeds them in justificatory structures—such as reasons and social practices—and (3) equips agents or institutions with procedures to reason from those structures in new contexts (Edelman et al., 3 Dec 2025). This sharply contrasts with alternative frameworks:

• Preferentist Modeling of Value (PMV): Encodes goals using utility functions $u: X\to\mathbb{R}$ or preference orderings $\succeq$ on outcomes $X$, but lacks information about the justificatory basis of these preferences, making no distinction between fleeting tastes, manipulative signals, and principled values.
• Values-As-Text (VAT): Records values as arbitrary text (e.g., "be helpful"), relying on downstream interpretation by LLMs with no internal structure for justification or differentiation.
• Thick Models of Value (TMV): Explicitly constrain value representation, enforce constitutive criteria, and operationalize graph-based, deontic, or policy-based structures that can be activated contextually and reasoned over.

This difference is crucial in the ability to support principled normative reasoning, represent collective goods, and filter out manipulative or ephemeral signals.

2. Core Formalism and Computational Structure

Thick models of value are formalized as the tuple: $\mathcal{M} = (V, N, G, A, R)$ where:

• $V$: Set of value concepts (e.g., honesty, autonomy).
• $N$: Set of norms or deontic constraints, each attached to a context $c \in C$, where $(c, S) \in N$ prescribes required/forbidden actions $S$ in $c$.
• $G = (V, E)$: Justificatory graph, with $(v_i \to v_j) \in E$ interpreted as "$v_i$ justifies $v_j$".
• $A = \{\alpha_h: H \to 2^V\}$: Family of attentional policies mapping histories or situations $H$ to relevant subsets of $V$.
• $R$: Normative reasoning procedure producing recommended actions by consulting $(V, N, G, A)$ (Edelman et al., 3 Dec 2025).

Special cases include:

• Values as Attentional Policies: $\alpha_h$ acts as a filter to identify constitutive criteria over raw slogans or unstructured preferences.
• Norm-Augmented Markov Games: Reward functions augmented by $N$ to instantiate social or community norms within agent objectives.
• Contractualist Reasoning: Universalization checks on proposals or norms, ensuring broad justifiability before adoption as shared constraints.

This architecture supports context-dependence, generalization, and explicit resolution of value conflicts.

3. Thick Models in Economic Valuation: Contextual Geometric Structures

In computational economics, thick models appear in the form of Contextual Geometric Structures (CGS) (Alicea, 2014). The formal setup is as follows:

• Let $D = \{d_1, ..., d_n\}$ be an $n$-dimensional feature space of culturally and perceptually salient attributes.
• Each agent is endowed with a CGS kernel $K = (\theta, a, g)$ where $\theta \in \mathbb{R}^n$ (prototypical coordinates), $a \in \mathbb{R}_+^n$ (scaling parameters), and $g$ (a tripartite genotype encoding dimension mutability, cryptographic acceptance hash, and flexibility).
• The soft classification function $C_K: \mathbb{R}^n \to [0,1]$,

$$C_K(x) = \exp\left\{-\sum_{i=1}^n \left(\frac{x_i - \theta_i}{a_i}\right)^2\right\}$$

or equivalently, $C_K(x) = \exp\{-d_K(x, \theta)^2\}$, embeds cognitive and cultural semantics into valuation.

• Kernel scaling allows agent conceptual spaces to dynamically evolve in response to novel or extreme stimuli via stochastic resizing of $a$.

Within market architectures, this enables:

• Minimal markets: Pairwise, thresholded bargaining over singleton goods, with evolving internal valuation bands partly determined by CGS.
• Compositional markets: Top-down auction mechanisms over bundled propositions, with community prices emerging from aggregate CGS-based accept/reject voting.

Valuation algorithms incorporate soft classification, historical price updating, symbolic/rarity premiums, and cryptographic transaction validation. This setup captures phenomena such as bubbles, non-transitive valuations, and propagating herding, which diverge systematically from efficient-markets pricing due to cultural, social, and cognitive embedding (Alicea, 2014).

4. Normative Reasoning and Principle Generalization

TMVs facilitate principled normative reasoning by building in auditability and transfer. Given explicit justificatory graphs, context-conditioned norm sets, and attentional policies, an agent can:

• Trace reasoning: Construct explicit chains of reasons for actions: $$(v_{i_1} \to \dots \to v_{i_k}), (c, S) \in N, \alpha_h(V)$$.
• Generalize across contexts: Apply procedures recursively through $G$ and adapt attentional focus via $A$, yielding robust behavior in novel or adversarial situations.
• Resolve conflicts: Identify higher-level unifiers or trade-off strategies in cases of value conflict (e.g., honesty vs. tact) by traversing $G$ for superordinate justificatory nodes.

For example, when tasked with content moderation, an agent using VAT may flip-flop, while a TMV-equipped agent with $V, N, G, A$ defined for learning, integrity, and helpfulness will provide normatively justified, context-sensitive responses, which are transparent and auditable (Edelman et al., 3 Dec 2025).

5. Modeling Collective Goods and Filtering Transient Preferences

TMVs natively account for collective goods and filter out fleeting or manipulative signals through their representational structure:

• Collective goods: Embedding shared norms in $N$ or leveraging collective decision procedures (e.g., contractualism, social-choice over value graphs) allows direct modeling of outcomes such as trust, fairness, or legitimacy. TMVs do not reduce these phenomena to mere extensions of individual utility functions, overcoming fundamental limitations of PMVs.
• Filtering ephemera: Values included in $V$ must satisfy non-instrumental, constitutive, and standpoint-invariant criteria, typically enforced by cross-perspective justification tests. Transient, manipulative, or irrational preferences ("momentary fads," "viral clickbait") are systematically excluded from $V$ as they lack justificatory persistence across agents and contexts (Edelman et al., 3 Dec 2025).

6. Applications in Artificial Intelligence, Institutions, and Economics

TMVs are foundational in several domains that require alignment, auditability, or robustness against incentive distortions:

• AI value stewardship: TMVs allow reflective elicitation of robust personal value graphs ($G_u \subseteq V$) and planning by attentional policies, counteracting engagement-maximization and supporting autonomy.
• Normative competence in agents: TMVs operationalize adaptive, context-resilient norm compliance (via norm-augmented reinforcement learning and contractualist modules) without rigid or naive rule execution.
• Negotiation and contract design: TMVs support value-based, justifiability-tested contracts, enabling credible win-win negotiation absent adversarial manipulation.
• Meaning-preserving economic mechanisms: Market intermediaries use TMVs to map supplier remuneration directly to clients’ constitutive values, rather than superficial proxies such as engagement or revenue.
• Democratic regulatory institutions: TMVs, when scaled via Moral Graph Elicitation and generative social choice, underpin rapid, community-traceable regulatory responses at institutional levels (Edelman et al., 3 Dec 2025, Alicea, 2014).

Representation Type	Ontology	Supports Normative Reasoning	Captures Collective Goods
PMV (Utility/Preference)	Utility/orderings	No	Limited, indirect
VAT (Values-As-Text)	Text strings	No	No
TMV (Thick Model)	Structured objects	Yes	Yes

7. Implications for Theory and Evolution

TMVs synthesize insights from economic theory, cultural evolution, cognitive neuroscience, and computational representation. In economic settings, TMVs account for the origins of bubbles, herding, and price divergence not as market pathologies, but as emergent properties of cognitively and culturally embedded agents exchanging symbolic, meaning-laden goods (Alicea, 2014). In institutional and AI contexts, TMVs provide the basis for full-stack alignment—ensuring that systems, agents, platforms, and regulators remain coherently and audibly aligned with complex human values. This multi-level approach bridges micro-foundational neural computation and macro-level social and economic patterns, offering an extensible framework for ongoing integration of normative structure into autonomous systems (Edelman et al., 3 Dec 2025, Alicea, 2014).