Towards Evolving Actor–Network Ontologies: Enabling Reflexive Digital Twins for Cultural Heritage

Abstract

This paper introduces the concept of evolving actor–network ontologies (EANO) as a new paradigm for cultural digital twins. Building on actor–network theory, EANO reframes ontologies from static representations into reflexive, dynamic structures in which semantic interpretations are continuously negotiated among heterogeneous actors. We propose a five-layer architecture that operationalizes this principle, embedding reflexivity, actor salience, and systemic parameters such as resistance and volatility directly into the ontological model. To illustrate this approach, we present minimal simulations that demonstrate how different actor constellations and systemic conditions lead to distinct patterns of semantic evolution, ranging from expert erosion to contested equilibria and balanced coexistence. Rather than serving as predictive models, these simulations exemplify how EANO captures semantic plurality and contestation within a transparent and interpretable framework. The contribution of this work is thus twofold: it provides a conceptual foundation for evolving ontologies in digital heritage and a lightweight demonstration of how such models can be instantiated and explored computationally.

1. Introduction

Digital twins have emerged as transformative instruments in cultural heritage, enabling the integration of spatiotemporal models, sensor data, expert knowledge, and AI reasoning to support documentation, risk management, and sustainable conservation strategies. As these systems evolve, their semantic backbones, typically expressed as ontologies, are expected to capture not only what a heritage asset is but also how it is perceived, threatened, and acted upon by a diversity of stakeholders. However, current digital twin architectures largely rely on static or narrowly evolving ontologies that reflect a fixed worldview, limiting their ability to accommodate the dynamic, contested, and actor-dependent nature of heritage interpretation and management.

In recent work, we introduced PANOPTES [1,2,3], a semantic framework for modeling heritage assets, threats, policies, and decision-making contexts through a structured ontology grounded in standards such as CIDOC-CRM [4], PROV-O [5], and SOSA/SSN [6]. This foundation was extended with the most recent Agentic PANOPTES architecture [7], which added temporal memory, policy responsiveness, and risk-aware reasoning by embedding the ICCROM/CCI ABC risk model [8]. These layers supported basic forms of AI-driven interpretation and reflex-like responses, but the core semantic structure remained essentially stable across time and stakeholder interactions.

In this paper, we propose a shift in paradigm: the integration of Actor–Network Theory (ANT) [9,10,11,12] as a conceptual and computational basis for evolving ontologies that adapt to changing configurations of meaning, relevance, and power within heritage contexts. ANT reframes knowledge not as a neutral reflection of reality, but as an emergent product of dynamic interactions between heterogeneous actors, including human and non-human, material, and discursive actors. Inspired by this view, we propose the construction of evolving actor–network ontologies capable of restructuring their own roles, relationships, and classification schemes as actors engage, negotiate, and influence the semantic environment over time.

This evolution enables a new generation of reflexive digital twins: systems that do not merely represent heritage statically but adaptively restructure their semantic fabric in response to shifting contexts, contested meanings, and emergent events. Reflexivity in this study refers to the system’s capacity to model not only the world but also the changing terms, through which it understands and organizes that world, making it epistemologically aware and semantically plastic. In contrast to adaptive systems that typically respond to new inputs, reflexive systems modify the very semantic structures they use to interpret those inputs based on shifting actor relationships and evolving contextual meaning.

To demonstrate this vision, we present a minimal simulation scenario showing how such a system may reclassify an asset or reinterpret a threat based on the interaction of stakeholders and policy feedback. We also outline the architectural, technical, and ontological implications of this shift, offering a roadmap for operationalization evolving actor–network ontologies in practice.

This work marks the next step in the PANOPTES trajectory, from static semantic modeling, through agentic reasoning, to structurally reflexive and epistemically aware cultural heritage AI systems. We argue that only by embracing semantic evolution can cultural digital twins truly align with the pluralistic, negotiated, and historically situated nature of heritage itself. It is important to emphasize that this paper is a theoretical contribution. The concepts, architectures, and simulations discussed herein are part of an emerging design framework, not a validated or deployed system nor the result of empirical implementation. Our aim is to articulate a forward-looking vision for semantically reflexive digital twins, grounded in both ontological engineering and actor–network epistemology. By proposing this conceptual synthesis, we seek to stimulate interdisciplinary dialogue and guide future development efforts at the intersection of semantic AI, heritage theory, and dynamically evolving knowledge systems.

2. Background and Theoretical Foundations

On the front of semantic digital twins and ontological foundations, digital twins in cultural heritage increasingly incorporate semantic technologies, ontologies, and knowledge graphs to represent assets, stakeholders, threats, and interventions in a machine-interpretable form. A systematic literature review of 82 research articles highlighted the relevance of ontologies in digital twin architectures across domains, including manufacturing, infrastructure, and cultural heritage, while also identifying persistent limitations in ontology evolution and adaptability in dynamic contexts [13].

Standards-based cultural heritage modeling remains grounded in CIDOC-CRM, the primary ontology for representing historical knowledge. It provides a formalized and extensible vocabulary for describing entities, events, actors, and relationships with semantics enabling machine-to-machine interoperability [4]. On this foundation, initiatives such as the Heritage Digital Twin Ontology (HDTO) and related frameworks have emerged to support interoperability, modularity, provenance, and diachronic modeling in heritage digital twins [14,15]. The HDTO defines classes such as HC1 Heritage Entity and HC2 Digital Twin and establishes properties (e.g., “is digital twin of”) to semantically interconnect representations, documentation, events, and preservation actions within a unified knowledge graph.

Sensor-aware digital twin architectures, such as the Reactive Heritage Digital Twin Ontology (RHDTO), extend these capabilities by explicitly encoding sensor data, activators, and decision workflows, allowing heritage digital twins to operate inductively over time through sensor-driven, policy-aware reasoning [16].

In parallel, studies implementing narrative-based knowledge representation frameworks, such as NKRL (Narrative Knowledge Representation Language), explore the modeling of immaterial and symbolic dimensions of heritage, enabling digital twins to encompass not only tangible features but also narratives, values, and emotive interpretations [17,18].

PANOPTES-based semantic infrastructures [2] further advanced this landscape by introducing a comprehensive model for cultural asset representation, integrating spatial, temporal, and threat dimensions with compatibility to standards like CIDOC-CRM [4], GeoSPARQL [19], OWL-Time [20], and SOSA/SSN [6]. Building on this foundation, the Agentic PANOPTES framework [7] embedded risk memory and adaptive behavior into digital twins through agent-based modeling of the ICCROM/CCI ABC methodology, enabling simulation of preventive conservation scenarios and reflex-like interventions.

While these ontological models offer structural richness and semantic interoperability, they still operate over largely static schemas with limited capacity for runtime evolution. Traditional solutions such as ontology versioning, modularization, and alignment permit schema updates but typically under centralized, manual control, lacking actor-driven or pluralistic reconfiguration mechanisms [21].

Meanwhile, heritage informatics and critical heritage studies emphasize the contested, value-laden, and negotiated nature of heritage interpretation, where classification, meaning, and significance shift over time across communities, institutions, and stakeholders [22]. These theoretical insights challenge assumptions of fixed ontological modeling and underscore the need for reflexive, adaptive semantics.

Actor–Network Theory (ANT), developed by Latour, Callon, Law, Muniesa, and others, posits that meaning, facts, and social structure emerge through negotiation within heterogeneous actor–networks that encompass human, material, institutional, and symbolic entities [9,10,11,12,23]. ANT-informed modeling avoids privileging any fixed interpretive center and favors translation, stabilization, and contestation as ongoing processes. ANT-informed ontology engineering remains a nascent but conceptually rich domain examining how actor interactions and sociotechnical negotiations can shape ontology structure. While concrete implementations in cultural heritage and digital twin systems are rare, foundational work in modular ontology engineering provides critical insights. Theoretical models of ontology modularization and logical structuring, especially within OWL and Description Logics, establish principles for scalable and adaptive semantic frameworks [24,25]. Broader surveys of knowledge graph evolution explore schema dynamics, semantic drift, and reasoning over heterogeneous knowledge sources [26]. More recent sociotechnical approaches advocate for modular designs grounded in social roles, institutional logic, and stakeholder diversity, aligning closely with actor–network-informed perspectives [27].

Building on this landscape, our evolving actor–network ontology approach merges the ontological formality of CIDOC-CRM and HDTO/RHDTO with the dynamic and reflexive imperatives of ANT-informed theory and the emerging concept of reflexive digital twins, which are systems that can reevaluate and restructure their own semantic foundations based on actor influence, evidence, and interpretive contestation. In contrast to static or manually versioned schemes, a reflexive digital twin is capable of monitoring classification uncertainty, semantic entropy, and disputed meanings and adjusting its ontological schema accordingly. This meta-reasoning extends beyond updating entity instances or property values: it enables reclassification of assets, schema bifurcation, introduction of new properties or roles, or even redefining classes, triggered by shifts in stakeholder significance, institutional authority, or emergent epistemic tension. By synthesizing established ontology engineering practices with ANT-style dynamics, we aim to operationalize evolving actor--network ontologies within cultural heritage digital twins, providing both semantic stability and pluralistic, contested adaptability.

3. Architecture Framework for Reflexive Ontologies

Evolving Actor–Network Ontologies (EANO) represent a foundational shift from static, taxonomic knowledge representation to dynamic, actor-configured semantic networks. Rather than modeling entities and relationships as fixed hierarchies, EANO supports a negotiated view of meaning, where roles, categories, and ontological commitments may shift in response to actor interactions, power dynamics, or epistemic reframings. To assess its feasibility, a minimal simulation is presented in Section 4, illustrating how actor influence, classification divergence, and reflexive schema restructuring unfold over time.

At the heart of this architecture lies a layered model comprising the following:

• Base Ontology Layer: This encodes core entity types and relationships, drawing from standards such as CIDOC-CRM, DUL, PROV-O, SOSA/SSN, and PANOPTES. It provides initial semantic grounding and interoperability.
• actor–network Layer: Maintains dynamic assemblages of human and non-human actors (e.g., institutions, stakeholders, sensors, documents), including alliances, oppositions, and influence relations. Actors may introduce or reinterpret concepts and propose reclassifications.
• Evolving Schema Layer: This maintains a mutable schema whose structure can be modified as a result of validated actor–network dynamics. It tracks classification change, schema evolution, and lineage of concepts.
• Reflexive Control Layer: This supervises semantic shifts, detects inconsistencies, and governs the admissibility of ontological change. It enables epistemic reflexivity by modeling how the system evaluates its own assumptions.
• Interaction Interface Layer: This supports annotation, conflict resolution, and actor-driven feedback mechanisms. It enables participation in meaning negotiation and schema co-evolution.

To implement this architecture, we map each conceptual layer to a concrete component:

• (1) Base Ontology Layer → Semantic Core Modules. These include

  -

  Asset and Context Schema: These are based on CIDOC-CRM, SOSA/SSN, and DUL, modeling cultural entities, threats, spatiotemporal extents, and measurements.

  -

  Agent and Policy Models: These encode actor roles, interventions, and institutional behavior.

  -

  Versionable Modules: Each semantic unit (e.g., classifications, threat taxonomies) is tracked over time with provenance and update history.

• (2) Actor–Network Layer → Actor Graph Engine. A dynamic, weighted graph in which

  -

  Nodes represent actors (institutions, communities, algorithms, etc.);

  -

  Edges represent influence, alignment, or antagonism;

  -

  Weights encode evolving actor influence based on observation, feedback, or reinforcement.

• (3) Evolving Schema Layer → Schema Evolution Mechanism. Implements structural adaptations such as

  -

  Reclassification of assets;

  -

  Splitting or merging categories;

  -

  Insertion of new properties, concepts, or actors.

  It is tracked via versionable modules and linked to actor claims, for example, through PROV-O.
• (4) Reflexive Control Layer → Ontology Adaptation Engine. A meta-level monitor that

  -

  Observes actor influence and classification divergence;

  -

  Triggers evolution rules (predefined or learned);

  -

  Logs rationales for change via a Change Justification Module.

• (5) Interaction Interface Layer → Participatory Interface Modules.

  -

  Enable semantic annotation and feedback;

  -

  Support conflict resolution workflows;

  -

  Visualize ontology evolution and schema change.

• (+) Agentic Reasoning Module. While not a separate conceptual layer, this component enhances decision-making by

  -

  Allowing agents to reason over both the current world state and evolving semantic schema;

  -

  Generating alerts or recommendations with awareness of classification uncertainty.

This alignment allows digital twins to not only reason over cultural heritage data but also to revise the very ontological structures through which meaning is derived. Reflexivity is thus operationalized as the capacity for controlled, actor-driven semantic restructuring. For example, if a conservation body reclassifies a monument from “vulnerable” to “critically endangered,” the actor–network registers the shift, the adaptation engine evaluates schema impacts, and the schema evolution module may restructure the classification taxonomy accordingly. These dynamics are explored in the accompanying simulation.

4. Simulation Methods and Validation

4.1. Design Goals and Realism Claims

Our goal is not to forecast future states of a site but to show that a reflexive twin can (i) remain anchored to a domain model used in practice (ICCROM/CCI ABC), (ii) use surface contested framings (Cultural vs. Functional) as bloc weights shift, and (iii) do so with parameters that a practitioner can set and audit. All simulations were implemented in an open Google Colab notebook, published for transparency and reproducibility (Code available at http://bit.ly/4mIFVeN, accessed on 1 October 2025). We therefore make three modest realism claims:

• Model grounding. Risk dynamics are driven by the ABC method, A (frequency/rate), B (loss per item), and C (fraction/items affected), each reported on the standard half-step grid $\{1,1.5,\dots ,5\}$; magnitude of risk $MR=A+B+C\in [3,15]$.
• Institutional transparency. We model four actors that reflect recognizable decision lenses in heritage governance: Expert (archaeologist, conservation scientist) emphasizes authenticity and technical integrity, responding strongly to loss-per-item (B) and extent (C); Community represents local attachment and social meaning, attentive to frequency of events (A) and to extent (C); Policymaker prioritizes thresholds and actionable signals, activating on $A\uparrow$, $\Delta MR>0$, and crossings of $M{R}_{\mathrm{high}}$; and the AI Agent serves as a prognostic module, detecting trends in $\Delta MR$ and widening exposure ($A+C$). Taken together, Expert+Community form a Cultural bloc, while Policymaker+AI form a Functional bloc. Near-parity between blocs is interpreted as a Hybrid or contested framing, reflecting observed tensions between cultural significance and functional risk management.
• Deterministic interpretability. All updates are deterministic functions of ABC deltas and $MR$ trends. Optional jitter is strictly positive and bounded (for sensitivity bands), never the driver of regime changes.

In short, regarding the actor blocs and framing intuition, Expert and Community foreground cultural significance, while Policymaker and AI foreground functional governance. The distinction is not arbitrary: it mirrors real heritage negotiations, where conservators and local actors emphasize meaning and authenticity, while decision-makers and predictive systems emphasize thresholds, risks, and resources. The Hybrid state is visible when neither side dominates and framing is contested.

4.2. From ABC to Evolving Actor Weights (EANO)

The EANO mechanism combines two layers: (i) the continuous scenario dynamics of the ICCROM/CCI ABC method and (ii) reflexive weight updates that translate these dynamics into negotiated actor influence.

Practitioners working with the ABC method record values only on the discrete half-step grid $\{1,1.5,\dots ,5\}$. For realism, however, our simulation first generates continuous trajectories $(A_c\left(t\right),B_c\left(t\right),C_c\left(t\right))$ with scenario-specific dynamics (e.g., a rise-and-fall flood peak, a slow cumulative fade, or a localized shock). These curves may include small, strictly positive jitter to avoid unrealistic stalls. The reported values are then

$$\left(A\left(t\right),B\left(t\right),C\left(t\right)\right)={\mathrm{round}}_{0.5}\left(A_c\left(t\right),B_c\left(t\right),C_c\left(t\right)\right)$$

(1)

and the magnitude of risk is

$$MR\left(t\right)=A\left(t\right)+B\left(t\right)+C\left(t\right),MR\left(t\right)\in [3,15]$$

(2)

For actor reactions we always use the continuous deltas $(\Delta A_c,\Delta B_c,\Delta C_c)$ so that weight shifts remain smooth and not limited by the half-step quantization.

Regarding actor weights and reflexive updates, we maintain a weight vector ${\mathbf{w}}_t\in {\mathbb{R}}^4$ (Expert, Community, Policymaker, AI Agent), renormalized to sum 1 at every step. The guiding intuition is that actors respond to what they “see” in the ABC curves:

• Zero-sum competition: No actor can grow without another losing ground, preventing runaway drift.
• Interpretability gates: Expert rises when losses per item B (or affected fraction C) increase; Community rises when frequency A or spread C widens; Policymaker activates only on actionable signals ($A\uparrow$, $\Delta MR>0$, or threshold crossings); and AI responds to risk trends and combined $A+C$ growth.
• Reflexive control: Mild entropy regularization prevents monopolies, and when overall risk $MR$ is low, the weights relax back toward their initial mix.

Another important aspect of the simulation relates to the formalization of pressure. Actor updates are driven by a notion of pressure, which aggregates how each actor reacts to observed changes in the ABC dynamics and overall risk level. Let $S\in {\mathbb{R}}^{4\times 3}$ be the sensitivity matrix whose rows correspond to actors (Expert, Community, Policymaker, AI) and columns to the continuous deltas $(\Delta A_c,\Delta B_c,\Delta C_c)$. For example,

$$S=\left[\begin{array}{ccc}0.06& 0.48& 0.32\\ 0.26& 0.05& 0.36\\ 0.18& 0.10& 0.10\\ 0.20& 0.15& 0.22\end{array}\right]$$

(3)

so the Expert reacts weakly to frequency changes $(\Delta A_c=0.06)$ but strongly to loss-per-item $(\Delta B_c=0.48)$ and extent $(\Delta C_c=0.32)$, while Community is more sensitive to frequency and extent.

Global responses to the magnitude of risk $MR$ are encoded in two 4D vectors of trend–sensitivity coefficients:

$${\theta }_{+}=(0.20,0.12,0.10,0.18),{\theta }_{-}=(0.08,0.06,0.10,0.08)$$

(4)

where entries correspond to (Expert, Community, Policymaker, AI Agent). These coefficients specify how each actor adjusts weight in response to risk dynamics. The vector ${\theta }_{+}$ applies when $MR$ is rising ($\Delta MR>0$), and ${\theta }_{-}$ applies when $MR$ is falling ($\Delta MR<0$). For example, the Expert ($0.20$) and AI Agent ($0.18$) respond most strongly to increases in $MR$, the Community is moderately responsive ($0.12$), while the Policymaker shows the least responsiveness to rises ($0.10$). During declines, the Policymaker becomes slightly more reactive ($0.10$) than the others, whereas the Community ($0.06$) relaxes most cautiously. Overall, Experts and AI are trend-sensitive, Policymakers are threshold-driven, and Communities are the most inertial.

The raw pressure vector is defined as

$${\delta }_t^{\mathrm{raw}}=S\left[\begin{array}{c}\Delta A_c\left(t\right)\\ \Delta B_c\left(t\right)\\ \Delta C_c\left(t\right)\end{array}\right]+{\theta }_{+}max(\Delta MR\left(t\right),0)+{\theta }_{-}min(\Delta MR\left(t\right),0)+{\gamma }_t$$

(5)

where $S\in {\mathbb{R}}^{4\times 3}$ maps instantaneous ABC deltas to actor-specific reactions, and ${\theta }_{\pm }\in {\mathbb{R}}^4$ encode global responsiveness to rising or falling $MR$. The ${\gamma }_t$ term introduces domain-specific rational boosts: (a) Experts gain on visible loss per item ($B\uparrow$), (b) Communities on extent ($C\uparrow$), (c) Policymakers on actionable signals ($A\uparrow$, $\Delta MR>0$, and small bonuses on threshold crossings), and (d) the AI Agent on trend detection ($\Delta MR$ and $A+C$). When $MR$ eases, recovery and give-back coefficients reverse these boosts, reallocating weight toward Experts and Communities while reducing Policymaker and AI influence. This ensures that all shifts remain interpretable and anchored in recognizable conservation and governance rationales. To preserve zero-sum balance, the mean pressure is subtracted so that total gain equals total loss:

$${\delta }_t={\delta }_t^{\mathrm{raw}}-\frac{1}{4}\sum _{i=1}^4{\delta }_{t,i}^{\mathrm{raw}}\mathbf{1}$$

(6)

where $\mathbf{1}$ is the 4D all-ones (unity) vector. This guarantees that actor weights remain competitive reallocations rather than drifting upward or downward in aggregate.

Having defined the zero-sum pressure ${\delta }_t$, the weights are updated in two stages. First, they are updated with step size $\eta$ and resistance $\rho$, and inertia and adoption are combined as

$${\tilde{\mathbf{w}}}_{t+1}=(1-\rho ){\mathbf{w}}_t+\rho max\left({\mathbf{w}}_t+\eta {\delta }_t,\mathbf{0}\right)$$

(7)

Then, we apply entropy regularization and relaxation to yield the finalized update:

$${\mathbf{w}}_{t+1}={\Pi }_{\Delta }\left((1-ϵ){\tilde{\mathbf{w}}}_{t+1}+ϵ\frac{1}{4}\mathbf{1}\right)$$

(8)

where ${\Pi }_{\Delta }$ renormalizes to sum 1 and $ϵ=0.02$ is a small entropy weight. Here ${\Pi }_{\Delta }$ denotes the projection onto the probability simplex ${\Delta }^3=\{\mathbf{w}\in {\mathbb{R}}^4\mid w_i\ge 0,{\sum }_i{w}_i=1\}$ [28], so that updated weights remain valid actor shares. This interpolation with the uniform distribution $(\frac{1}{4},\frac{1}{4},\frac{1}{4},\frac{1}{4})$ corresponds to maximizing the Shannon entropy of the weight vector, preventing monopolization by a single actor. In practical terms, entropy ensures that all actors retain a minimal presence, making recoveries possible and reflecting the plural character of heritage governance. If $MR\left(t\right)<M{R}_{\mathrm{low}}$, an additional relaxation nudges ${\mathbf{w}}_{t+1}$ back toward its baseline ${\mathbf{w}}_0$.

Finally, bloc-level framings are derived from actor weights. The Cultural share is defined as $S_{\mathrm{cult}}=w_{\mathrm{Exp}}+w_{\mathrm{Comm}}$, the Functional share as $S_{\mathrm{func}}=w_{\mathrm{Pol}}+w_{\mathrm{AI}}$, and their difference $\Delta S=S_{\mathrm{cult}}-S_{\mathrm{func}}$ determines the state. With a tie-band $\tau$, the framing state is given by

$${\mathrm{state}}_t=\left\{\begin{array}{cc}\mathrm{Cultural},\hfill & \Delta S_t>\tau ,\hfill \\ \mathrm{Hybrid},\hfill & |\Delta S_t|\le \tau ,\hfill \\ \mathrm{Functional},\hfill & \Delta S_t<-\tau \hfill \end{array}\right.$$

(9)

so near-parity is classified as Hybrid rather than producing brittle binary flips.

Taken together, Equations (1)–(9) specify a reflexive mechanism. Actors shift their influence in response to interpretable cues from the ABC dynamics and overall risk trends, while zero-sum redistribution, entropy regularization, and low-risk relaxation prevent collapse into single-actor dominance. The resulting framing transitions emerge endogenously from the scenario logic, remain deterministic and reproducible under fixed presets, and are auditable by practitioners through the same parameters they use in real-world assessments.

4.3. Scenarios and User-Visible Parameters

To make the simulation interpretable and realistic, we define three canonical scenarios that correspond to recognizable patterns in heritage risk assessment practice. These are not arbitrary: each mirrors examples discussed in the ICCROM/CCI ABC manual and has direct conservation relevance. The corresponding preset values used in the simulations are listed in

Table 1. Scenario presets (values match the interactive sliders). All weights are renormalized to sum 1.

Scenario	${\mathit{A}}_0$	${\mathit{B}}_0$	${\mathit{C}}_0$	${\mathit{w}}_0$	$\mathit{\rho }$	$\mathit{\tau }$	${\mathit{M}\mathit{R}}_{\mathbf{low}}$	${\mathit{M}\mathit{R}}_{\mathbf{high}}$	${\mathit{\sigma }}_{+}(\mathit{A},\mathit{B},\mathit{C})$
Seasonal Flooding (event)	3.0	2.5	3.0	(0.30, 0.35, 0.20, 0.15)	0.30	0.16	7.4	9.6	(0.05, 0.02, 0.04)
Light Fading (cumulative)	2.0	3.5	2.0	(0.42, 0.28, 0.18, 0.12)	0.22	0.12	7.8	9.4	(0.05, 0.02, 0.04)
Theft/Vandalism (shock)	1.5	3.5	1.5	(0.35, 0.25, 0.30, 0.10)	0.20	0.12	7.2	9.0	(0.05, 0.02, 0.04)

.

• Seasonal Flooding represents an event-driven hazard: frequency (A) and extent (C) increase to a mid-run peak, while loss-per-item (B) shows a temporary surge before easing. This captures situations where water ingress damages sites, but subsides after the flood peak.
• Light Fading illustrates a cumulative hazard: per-item loss (B) is persistently high, while frequency (A) and extent (C) rise gradually. This corresponds to slow but irreversible deterioration (e.g., pigment fading and material degradation) where Expert and Community concerns dominate and policy intervention remains mild.
• Theft/Vandalism models a localized shock: frequency (A) spikes abruptly in a short window while loss-per-item (B) is high but extent (C) remains low. This captures episodic but severe disturbances that trigger policy activation and vigilance, followed by relaxation once the immediate shock passes.

These three scenarios cover the main topologies of risk trajectories—event peaks, cumulative drifts, and localized shocks—ensuring that our tests reflect realistic dynamics practitioners encounter in heritage governance.

The interface exposes only these slider-level quantities. Scenario-specific gating constants are embedded in the presets and are not user-tunable. These correspond to coefficients that amplify or dampen particular reactions, e.g., policy boosts on $A\uparrow$ and $\Delta MR>0$, a threshold bonus when $MR$ crosses $M{R}_{\mathrm{high}}$, and recovery/giveback coefficients that reallocate influence toward the Expert and Community once $MR$ eases. Formally, these appear in the gate vector ${\gamma }_t$ of (5) as fixed entries, complementing the sensitivity matrix S and the trend-response vectors ${\theta }_{+}$ and ${\theta }_{-}$. The domain rationale is as follows: the Expert increases weight when B (loss per item) or C (extent) rises; the Community responds to A (frequency) and C; the Policymaker gains influence only on actionable signals ($A\uparrow$, $\Delta MR>0$) and receives a small bonus on crossings of $M{R}_{\mathrm{high}}$; and the AI Agent responds to trends in $\Delta MR$ and widening exposure ($A+C$). These constants are hard-coded in the implementation but remain conceptually transparent and interpretable.

To make “realistic” dynamics precise and reproducible, each scenario is required to pass three tests: a face validity test, a topology test, and a falsification stress test. All runs are deterministic under the presets in Table 1. The face validity test checks whether trajectories match domain expectations. In the Flooding scenario, as $MR$ rises to a clear peak, the framing should leave Cultural (entering Hybrid/Functional) at least once and return toward Cultural/Hybrid as the peak resolves. In Light Fading, with cumulative drift and high B, framing should remain Cultural or Hybrid for most of the run (Functional, if it appears, is episodic). In Theft/Vandalism, the shock window should trigger an early Functional run (policy activation), followed by de-escalation toward Hybrid/Cultural as vigilance eases. The topology test requires that the state timeline exhibit ordered regimes rather than noise: long plateaus are separated by few transitions, with Hybrid segments concentrated around $MR$ thresholds or just after peaks. The falsification stress tests check whether small, directed parameter changes degrade behavior in intelligible ways. Setting $\tau =0$ removes the Hybrid band, producing chattering at parity. Setting $\rho \to 0$ removes inertia, producing brittle jumps. Raising $M{R}_{\mathrm{high}}$ by $+1$ in the Flooding scenario delays policy activation and shortens the Functional plateau. These degradations confirm that observed behaviors are caused by the intended mechanisms. Under these tests, the three scenarios produce the expected phenomena: peak-driven re-framing and recovery (event), slow drift with cultural persistence (cumulative), and policy-led functional spikes that ease with time (shock). Crucially, all are obtained without stochastic forcing, parameter overfitting, or black-box learning. The benefit of EANO is that the ontology’s framing is endogenously negotiated by actor sensitivities to ABC dynamics, while the tie-band and resistance provide reflexive control that prevents runaway monopolies and enables recoveries.

Figures presented in subsequent sections were generated with the presets in Table 1, namely 80 steps (the notebook default) and a fixed random seed of 42. The strictly positive jitter ${\sigma }_{+}(A,B,C)$ is sampled independently from $\mathcal{U}(0,{\sigma }_{+})$ and added to the continuous ABC trajectories before half-step quantization to $\{1,1.5,\dots ,5\}$. Regime changes (Cultural/Hybrid/Functional) are deterministic given the scenario preset, seed, and steps. The EANO update uses continuous deltas $(\Delta A_c,\Delta B_c,\Delta C_c)$ to avoid quantization stalls; plots report the half-step $(A,B,C)$ and $MR=A+B+C$ as practitioners would record them.

The following fixed implementation constants are hard-coded across scenarios and not adjusted per run: adoption step $\eta =0.6$, entropy regularization $ϵ=0.02$, and low-risk relaxation $=0.05$ toward $w_0$ when $MR<M{R}_{\mathrm{low}}$. Scenario-specific elements (sensitivity matrix S, governance bands $(M{R}_{\mathrm{low}},M{R}_{\mathrm{high}})$, and gating vector ${\gamma }_t$) are embedded in the presets and are not user-tunable. Resistance $\rho$ and tie-band width $\tau$ are the slider-exposed values reported in Table 1. The trend-response vectors are fixed across scenarios as defined in (4).

4.4. Results

For each scenario we include the six standard plots: (P1) ABC + $MR$, (P2) actor weights, (P3) framing share with tie-band, (P4) state timeline and (P5) time-in-state bars, and (P6) normalized entropy of actor influence. Unless stated otherwise, runs use the presets in Table 1, namely 80 steps, seed $=42$, and strictly positive pre-quantization jitter $(0.05,0.02,0.04)$, to avoid quantization stalls while preserving determinism of regime changes.

Graphical results for Seasonal Flooding (event peak with partial easing) are presented in a set of graphs in Figure 1. In this figure,

• P1: ABC + $MR$. A (frequency) rises stepwise to a mid-run plateau, B (loss/item) remains mostly stable before a small end decline, and C (extent) increases stepwise until it levels off. The resulting $MR=A+B+C$ rises to a clear peak and then partially declines.
• P2: Weights. The Cultural bloc (Expert+Community) dominates early but declines gradually through the run. The Policymaker gains modestly via the $A\uparrow$ and $\Delta MR>0$ gates, while the AI Agent rises steadily and eventually converges with Policymaker, together balancing the Cultural bloc by the end.
• P3: Framing share. The Cultural share declines into the tie-band and stabilizes there, reflecting contested influence as $MR$ peaks and then eases.
• P4: State timeline. A long initial Cultural stretch is followed by a stable Hybrid plateau that persists until the end of the run. No Functional regime emerges under the preset parameters, and the transition occurs cleanly without oscillations.
• P5: Time-in-state counts. The aggregate distribution shows that Hybrid occupies the largest share of steps, followed by Cultural, while no Functional state occurs under the preset parameters. This statistical summary complements the timeline view by confirming the overall dominance pattern.
• P6: Entropy. Entropy rises steadily to nearly 1.00, indicating diversification of influence and the absence of single-actor dominance.
• Validation: T1 (gradual decline from Cultural into Hybrid, with stabilization in the tie-band), T2 (ordered, smooth trajectory without noisy oscillations), and T3 (raising $M{R}_{\mathrm{high}}$ shortens the Hybrid regime and delays Policymaker activation).

Graphical results for Light Fading (cumulative drift, high/stable B) are presented in Figure 2. In this figure,

• P1. A rises stepwise and then saturates, B remains stably high, and C increases gradually before flattening. The resulting $MR$ shows a smooth cumulative drift that plateaus without shocks or reversals.
• P2. The Expert is dominant at first (reflecting sensitivity to high B) but declines steadily over the run, remaining influential though no longer predominant by the end. The Community stays mid-level with only slight drift, while Policymaker and AI weights grow gradually; by the final steps, AI nearly matches Policymaker in influence.
• P3. The Cultural bloc share steadily decreases, remaining above the Hybrid band for most of the run and dipping into it only at the very end, reflecting gradual erosion of Cultural dominance without full transition to Hybrid.
• P4. The timeline shows a long uninterrupted Cultural plateau up to about step 65, followed by a Hybrid segment; no Functional state occurs under the preset parameters.
• P5. The aggregate distribution shows that Cultural framing dominates the run, with Hybrid accounting for a small minority of steps and no Functional state appearing under the preset parameters. This statistical summary complements the timeline by confirming the persistence of Cultural framing in cumulative hazards.
• P6. Entropy rises steadily to nearly 1.00, indicating a steady move toward full balance of influence and confirming the absence of single-actor lock-in.
• Validation: T1 (cultural persistence), T2 (coherent plateau), and T3 (lower $\rho$ or higher $\tau$ increases Hybrid time).

Graphical results for Theft/Vandalism (localized shock with vigilance easing) are presented in a set of graphs in Figure 3. In this figure,

• P1. A exhibits a brief mid-run spike and then returns to baseline; B remains stably high and C stays low. The resulting $MR$ shows a localized shock followed by full easing back to its initial level.
• P2. Policymaker and AI both rise into the mid-run shock, with Policymaker peaking briefly and AI showing a steady climb before easing slightly. Expert declines during the shock but regains some share as $MR$ falls, while Community remains comparatively stable with only minor fluctuation.
• P3. During the mid-run shock, the Cultural bloc share dips into the Hybrid band, producing a bounded Hybrid interval. As vigilance eases and $MR$ declines, the share rises back above the band and stabilizes in Cultural framing.
• P4. A single, clearly bounded Hybrid patch appears in mid-run, bracketed by Cultural segments before and after; no Functional regime occurs under the preset parameters.
• P5. The aggregate distribution shows Cultural framing dominating the run, with Hybrid confined to a short interval and no Functional state appearing under the preset parameters. This statistical view confirms that the shock produces only a temporary shift without long-term functional lock-in.
• P6. Entropy rises sharply during the shock, peaking near full balance as actor influence temporarily equalizes, then declines slightly to a stable plateau. This indicates that the shock induces a transient diversification of influence, after which the system settles into a plural but non-maximal distribution.
• Validation: T1 (shock to recovery), T2 (coherent plateau), and T3 (wider shock window lengthens the Hybrid patch; higher $M{R}_{\mathrm{high}}$ truncates it).

Taken together, the six plots show that the framing transitions observed across all three scenarios are endogenous, interpretable outcomes of the EANO mechanism. By grounding dynamics in the ABC method and applying transparent gating rules, EANO generates distinct but realistic trajectories for event, cumulative, and shock hazards. The tie-band and resistance provide reflexive control, preventing brittle oscillations and runaway dominance while enabling recovery when risks ease. The entropy trace further certifies that plurality of influence is preserved, with no single actor monopolizing control. In this way, EANO produces realistic, reproducible, and auditable patterns of framing evolution that reflect contested negotiations in heritage governance.

5. Discussion

The three minimal scenarios in Section 4 demonstrate how EANO enables a reflexive twin to surface and manage contested framings under domain-grounded risk dynamics. Because updates are deterministic functions of ABC deltas and $MR$ trends (with strictly positive, bounded pre-quantization jitter that never drives regime changes), the observed transitions are interpretable and reproducible.

In the case of Seasonal Flooding, a mid-run peak in $MR$ produces the expected re-framing: early Cultural dominance gives way to a sustained Hybrid interval near the peak, triggered by policy gating on $A\uparrow$ and $\Delta MR>0$. As severity eases, Cultural and Hybrid framing stabilize in balance. The state timeline shows a few long plateaus rather than oscillations, and entropy steadily rises, confirming diversification of influence without lock-in.

For Light Fading, high and stable B keeps Expert influence salient, while gradually rising C supports Community. Policymaker and AI activation remain mild in the absence of strong $A\uparrow$ or threshold crossings. The framing stays Cultural for most of the run, dipping into Hybrid only at the very end. Entropy grows steadily, reflecting gradual pluralization under cumulative pressures.

In the case of Theft/Vandalism, a short A spike deterministically produces a mid-run Hybrid interval via the policy gates and trend response. As vigilance eases and $MR$ declines, the system returns to Cultural framing. Entropy rises sharply during the shock—indicating a temporary diversification of influence—then settles to a stable plateau, reflecting recovery without long-term disruption.

Taken together, these patterns show how bloc shares (Expert+Community vs. Policymaker+AI) respond coherently to ABC mechanisms that practitioners recognize. The tie-band $\tau$ and resistance $\rho$ provide reflexive control that prevents runaway dominance and enables recovery, while directed parameter changes yield intelligible degradations (e.g., removing the tie-band produces chattering at parity; setting $\rho \to 0$ yields brittle jumps; raising $M{R}_{\mathrm{high}}$ delays or shortens policy-led Hybrid patches). These tests support our realism claims without resorting to stochastic forcing or opaque machine learning.

Beyond the specific scenarios, EANO offers a general mechanism for embedding reflexivity into cultural heritage digital twins. By grounding dynamics in the ABC method, it ensures that simulated framings remain auditable by practitioners rather than opaque outputs. By combining actor sensitivities, gating, resistance, and entropy regularization, it demonstrates how contested framings can emerge, stabilize, and recover under shifting risks. In this way, EANO provides an explainable simulation of governance dynamics: it shows how actors and blocs negotiate influence, why transitions occur when they do, and how recovery from shocks is structurally enabled. The aim is not to forecast site futures, but to provide a transparent laboratory for testing policies, thresholds, and actor interactions in a reproducible way.

Nonetheless, the current formulation has limitations. It omits actor memory and credibility hierarchies that are central to real governance processes. Future work will extend the model with provenance-weighted influence, temporal discounting, and learned gates. Empirical calibration on case studies will further test whether observed site negotiations align with the trajectories modeled by EANO.

In sum, EANO bridges formal risk models and reflexive governance simulation, offering an interpretable and reproducible path forward for cultural heritage digital twins.

6. Conclusions

This paper introduced evolving actor–network ontologies (EANO) as a novel framework for rethinking cultural digital twins. Unlike static ontologies that presuppose fixed categories, EANO embeds reflexivity and actor heterogeneity at the core of its design, enabling semantic interpretations to evolve as negotiated outcomes rather than predetermined classifications. The proposed architecture treats resistance, sensitivity, and redistribution not as afterthoughts but as structural features of the ontology itself, turning digital twins into adaptive instruments that mirror the contested and plural nature of heritage meaning.

To ground these ideas, we implemented a set of minimal deterministic simulations across three canonical scenarios—event peaks, cumulative drifts, and localized shocks—each drawn from recognizable heritage risk patterns. These runs were not intended as predictive tools, but as illustrative demonstrations of how the EANO mechanism generates realistic, reproducible, and interpretable trajectories. In this way, the simulations showcase the benefits of the framework without claiming empirical calibration.

By offering a transparent and auditable mechanism for modeling contested framings, EANO lays the foundation for digital twins that act not as static repositories of information but as evolving cognitive instruments. Future extensions will incorporate actor memory, credibility hierarchies, and adaptive learning, enabling even richer negotiations of meaning. In this sense, the paper is explicitly theoretical: it positions EANO as a foundation for explainable, adaptive simulation of governance dynamics, where cultural digital twins can engage with the dynamic realities of interpretation and policy mediation.