Deontic Cognitive Event Calculus (DCEC)

• DCEC is a sorted, quantified modal logic that integrates ethical principles with event calculus to model dynamic agent actions and world states.
• It formalizes doctrines such as the double and triple effect via precise constraints on intentions, outcomes, and ethical obligations.
• The framework employs advanced proof automation, including shadowing techniques, to verify ethical compliance in both new and existing AI systems.

The deontic cognitive event calculus (DCEC) is a sorted, first-order, quantified modal logic designed to formalize ethical principles within autonomous systems. DCEC uniquely integrates a multi-modal approach—including intensional attitudes such as knowledge, belief, intention, obligation, and permission—layered over event-calculus representations of agentive actions and dynamic world states. Its architecture is explicitly crafted to address requirements in ethical reasoning, notably for doctrines such as the doctrine of double effect (DDE), triple effect, akrasia, and aspects of virtue ethics. DCEC's design supports both direct embedding within new AI systems and the overlay of "DDE-compliance layers" that can verify ethical properties of existing black/gray-box systems, provided certain interfaces are exposed (Govindarajulu et al., 2017, Govindarajulu et al., 2019).

1. Logical Architecture and Syntax

DCEC employs a sorted signature comprising key sorts: Agent, Moment (Time), ActionType, Action, Event, Fluent, and Boolean. The functional and predicate vocabulary originates in the event calculus, covering foundational facts about world states (holds), events (happens), causation (initiates/terminates), and temporal relations (prior). Modal vocabulary is standardized, with operators for perceptual (P), epistemic (K), doxastic (B), intentional (I), desiderative (D), communal (C), communicative (S), and deontic (O, P) attitudes. Formulas combine these modalities with first-order quantifiers and standard logical connectives, enabling highly expressive representations for temporally indexed, agent-relative, and utility-sensitive reasoning (Govindarajulu et al., 2017, Govindarajulu et al., 2019).

2. Semantics and Inference Principles

DCEC utilizes standard Tarski-style semantics for its first-order fragment, grounded in a discrete-time event calculus. Modal semantics are either proof-theoretic (as in (Govindarajulu et al., 2017))—employing natural-deduction inference schemata—or possible-worlds/Kripke-style (as in (Govindarajulu et al., 2019)), with accessibility relations $R_{K_a}$, $R_{B_a}$, $R_{I_a}$, $R_{O_a}$ indexed by agent and attitude. Core inference principles allow propagation of beliefs, knowledge, intentions, and obligations across time and situations via rule schemas (R₁–R₁₄, see (Govindarajulu et al., 2017)). For automation, DCEC leverages “shadowing”—a process that replaces modal subformulas with propositional atoms, thereby permitting (when appropriate) use of first-order theorem provers without risk of unsound substitution of non-rigid designators in modal scopes (Govindarajulu et al., 2019).

3. Event Calculus Dynamics and Mental States

World evolution in DCEC is governed by event calculus dynamics: events initiate or terminate fluents (state-properties) at specified moments, with inertia managed via a “clipped” predicate. Agentive actions are modeled as instantiations of action types by agents at given times. Mental attitudes are primitive: for instance, $K(a,t,\varphi)$ for agent $a$ knowing $\varphi$ at time $t$, $B(a,t,\varphi)$ for believing, $I(a,t,\varphi)$ for intending. Modal inference rules dictate how perception leads to knowledge, how knowledge gives rise to beliefs and intentions, and how obligations can transform into goals. Deontic modalities encode obligations ($O$), permissions ($P$), and prohibitions, with explicit representation and inference mechanisms (Govindarajulu et al., 2017, Govindarajulu et al., 2019).

4. Formalization of Ethical Principles: Doctrine of Double and Triple Effect

DCEC provides an explicit, formal framework for the doctrine of double effect (DDE) and its extensions. The four classical DDE constraints (and an optional counterfactual fifth) are encoded as conjunctive conditions on actions, agent intentions, and outcome utilities:

• Non-forbidden: The action is not explicitly prohibited.
• Net Positivity: The cumulative utility of all initiated and terminated fluents by the action exceeds a threshold $\gamma$.
• Intention Constraints: Agent must intend only good effects (F₃ₐ) and not intend any bad effects (F₃_b).
• Means-vs-Side-Effect: No bad effect is used as a means to produce a good effect (F₄). Kamm’s triple effect is subsumed by selectively relaxing F₄ conditioned on the absence of intentionality regarding the harm. This formalism is extensible to handle agent indexicality, temporal progression, and utility aggregation. Specific operators are defined—for example, $\triangleright(holds(f,t_1), holds(g,t_2))$ formally distinguishes intentional means from side-effects (Govindarajulu et al., 2017).

5. Automated Reasoning and Proof Automation

Proof automation in DCEC is accomplished through modal-theorem-proving techniques tailored for sorted, quantified, multi-modal formulae. The shadowing-based algorithm in (Govindarajulu et al., 2019) alternates between first-order reasoning on propositional "shadow" formulas and explicit inference in the modal layer. Core rules include first-order resolution ($I_R$), belief propagation ($I_B$), and obligation-to-goal translation ($I_O$). This mechanism ensures soundness (blocking improper substitution into modal scopes) while retaining efficiency by leveraging mature first-order provers where possible. Modal resolution operates via indexed accessibility relations and context-sensitive proof patterns (Govindarajulu et al., 2019).

6. Applications: Compliance Layers and Scenario Analysis

DCEC supports two principal application modes:

• From-scratch construction: Ethical constraints are embedded directly into the action-selection architecture of autonomous systems.
• Verification/Compliance Layer: A DDE layer "glues" onto existing architectures (planners, DNNs, Bayesian nets, etc.) when minimal interfaces are provided; specifically, systems must expose intentions ($I(a,t,\varphi)$) and deontic constraints ($O(a,t,\dots)$). Illustrative simulations include variants of the trolley problem, with discrete-time analysis. Modal checks (F₁–F₄) and event calculus simulation have reported runtimes in the sub-second range on practical instances. DCEC can interface with STRIPS-style planners by mapping goals and ethical hierarchies to modal intentions and obligations; for POMDP/MDPs, intended goals and causal dependencies form the basis for assessing ethical compliance (Govindarajulu et al., 2017).

7. Open Problems and Limitations

Soundness and completeness of DCEC in full generality remain as open theoretical problems. The current framework does not natively accommodate probabilistic uncertainty, counterfactual reasoning (full formalization of F₅), nor explanation generation beyond proof traces. Black-box systems that do not expose intentions to the compliance layer cannot be verified; a gray-box architecture is essential. The permission operator’s formal inference rules are under development, and learning utility functions ($\mu$) with complex dependencies may be intractable unless provided or learned offline (Govindarajulu et al., 2017, Govindarajulu et al., 2019). A plausible implication is that future research on probabilistic DCEC, integration of subjunctive logic, and broader empirical benchmarks is anticipated.

---

In summary, the deontic cognitive event calculus offers a rigorously formal, multi-modal, time- and agent-indexed framework for ethical reasoning in autonomous systems. Its event calculus core, richly sorted modal architecture, and proof automation strategies position it as a foundational tool for both the design and verification of ethically sensitive AI applications (Govindarajulu et al., 2017, Govindarajulu et al., 2019).