Separation of Justifications in Logic & Decision

• Separation of justifications is a formal method that distinguishes different classes of evidence to maintain consistency in logical, epistemic, and decision-theoretic systems.
• It ensures that contradictory claims cannot receive high-quality justification simultaneously, preserving semantic soundness in various formal models.
• The approach underpins models in rational choice theory, Bayesian testing, and epistemic frameworks, offering clear implications for empirical and theoretical research.

Separation of justifications refers to the formal, semantic, or operational distinction between different classes or sources of justification within a logical, epistemic, or decision-theoretic system. Separation arises both as a design principle in logic (avoiding collapse of epistemic, normative, or decision-theoretic justifications) and as a mathematical consistency condition (precluding contradictory high-quality justifications for both a claim and its negation). This concept features centrally in the behavioral foundations of rational choice under justifiable constraints, the formal modeling of epistemic belief/knowledge, the syntactic/semantic structure of modal and justification logics, and the foundations of Bayesian testing procedures.

1. Consistency and Separation in Justification Systems

In general justification theory, the separation property formalizes a critical form of consistency: for any statement $x$ and its complement $\bar x$, the best attainable quality of justification for $x$ cannot exceed that for $\bar x$, and vice versa. This is made precise in the setting of tree-like justification systems, which are formal structures encoding rules, facts, and justifications over a three-valued logic $\{T, F, U\}$.

Let $\mathcal F$ be a fact space with an involution $\bar{\cdot}$ and $\mathcal F^d$ the set of defined facts (closed under involution), and $\mathcal F^o$ the set of open facts. A justification frame is a triple $(\mathcal F^d, \mathcal F^o, R)$, where $R$ is a set of rules $x \leftarrow A$.

A tree-like justification is a finite or countably branching forest whose node labels correspond to facts, with root and child labels conforming to the rule set $R$. For any interpretation $I : \mathcal F \to \{T, F, U\}$, the supported value $supp(x,I)$ is defined as the maximal value (in the order $F < U < T$) attainable from all justifications for $x$, with each evaluated via the minimal value over its branches.

Main Consistency Result: If the rule frame is complementary (any "block" for $x$ induces a rule for $\bar{x}$, and vice versa) and the branch evaluation commutes with negation, then for every $x$ and every interpretation $I$,

$supp(x, I) = supp(\bar{x}, I)$

Thus, no system state can provide a better justification simultaneously for a statement and its negation. This enforces a separation of justifications, ensuring the mutual exclusion of high-quality explanations for contradictory claims. This property secures the semantic well-definedness of non-monotonic logics with justifications, especially in systems where nested or modular combinations occur. Extension of this property to graph-like justifications is more subtle but follows under additional constraints (Marynissen et al., 2022).

2. Epistemic and Normative Separation in Modal and Justification Logics

Formal justification logics introduce distinct syntactic and semantic types for justifications, particularly separating epistemic justifications (reasons for knowledge) from deontic justifications (reasons for obligation). In the temporal logic JTO, two disjoint sorts of justification terms are defined: $T\!m^E$ (epistemic) and $T\!m^D$ (deontic), with separate application and sum operations.

Formulas $t\!:\_E\,\varphi$ assert that epistemic term $t$ justifies $\varphi$ as knowledge, while $s\!:\_D\,\varphi$ asserts that deontic term $s$ justifies $\varphi$ as an obligation. The axiomatic structure and semantics keeps these two justification classes strictly disjoint:

• Term disjointness: Operators act within their own justification type; mixing is syntactically forbidden.
• Semantic separation: Evidence functions for epistemic and deontic reasons are mapped to different slots in the model and are governed by disjoint closure, introspection, and factivity requirements.
• No cross-lifting: There is no axiom or rule that directly translates epistemic into normative justifications or vice versa; their influence only propagates through propositional or temporal context.

This strict separation ensures that reasons for knowing and reasons for obligation are not conflated and that permission operators behave as proper duals of obligation within the deontic fragment only (Ghari, 2021).

3. Behavioral Characterizations and Decision-Theoretic Models

In rational choice theory, separation of justifications models agents who, while having a true underlying utility preference $u$, restrict their feasible choices to alternatives that are maximally ranked according to some accepted set of justifiable preferences $\mathcal J$. These justifiable preferences are themselves complete, transitive, and antisymmetric relations (often with utility representations $w$), and the true choice is only made among those alternatives that are maximal for some $w\in W$.

For any menu $A$,

$$M_J(A) = \bigcup_{w\in W} \operatorname{arg\,max}_{x\in A} w(x)$$

The DM then selects $u$-best among $M_J(A)$. This structure separates the "justification constraint" (external rationales, e.g., moral or legal) from "true preference" (internal). Two main behavioral axioms underpin this model:

• Optimization: DMs are utility maximizers among feasible options.
• Irrelevance of Unjustifiable Alternatives (IUA): Once an alternative is excluded under all justifications, its presence in larger menus is behaviorally inert.

This paradigm generalizes standard utility maximization (singleton $\mathcal J$), admits menu-dependent reversals (due to multiple justifications), and provides clean identification results. By requiring each justification to be a full preference, the model improves tractability, interpretation, and empirical identification compared to models with arbitrary, structureless rationales (Ridout, 2020).

4. Epistemic Modeling: Accepted vs. Knowledge-Producing Justifications

Justification Epistemic Models (JEMs) enforce a separation between accepted justifications (those actually endorsed by the agent) and knowledge-producing justifications (those that are factive in the model). In formal terms, a JEM is a triple $(\ast,\mathcal{A},\mathcal{E})$ where:

• $\mathcal{A} \subseteq \text{Tm}$: accepted justifications (properly closed under application),
• $\mathcal{E} \subseteq \text{Tm}$: knowledge-producing justifications (properly closed, factive).

Derived notions:

• A formula $F$ is believed iff $\exists\,t\in\mathcal{A}$, $\ast\models t\!:\!F$.
• Known iff there exists $t\in\mathcal{A}\cap\mathcal{E}$ with $\ast\models t\!:\!F$.

The distinction is crucial in capturing Gettier cases and similar epistemic scenarios: an agent may accept a (non-factive) justification and believe a true statement, yet lack knowledge because the justification is not knowledge-producing. Classical modal logics, lacking this fine-grained distinction, treat justifications only implicitly and cannot represent this separation (Artemov, 2017).

5. Separation in Bayesian Decision-Theoretic Justification

In Bayesian testing, decision-theoretic justification for the widely used “inverting credible sets” rule emerges from considering decision problems where directional conclusions are penalized asymmetrically. The optimal Bayes rule for choosing among $\{\theta < \theta_0\}$, $\theta = \theta_0$, and $\theta > \theta_0$ with specified losses separates the justification for each action in terms of posterior quantities.

The central result is that accepting $H_0$ is justified precisely if $\theta_0$ is inside the central $(1-\alpha)$ credible interval; this is not a mere ad-hoc convention but follows formally by minimizing expected loss under explicit modelled consequences. This conceptual separation—between traditional justification by ad-hoc inversion and rigorous decision-theoretic justification—clarifies the logical basis of Bayesian procedures and distinguishes them in terms of optimality and specific error consequences (Thulin, 2012).

6. Illustrative Examples and Implications

Non-monotonic Logic (Justification Systems): In tree-like justification systems, the separation property ensures that the assignment of truth values via justifications respects involutive consistency: a fact and its negation cannot be simultaneously justified "better" than each other for any interpretation.

Rational Choice with External Constraints: Empirical patterns violating classical consistency (e.g., menu-dependent reversals in social preference experiments) are explained by separation of justifiable from true preferences. For example, observed choice cycles in social experiments (such as the disabled-stranger scenario) directly align with the justifiability model but not with single-preference or arbitrary-rationale approaches (Ridout, 2020).

Epistemic Scenarios (Russell's Prime Minister example): The distinction between accepted and knowledge-producing justifications (JEM model) formalizes situations where agents have justified true belief (via accepted, but non-factive, justifications) yet lack knowledge. The rigorous separation at the model level encodes scenarios central to modern epistemology and the analysis of Gettier cases (Artemov, 2017).

Modal–Deontic Logic: In logics combining epistemic and normative justification modalities, strict syntactic and semantic separation prevents mixing of reasons for knowledge and reasons for obligation, ensuring model-theoretic clarity and supporting robust temporal and deontic reasoning (Ghari, 2021).

7. Comparative Perspective and Research Frontiers

Separation of justifications underpins and distinguishes various formal approaches:

• Models with required separation (tree-like justification systems, JTO logic, JEMs) offer consistency, interpretability, and identification advantages over approaches allowing arbitrary or collapsed justification structures.
• The principle is essential for tractability in behavioral identification, logical soundness in formal reasoning, and empirical alignment in decision experimentation.
• Current research investigates the extension of separation results to more general (graph-like) justification structures and the interaction with further modalities (temporal, probabilistic).

A plausible implication is that the explicit separation of justification classes, when sufficiently structured, undergirds the semantic soundness, empirical identifiability, and operational robustness of complex reasoning systems across logic, epistemology, and decision theory.