Justified Models: Foundations & Applications

• Justified models are conceptual frameworks in which beliefs, predictions, or choices are accepted only when supported by explicit justification, ensuring an onto-relational fit between theory and evidence.
• In logic programming and machine learning, these models employ causal and support graphs or calibrated evidence to provide consistent explanations and robust performance under uncertainty.
• Across disciplines like epistemic planning, collective choice, and applied analysis, justified models systematically filter decisions through validated reductions and justification sets, enhancing theoretical soundness and practical reliability.

“Justified models” is not a single formalism but a recurring research motif: a model, belief state, prediction, committee, or reduced equation is accepted only when it is constrained by an explicit notion of justification. In the literature surveyed here, this motif appears in epistemological realism and justification logic, in causal and support-based semantics for logic programs, in machine learning systems that tie prediction quality to evidence and calibrated reliance, in epistemic planning via justified perspectives, and in domain-specific notions of justified choice, justified representation, and fully justified asymptotic reduction (Andrews, 2012, Cabalar et al., 2014, Akula et al., 2019, Li et al., 2024).

Domain	What is justified	Primary mechanism
Epistemology and justification logic	Belief, knowledge, necessity	Onto-relations, explicit justification terms, accepted vs knowledge-producing reasons
Logic programming	True atoms or whole models	Causal graphs, support graphs, stable-model refinements
Machine learning and XAI	Predictions, trust, reliance	Neighborhood support, dialog explanations, evidence retrieval, faithful subspace explanations
Planning and collective choice	Nested beliefs, committees, choices	Justified perspectives, representation axioms, justifiable feasible sets
Applied analysis	Reduced equations or training objectives	Consistency, well-posedness, convergence, admissible weighted losses

1. Epistemic and justificatory foundations

A central philosophical lineage treats justification as the condition that links belief to truth. Under epistemological-scientific realism, knowledge is framed by the familiar schema

$K(p) \leftrightarrow B(p) \land J(p) \land T(p),$

but justification is not autonomous: it is constrained by an onto-relationship between reality and belief, so that epistemology “follows from” ontology. On this view, inferential and non-inferential justification are both legitimate only insofar as they are reality-tracking, and the paper implies a hybrid architecture of foundational cores plus coherent inferential networks, each checked against onto-relational fit (Andrews, 2012).

A related line in non-Fregean logic makes justification primitive and defines necessity from it:

$$\square \varphi \leftrightarrow \exists u\,(\varphi:u).$$

Here justifications are treated extensionally as sets of true propositions, and necessity is the union of all evidence sets. The resulting semantics is algebraic rather than relational, yet can capture modal principles corresponding to S4 and S5 when the corresponding constraints are added (Lewitzka, 2012).

Justification Epistemic Models sharpen the epistemic role of reasons by distinguishing accepted justifications from knowledge-producing justifications. Belief is derived from the former, knowledge from their intersection. This makes it possible to represent Russell-style cases in which a proposition is true and believed for a wrong reason, but not known (Artemov, 2017). At the semantic level, basic justification models further isolate two structural properties. Sharpness requires that application coincide exactly with modus-ponens composition,

$(s \cdot t)^* = s^* \rhd t^*,$

while injectivity requires each justification term to support at most one formula. The referential logic Jref axiomatizes these properties and is sound, complete, and decidable for the class of sharp injective justification models (Krupski, 2017).

2. Logic-programming semantics and justified truth assignments

In logic programming, a justified model is a truth assignment annotated with explicit derivational structure. The causal-graph approach associates each true atom with a causal value built algebraically from rule labels using addition for alternative causes, product for joint causation, and non-commutative application for rule application:

$$U+U' = U \cup U', \qquad U*U' = U \cap U', \qquad U \cdot U' = \{\,G \cdot G' \mid G \in U,\ G' \in U'\,\}.$$

For positive programs, the least causal model is obtained semantically, yet its maximal causal graphs correspond exactly to the non-redundant syntactic proofs of the atom (Cabalar et al., 2014).

This semantics extends to disjunctive programs and to causal-choice rules. The extension preserves the classical stable-model projection: each standard stable model corresponds to a disjoint class of causal stable models sharing the same truth assignment while differing in their explanations. Causal-choice rules are introduced precisely because ordinary choice rules conflate “may cause” with “does cause” once the head becomes true by any route (Cabalar et al., 2016).

A different but closely related notion uses support graphs. A classical model is justified iff it admits at least one acyclic support graph explaining all of its true atoms. This yields two sharp structural facts:

$$SM(P) \subseteq JM(P), \qquad P\ \text{non-disjunctive} \Rightarrow SM(P)=JM(P).$$

Thus every stable model is justified, but justified need not imply stable for disjunctive programs (Cabalar et al., 2023). At a more abstract level, tree-like justification theory proves that if the rule frame is complementary and the branch evaluation respects negation, then support is automatically consistent: the best supported value of a fact is always the complement of the best supported value of its negation. This resolves the consistency problem for tree-like systems and guarantees that justified interpretations cannot simultaneously support a fact and its complement (Marynissen et al., 2022).

3. Machine learning: justification as calibrated support, evidence, and trust

In machine learning, justified models typically do not mean transparent models in the classical symbolic sense. Instead, they are predictors whose outputs are constrained by support, evidence, or calibrated interaction. One line extends justified true belief to supervised classification. A classifier’s prediction is treated as belief, and its justification is constructed from neighborhood support in both input and latent spaces. The resulting output is triaged into IK (“I know”), IMK (“I may know”), and IDK (“I don’t know”), exposing epistemic uncertainty rather than merely reporting softmax confidence. Under adversarial noise on MNIST, the baseline had $F_{IK}=0.851$ and $A_{IK}=0.026$, whereas the $\epsilon$-NN justified model had $F_{IK}=0.000$ and $A_{IK}=0.000$; under large uniform noise on Grid Stability, the baseline had $$\square \varphi \leftrightarrow \exists u\,(\varphi:u).$$0 and $$\square \varphi \leftrightarrow \exists u\,(\varphi:u).$$1, while the $$\square \varphi \leftrightarrow \exists u\,(\varphi:u).$$2-NN justified model had $$\square \varphi \leftrightarrow \exists u\,(\varphi:u).$$3 and $$\square \varphi \leftrightarrow \exists u\,(\varphi:u).$$4 (Virani et al., 2019).

A second line treats justification as an interaction protocol. X-ToM models explanation as dialog, explicitly tracking the human’s intention, the human’s model of the machine, and the machine’s model of the human. In a study with 120 human subjects, X-ToM explanations were rated significantly higher than QA and Saliency Maps on usefulness, sufficiency, and appropriated detail ($$\square \varphi \leftrightarrow \exists u\,(\varphi:u).$$5), with no significant difference in response time (Akula et al., 2019). A fairness-oriented variant defines a justified summary as one that is sufficient for predicting the outcome and independent of protected attributes. FairSum uses a multi-task text model, Integrated Gradients, and the inclusion score

$$\square \varphi \leftrightarrow \exists u\,(\varphi:u).$$6

to retain high-utility, low-leakage sentences. On the Chicago food inspection dataset, utility stayed essentially unchanged (Micro-F1 $$\square \varphi \leftrightarrow \exists u\,(\varphi:u).$$7) while demographic leakage fell from Micro-F1 $$\square \varphi \leftrightarrow \exists u\,(\varphi:u).$$8 to $$\square \varphi \leftrightarrow \exists u\,(\varphi:u).$$9 and Macro-F1 $(s \cdot t)^* = s^* \rhd t^*,$0 to $(s \cdot t)^* = s^* \rhd t^*,$1 (Keymanesh et al., 2021).

More recent work separates interpretability from transparency. FaithfulDefense constructs explanations for logical models that are completely faithful for an entire positive subspace around a query point while revealing as little of the decision boundary as possible. The selection problem reduces to maximum coverage, so greedy selection inherits the $(s \cdot t)^* = s^* \rhd t^*,$2 approximation guarantee. Empirically, FaithfulDefense achieved FPR $(s \cdot t)^* = s^* \rhd t^*,$3 on FICO, German, and Loan, whereas LIME yielded FPR $(s \cdot t)^* = s^* \rhd t^*,$4, $(s \cdot t)^* = s^* \rhd t^*,$5, and $(s \cdot t)^* = s^* \rhd t^*,$6 respectively (Zhong et al., 26 Feb 2025). In legal NLP, this evidential conception becomes explicit: Justifiable Artificial Intelligence shifts emphasis from explaining internals to providing evidence from trustworthy sources for and against a claim, with retrieval, entailment classification, provenance, and human validation as the core workflow (Wehnert, 2023).

4. Justified perspectives in epistemic planning

In epistemic planning, justified models are state-based devices for computing what agents are justified to believe from their observation history. The Justified Perspective model defines belief over a state sequence $(s \cdot t)^* = s^* \rhd t^*,$7 by

$(s \cdot t)^* = s^* \rhd t^*,$8

where $(s \cdot t)^* = s^* \rhd t^*,$9 reconstructs agent $$U+U' = U \cup U', \qquad U*U' = U \cap U', \qquad U \cdot U' = \{\,G \cdot G' \mid G \in U,\ G' \in U'\,\}.$$0’s perspective from observations and retrieval. The semantics is ternary—true, false, unknown—and supports arbitrary nesting by functional composition rather than explicit Kripke-model expansion (Li et al., 2024).

The Predictive Justified Perspective model removes JP’s static-environment assumption. Variables become processual,

$$U+U' = U \cup U', \qquad U*U' = U \cap U', \qquad U \cdot U' = \{\,G \cdot G' \mid G \in U,\ G' \in U'\,\}.$$1

and belief reconstruction uses predictive retrieval functions $$U+U' = U \cup U', \qquad U*U' = U \cap U', \qquad U \cdot U' = \{\,G \cdot G' \mid G \in U,\ G' \in U'\,\}.$$2 specialized to the variable’s changing type. If the observation history is too sparse, PJP defaults to the static case; otherwise it predicts current values from past observations. This is the key move that allows justified belief to remain valid when variables evolve even without direct action or observation (Li et al., 2024).

The empirical effect is substantial. In Big Brother Logic instance B3, PJP solved the problem with $$U+U' = U \cup U', \qquad U*U' = U \cap U', \qquad U \cdot U' = \{\,G \cdot G' \mid G \in U,\ G' \in U'\,\}.$$3, $$U+U' = U \cup U', \qquad U*U' = U \cap U', \qquad U \cdot U' = \{\,G \cdot G' \mid G \in U,\ G' \in U'\,\}.$$4, and $$U+U' = U \cup U', \qquad U*U' = U \cap U', \qquad U \cdot U' = \{\,G \cdot G' \mid G \in U,\ G' \in U'\,\}.$$5, whereas JP timed out after $$U+U' = U \cup U', \qquad U*U' = U \cap U', \qquad U \cdot U' = \{\,G \cdot G' \mid G \in U,\ G' \in U'\,\}.$$6 expanded nodes. In the Number domain, PJP solved the second-order-polynomial instance N8 with $$U+U' = U \cup U', \qquad U*U' = U \cap U', \qquad U \cdot U' = \{\,G \cdot G' \mid G \in U,\ G' \in U'\,\}.$$7 and $$U+U' = U \cup U', \qquad U*U' = U \cap U', \qquad U \cdot U' = \{\,G \cdot G' \mid G \in U,\ G' \in U'\,\}.$$8, while JP failed. In Grapevine instance G2, PJP found a plan with $$U+U' = U \cup U', \qquad U*U' = U \cap U', \qquad U \cdot U' = \{\,G \cdot G' \mid G \in U,\ G' \in U'\,\}.$$9 and $$SM(P) \subseteq JM(P), \qquad P\ \text{non-disjunctive} \Rightarrow SM(P)=JM(P).$$0, whereas JP returned False (Li et al., 2024). The framework is therefore justified in a double sense: beliefs are justified by perspective-sensitive evidence, and the planning semantics remains operational in dynamic multi-agent environments.

5. Collective choice and constrained decision-making

Outside logic and AI explanation, justified models also appear in social choice and behavioral decision theory. In Sub-Committee Voting, justified representation is generalized through Intra-wise JR (IW-JR), Span-wise JR (SW-JR), and weak-SW-JR. SW-JR is the global analogue of classical justified representation, but it may fail to exist even with two sub-committees of quota $$SM(P) \subseteq JM(P), \qquad P\ \text{non-disjunctive} \Rightarrow SM(P)=JM(P).$$1. By contrast, IW-JR and weak-SW-JR always admit a committee, and the paper gives a polynomial-time constructive algorithm that simultaneously guarantees both. The existence problem for an SW-JR committee is NP-complete (Aziz et al., 2017).

In decision theory, the justifiability model formalizes a decision maker who has a true preference but chooses only among outcomes top-ranked by some admissible justification. For a menu $$SM(P) \subseteq JM(P), \qquad P\ \text{non-disjunctive} \Rightarrow SM(P)=JM(P).$$2,

$$SM(P) \subseteq JM(P), \qquad P\ \text{non-disjunctive} \Rightarrow SM(P)=JM(P).$$3

This produces a lexicographic structure: first remove unjustifiable options, then maximize true utility over the surviving set. In the general model, representability is characterized by Optimization and Irrelevance of Unjustifiable Alternatives; in the expected-utility special case, Independence, Continuity, Monotonicity, and Convexity are added, and the set of admissible justifications becomes identifiable from choice behavior under the stated conditions (Ridout, 2020).

These literatures give “justified model” a normative rather than explanatory role. The object being justified is not a proposition or prediction but a collective decision or individual choice rule. Even so, the same structural idea persists: admissibility is mediated by an explicit justification set rather than by preference, support, or feasibility alone.

6. Fully justified reductions and theoretically justified training objectives

In applied analysis, “justified model” often means a reduced model that is not merely formally derived but rigorously validated. For one-dimensional internal waves in the Camassa–Holm regime, a new Green–Naghdi-type system is called fully justified because it is consistent, well-posed, and close to the full Euler system on the relevant time scale. The consistency theorem yields the residual bound

$$SM(P) \subseteq JM(P), \qquad P\ \text{non-disjunctive} \Rightarrow SM(P)=JM(P).$$4

and the convergence theorem gives

$$SM(P) \subseteq JM(P), \qquad P\ \text{non-disjunctive} \Rightarrow SM(P)=JM(P).$$5

The same framework also justifies lower-order approximations, including the Constantin–Lannes model (Duchene et al., 2013).

A related but distinct usage appears in generative modeling. In Generator Matching, Flow Matching, Diffusion Models, and Edit Flows, time-dependent loss weighting is theoretically justified because both the Bregman loss and the linear parameterization of the generator may depend on time and state:

$$SM(P) \subseteq JM(P), \qquad P\ \text{non-disjunctive} \Rightarrow SM(P)=JM(P).$$6

If $$SM(P) \subseteq JM(P), \qquad P\ \text{non-disjunctive} \Rightarrow SM(P)=JM(P).$$7 and the weight is positive almost everywhere, reweighting can be absorbed into a change of time law without changing minimizers. The note’s point is not that a particular model is justified by explanation, but that a class of training heuristics already used in practice is justified by the underlying theory (Billera et al., 20 Nov 2025).

Across these literatures, a recurring pattern is that justification is not a decorative explanation layered onto an independently defined model. It is the condition that determines which beliefs count as knowledge, which truth assignments count as stable or supported, which predictions warrant reliance, which plans encode defensible higher-order belief, which committees count as representative, which choices are admissible, and which reduced dynamics or losses count as theoretically sound. This suggests that “justified models” are best understood as models whose semantics or admissibility conditions are indexed by explicit supporting structures rather than by outputs alone.