---
title: 'TAXAL: Triadic Explainability Framework'
url: https://www.emergentmind.com/topics/taxal
type: topic
---

# TAXAL: Triadic Explainability Framework

TAXAL, short for **Triadic Alignment for eXplainability in Agentic LLMs**, is a conceptual and practical framework for designing, evaluating, and deploying explanations for large language models, particularly when they operate as agentic systems in high-risk sociotechnical domains such as law, healthcare, education, public services, and safety systems. Its central claim is that explainability is not a unitary property but a **triadic fusion** of three partially conflicting dimensions—**cognitive**, **functional**, and **causal**—that must be jointly considered and aligned with stakeholder roles and domain constraints. In this formulation, explanation is treated simultaneously as communication, workflow infrastructure, and faithful reflection of model behavior, rather than as a post-hoc narrative attached to outputs [2509.05199].

## 1. Definition and conceptual basis

TAXAL defines explainability for agentic LLMs through three complementary axes: **cognitive** concerns how explanations are understood by humans, **functional** concerns how explanations are used in practice, and **causal** concerns how faithfully explanations reflect the model’s actual reasoning or decision pathways. The framework is explicitly motivated by the claim that traditional explainability methods, when focused on surface outputs, do not adequately capture the reasoning pathways, planning logic, and systemic impacts of agentic LLMs [2509.05199].

The term **triadic fusion** denotes more than classification. Explainability goals are first decomposed into the three axes, but TAXAL requires that they then be **simultaneously aligned** in a role-sensitive manner. Individual methods such as chain-of-thought prompting, SHAP, dialogic interfaces, attribution-based debugging, or counterfactuals typically optimize one or two dimensions rather than all three. TAXAL therefore recommends **composite** explanation strategies that combine multiple techniques, and **layered** explanation interfaces that present different views to different stakeholders.

This orientation distinguishes TAXAL from single-focus XAI paradigms. Feature-attribution approaches may offer comparatively strong causal grounding while remaining weak on user comprehension or institutional usability. User-centric rationales may be cognitively effective while lacking faithfulness to the internal computation. Formal causal methods may be strong on auditability but costly or impractical in deployed workflows. TAXAL treats these tensions as a design problem rather than an anomaly.

The framework is also specific to **agentic LLMs** rather than prompt-response systems alone. In TAXAL’s scope, agentic LLMs are systems that plan, decompose tasks into sub-goals, call tools and external APIs, maintain internal state over time, and act autonomously or semi-autonomously within multi-step workflows. This broadens the target of explanation from isolated predictions to plans, action sequences, tool choices, temporal state, and system behavior over time.

## 2. The triadic dimensions

The three dimensions have distinct explanatory objects, operational mechanisms, and evaluation criteria.

| Dimension | Central concern | Representative methods |
|---|---|---|
| Cognitive | How explanations are understood by humans | Natural language rationales, dialogic XAI, progressive disclosure |
| Functional | How explanations are used in practice | Hallucination detection, explanation-augmented supervision, LLM-as-evaluator |
| Causal | How explanations faithfully reflect model behavior | SHAP, LIME, attention rollout, causal tracing, counterfactual reasoning |

The **cognitive dimension** addresses plausibility, clarity, mental-model alignment, and trust calibration. It asks whether an explanation is comprehensible to its intended audience, whether it matches users’ conceptual frameworks and capacities, and whether it helps users form accurate beliefs about system behavior and limits. Operational forms include natural-language rationales, conversational explanation interfaces, personalization by expertise level, and progressive disclosure in which a simple default view is supplemented by optional detail. TAXAL associates cognitive evaluation with plausibility, Human Reasoning Agreement (HRA), comprehension tests, and trust-related measurements.

The **functional dimension** treats explanations as instruments embedded in technical and institutional workflows. Here explanations are expected to support diagnosis, debugging, quality assurance, training, governance, safety, and task performance. Functional forms include attribution-guided debugging, hallucination detection, explanation-aware safety filters, explanation-augmented supervision, and LLMs used as evaluators of other models’ outputs or explanations. This dimension is evaluated through task-based efficacy, decision impact, robustness under perturbation, and consistency across runs.

The **causal dimension** concerns faithfulness to the model’s actual computation and sensitivity to interventions. It includes token-level attribution, perturbation and ablation, causal tracing through layers or attention heads, counterfactual explanations, and ante-hoc interpretable architectures such as concept bottleneck models, neural additive models, and self-rationalizing architectures. In this dimension, explanations are valuable because they support auditability, fairness analysis, contestability, and compliance in regulated settings.

TAXAL emphasizes that the dimensions are neither reducible to one another nor fully harmonizable. A highly plausible explanation can be unfaithful. A technically faithful explanation can be too complex for affected users. A diagnostic explanation useful to developers can be functionally excellent yet cognitively unsuitable for laypersons. These conflicts are constitutive of the framework rather than edge cases.

## 3. Alignment mechanics and method profiling

TAXAL operationalizes triadic fusion through a stepwise workflow. First, the relevant stakeholder role is identified, such as developer, regulator, domain expert, auditor, or end user. Second, the most salient explanatory dimension or dimensions are selected for that role. Third, explanation methods are mapped to those needs. Fourth, trade-offs among the dimensions are explicitly balanced. Fifth, explanations are iterated in real context using domain-specific feedback and metrics [2509.05199].

This process is formalized through a triadic alignment profile for each explanation method:
$$
A(m) = \big(a_{\text{cog}}(m),\ a_{\text{func}}(m),\ a_{\text{caus}}(m)\big).
$$
In TAXAL’s formulation, each component may be expressed qualitatively as High, Medium, or Low, or more quantitatively when metrics are available. The design problem is then to select a set of methods appropriate to a stakeholder role and domain rather than to search for a universally optimal explanation form.

The framework’s illustrative scoring matrix shows how common methods occupy different positions in this triadic space. **Chain-of-Thought** is characterized as high on cognitive and functional dimensions but medium on causal alignment, because explicit reasoning steps may improve transparency and task performance without guaranteeing faithfulness to internal causal processes. **Token attribution methods** such as SHAP and LIME are characterized as medium on cognitive alignment, high on functional utility, and high on causal alignment, reflecting their value for debugging and tracing influence while also requiring expertise to interpret. **Interactive dialogic XAI** is high on cognitive and functional axes and medium on causal grounding, since its explanatory value depends partly on the underlying attribution or tracing substrate [2509.05199].

This method profiling makes TAXAL a **meta-taxonomy** as well as a design framework. Existing and future XAI methods can be located within the triad, not only in terms of what they explain but also in terms of which stakeholders they serve, which institutional requirements they satisfy, and what trade-offs they impose. A plausible implication is that TAXAL’s main novelty lies less in any single explanatory algorithm than in its insistence that explanation design must be multidimensional, role-sensitive, and pipeline-oriented.

## 4. Agentic LLMs and role-sensitive applications

TAXAL is framed around the explanatory challenges of agentic LLMs. Because such systems plan, select tools, maintain temporal state, and interact with multiple actors, explanation must account for **action sequences**, **subgoal decomposition**, **tool invocation**, and **behavior over time**, not merely single textual outputs. TAXAL therefore extends explanation targets from “why this answer” to “why this plan,” “why this action sequence,” and “why this tool was invoked under these conditions” [2509.05199].

The framework’s case studies illustrate how explanation strategies vary by domain and stakeholder.

In **law**, the example is a legal document review assistant that flags risky clauses and suggests edits. Junior associates require cognitively accessible summaries of legal language, whereas internal lawyers and auditors require higher causal fidelity, detailed contractual reasoning, and counterfactuals such as the effect of removing an indemnity clause. Clients, by contrast, receive simplified risk summaries with less causal detail.

In **education**, the example is an intelligent tutoring system. Students benefit from tailored explanations and progressive hints, emphasizing cognitive clarity and encouragement. Teachers require functional oversight and moderate causal insight into assessment logic. Here TAXAL treats explanation as both pedagogical support and workflow infrastructure.

In **public services**, the example is an eligibility chatbot. Citizens need simple, understandable reasons such as an income threshold explanation, alongside contestability mechanisms. Auditors and policymakers require deeper causal paths, logs, and counterfactual trails for oversight. This case places cognitive clarity and fairness concerns alongside stronger causal requirements for institutional review.

In **HR**, the example is an AI-assisted hiring platform. Recruiters receive functional summaries of candidate strengths and weaknesses to accelerate screening and document decisions. Internal compliance roles require causal traces and fairness checks. Candidates receive accessible reasons or advice with constrained causal detail, reflecting the domain’s legal sensitivities.

In **mental health screening**, the framework describes clinician-facing rationales such as references to loss of interest, hopelessness, or mention of self-harm, alongside contrastive explanations between flagged and non-flagged entries. Clinicians require higher causal insight into changes in risk, whereas patient-facing feedback should remain simplified and care-oriented rather than diagnostic.

In **safety and jailbreak detection**, TAXAL recommends strong functional and causal detail for internal security teams, including traces to token patterns and perturbations, while limiting external explanations to policy-based statements so as not to leak detection logic. This is one of the clearest illustrations of TAXAL’s claim that the same system should expose different explanation slices to different stakeholders.

Across these cases, the framework repeatedly shows that explanation is not only user-relative but institution-relative. Law and HR emphasize counterfactuals and auditability; education emphasizes cognitive and functional support; public services and healthcare combine intelligibility with regulatory causal requirements; jailbreak defense prioritizes internal functional and causal visibility while minimizing external disclosure.

## 5. Design patterns, methodology, and evaluation

TAXAL proposes several recurring design patterns. **Layered explanation interfaces** present simple rationales to end users while exposing deeper technical and causal layers to experts and auditors. **Explanation role routing** adapts granularity and emphasis according to who is asking. **Counterfactual feedback modules** embed what-if tools in domains such as hiring, eligibility, and medical risk assessment. **Explanation-aware prompt templates** ask the LLM to produce user-level rationales, procedural steps, and contrastive or counterfactual perspectives within a single explanation pipeline [2509.05199].

The methodological pipeline begins with **context and stakeholder analysis**, including domain, risk profile, stakeholders, and regulatory constraints. It proceeds to **dimension profiling**, in which each stakeholder is associated with weights over cognitive, functional, and causal priorities. It then performs **technique mapping** via triadic method profiles, followed by **architecture and interface design** that combines methods such as chain-of-thought, rationales, SHAP, and counterfactuals into role-specific views. The process concludes with **evaluation and iteration** in real workflows.

Evaluation in TAXAL is explicitly plural rather than unified by a single scalar score. For the **cognitive** axis, the framework highlights plausibility, comprehension, trust calibration, and HRA, conceptually represented as
$$
\text{HRA}(E,R) \approx \text{similarity}(E,R),
$$
where \(E\) is the explanation and \(R\) is a human rationale. For the **functional** axis, it emphasizes task-based efficacy, time to decision, confidence shifts, error-rate changes, robustness under small perturbations, and consistency across runs. For the **causal** axis, it emphasizes faithfulness, attribution sensitivity, causal tracing, counterfactual correctness, and contrastivity.

Counterfactual explanation is represented in TAXAL through a standard constrained optimization form:
$$
x' = \arg\min_{x'} \ \text{dist}(x,x') \quad \text{s.t.} \quad f(x') \neq f(x).
$$
This expresses the search for a minimally changed input that crosses the decision boundary and supports recourse or fairness analysis. TAXAL places particular emphasis on such counterfactuals in domains where contestability and accountability are central.

The framework does not propose a single global explainability metric. Instead, it recommends **triangulating** metrics and interpreting them relative to stakeholder role and domain. This reflects one of TAXAL’s core commitments: explanation quality is inseparable from explanatory purpose.

## 6. Theoretical significance, limitations, and open problems

TAXAL’s theoretical contribution is to recast explainability as a **triadic sociotechnical construct**. In this formulation, the cognitive dimension corresponds to explanation as communication, the functional dimension to explanation as infrastructure or tooling, and the causal dimension to explanation as faithful reflection for accountability. The framework explicitly links these dimensions to broader capacities of **trust**, **contestability**, and **accountability**, and situates explainability within governance and institutional practice rather than treating it as a purely technical property [2509.05199].

This positioning yields several normative recommendations. Systems should implement **role-aware explanation design by default**. Documentation should specify which TAXAL dimensions are being targeted for which stakeholders. Public and enterprise procurement processes should incorporate triadic explainability criteria. Participatory evaluation should include not only technical experts but also affected stakeholders and laypeople when fairness and usability are at issue.

The framework is also explicit about its limitations. It is presently **conceptual** and requires broader empirical validation across domains and tasks. Its mappings from techniques to dimensions are largely **qualitative**, creating a need for standardized scoring rubrics and quantitative alignment measures. The paper notes **empirical evaluation gaps**, including the absence of large-scale quantitative comparisons and the need for benchmark-based studies. It also recognizes that evolving modalities—multimodal systems, multi-agent systems, and tool-using agents—may require refinement or extension of the triad.

A central unresolved issue is the **faithfulness versus plausibility trade-off**. Technically faithful explanations may be too complex for many users, while simple narratives may be cognitively effective but misleading. TAXAL identifies this as an open design and evaluation problem rather than claiming to solve it. Additional open problems include semantically complex or long-context tasks, temporal causal chains in agentic plans, multi-agent coordination, and the evolution of goals over time.

Future work proposed within the framework includes building TAXAL-based toolkits and deployment pipelines, conducting user studies that directly measure triadic alignment, and extending the framework to temporal, systemic, and multi-agent explanation structures. This suggests that TAXAL is best understood not as a closed theory of explainability but as a structured research program for agentic LLM explanation in high-risk sociotechnical settings.

Source: https://www.emergentmind.com/topics/taxal