Organizations: Department of Statistical and Actuarial Sciences, Western University, London, Canada · Department of Computer Science, Western University, London, Canada · University of Southampton Business School, University of Southampton, Southampton, United Kingdom · Centre for Operational Research, Management Sciences and Information Systems, University of Southampton, Southampton, United Kingdom · School of Mathematical Sciences, University of Southampton, Southampton, United Kingdom · Department of Computer Science, Reykjavik University, Reykjavik, Iceland
Credit decisioning is a high-stakes task in which model outputs must be accurate and explainable to support compliant decisions. Although modern credit risk models such as eXtreme Gradient Boosting (XGBoost) and Graph Neural Networks (GNNs) improve predictive performance, their explanations are often too technical for stakeholders creating communication gaps that can shape approvals, denials, and fairness judgments. We examine whether Large Language Models (LLMs) can serve as explanation layers that translate post-hoc explanation artefacts into stakeholder-appropriate risk narratives. Using Freddie Mac single-family loan-level data, we develop three pipelines: standard tabular (XGBoost + SHAP), and two with alternative data, a pure network-based (GNN + GNNExplainer), and a bimodal one (combining tabular and network data). We generate narratives with three LLM configurations: a small fine-tuned LLM (Gemma 3 4B), a large fine-tuned LLM (DeepSeek R1 70B), and a zero-shot commercial LLM (Gemini 2.5). Explanation quality is evaluated through automated checks across all pipelines and a human study of bimodal explanations comparing credit risk professionals and non-professionals on eight decision-relevant dimensions. We have three main findings. First, the pipeline accounts for higher variance in evidence-grounding scores than the language model, meaning that the binding constraint on explanation quality is the evidence representation, not the model used. Second, the explanation narratives reliably name the influential factors but are less reliable when stating the direction of influence, which may be consequential for adverse-action communication. Finally, professionals apply stricter evidentiary standards than non-professionals. We discuss implications for the governance of risk models, including deployment considerations and the value of domain-aligned LLMs in regulated credit settings.
Figures & tables
LLM
Size
Fine-tuned
Deployment Mode
Source
Gemma 3
Small (4B)
Yes
Open-source
Google DeepMind
DeepSeek R1
Large (70B)
Yes
Open-source
DeepSeek AI
Gemini 2.5
Undisclosed
No
API-based
Google
Table 1: LLMs used for narrative generation.
Figure 1: The architecture of the three explanation pipelines.
Pipeline
Prompt Components
Tabular- based
Predicted outcome; top- k SHAP feature attributions with descriptions; instructional step-by-step reasoning; guided subgoal structure; borrower-level counterfactual prompt; output format constraint.
Network- based
Predicted outcome; network connectivity analysis; instructional step-by-step reasoning for structural and relational reasoning; guided subgoal structure; network-level counterfactual prompt; output format constraint.
Bimodal
Predicted outcome; top- k SHAP feature descriptions; GNNExplainer subgraph descriptions; combined instructional reasoning linking borrower and relational characteristics; combined borrower-level and network-level counterfactual prompt; disambiguation hint for conflicting contributions; output format constraint.
Table 2: Prompt design components for each explanation pipeline.
Figure 2: Workflow for fine-tuning Gemma 3 and DeepSeek R1 using LoRA-based instruction tuning on 8-bit quantised base models. SHAP and GNNExplainer outputs are paired with ChatGPT-4o-generated explanations to create the synthetic supervision dataset.
Metric
Acronym
CRP Wording
NCRP Wording
Understandability
UND
The explanation is easy to understand.
The explanation is easy to understand.
Trustworthiness
TRU
The explanation can be trusted because it presents sufficient and reliable evidence to support the loan decision.
I trust this explanation given by the system.
Insightfulness
INS
This explanation reveals insightful risk factors that influence loan decisions.
This explanation provides useful insight into why the decision was made.
Satisfaction
SAT
The level of detail satisfies my expectations for a risk assessment.
I am satisfied with how the explanation addressed my concerns about the decision.
Confidence
CON
This explanation increases my confidence for loan approval/denial.
I am confident in the explanation provided for the system’s decision.
Convincingness
CVN
The justification is convincing and argues for the decision using well-weighted risk evidence.
The explanation is convincing and makes the decision seem reasonable.
Table 3: Perceived explanation-quality criteria across CRPs and NCRPs.
Pipeline
LLM
TFC
TDC
NFC
NDC
Tabular- based
Gemma 3
3.77 ± 0.12
3.68 ± 0.15
1.73 ± 0.09
2.14 ± 0.17
DeepSeek R1
4.35 ± 0.10
3.48 ± 0.13
1.58 ± 0.10
1.91 ± 0.17
Gemini 2.5
3.91 ± 0.11
3.94 ± 0.15
1.90 ± 0.09
2.42 ± 0.15
Network- based
Gemma 3
1.02 ± 0.03
1.10 ± 0.07
4.83 ± 0.07
3.20 ± 0.09
DeepSeek R1
1.00 ± 0.00
1.15 ± 0.10
4.46 ± 0.11
3.22 ± 0.09
Gemini 2.5
1.00 ± 0.00
1.01 ± 0.02
4.88 ± 0.07
2.89 ± 0.10
Table 4: LLM-as-a-judge scores across pipeline–LLM configurations. Boldface indicates the highest mean per column within each pipeline.
Metric
Fpipe
ηpipe2
Fllm
ηllm2
Fint
ηint2
TFC
328.96
0.391
11.81
0.014
10.04
0.024
TDC
299.02
0.352
26.08
0.031
22.07
0.052
NFC
354.27
0.377
62.95
0.067
21.66
0.046
NDC
39.19
0.069
28.94
0.051
11.63
0.041
Table 5: Two-way ANOVA results for LLM-as-a-judge fidelity metrics. All effects are statistically significant ( p<10−6 ).
Metric
Gemma 3
DeepSeek R1
Gemini 2.5
UND
4.65 ± 0.14
4.24 ± 0.23
4.41 ± 0.19
TRU
4.23 ± 0.21
3.93 ± 0.25
3.98 ± 0.22
INS
4.09 ± 0.25
3.78 ± 0.25
3.88 ± 0.23
SAT
3.97 ± 0.24
3.75 ± 0.29
3.57 ± 0.27
CON
4.10 ± 0.22
3.70 ± 0.27
3.87 ± 0.24
CVN
4.08 ± 0.22
3.64 ± 0.29
3.74 ± 0.25
Table 6: Human (CRP) mean Likert ratings across explanation-quality metrics. Boldface indicates the highest mean per column.
Metric
Gemma 3
DeepSeek R1
Gemini 2.5
UND
4.36 ± 0.11
4.22 ± 0.15
4.18 ± 0.17
TRU
4.24 ± 0.11
4.19 ± 0.14
4.01 ± 0.16
INS
4.11 ± 0.13
4.03 ± 0.15
4.06 ± 0.16
SAT
4.08 ± 0.14
4.12 ± 0.14
4.01 ± 0.17
CON
4.24 ± 0.12
4.17 ± 0.15
4.11 ± 0.16
CVN
4.14 ± 0.13
4.01 ± 0.15
4.04 ± 0.16
Table 7: Human (NCRP) mean Likert ratings across explanation-quality metrics. Boldface indicates the highest mean per column.
Figure 3: Between-cohort comparisons by LLM.
CRPs
NCRPs
Metric
Gemma 3
DeepSeek R1
Gemini 2.5
Gemma 3
DeepSeek R1
Gemini 2.5
UND
4.05 (0.87)
3.90 (0.88)
3.94 (0.88)
4.10 (0.76)
3.78 (1.04)
4.07 (0.85)
TRU
4.17 (0.82)
3.92 (1.00)
3.68 (1.05)
4.30 (0.60)
4.01 (0.73)
4.07 (0.78)
INS
4.31 (0.82)
3.77 (1.08)
3.68 (0.99)
4.39 (0.55)
4.14 (0.67)
4.07 (0.88)
SAT
4.40 (0.67)
3.81 (1.03)
3.65 (0.93)
4.38 (0.62)
4.16 (0.71)
4.09 (0.79)
CON
4.35 (0.71)
3.62 (1.16)
3.68 (0.92)
4.36 (0.58)
4.11 (0.72)
4.15 (0.83)
Table 8: Participant-level mean Likert ratings by cohort, LLM, and metric, with standard deviations in parentheses
CRPs
NCRPs
Metric
Mean diff.
SD diff.
pHolm
pBH
Mean diff.
SD diff.
pHolm
pBH
UND
− 0.050
0.468
1.000
0.681
− 0.121
0.583
1.000
0.733
TRU
+ 0.350
0.489
0.250
0.250
+ 0.075
0.657
1.000
0.765
INS
+ 0.150
0.648
1.000
0.732
− 0.045
0.510
1.000
0.765
SAT
+ 0.250
0.635
1.000
0.517
− 0.288
0.455
0.722
0.413
CON
+ 0.250
0.687
1.000
0.517
− 0.150
0.555
1.000
0.733
Table 9: Paired differences between DeepSeek R1 and Gemini 2.5 by metric and cohort, with Holm-adjusted and Benjamini–Hochberg FDR-adjusted p -values. Positive values indicate higher ratings for DeepSeek R1.
Component
Variance
Share (%)
Participants (leniency/strictness)
0.507
46.3
Explanation cases (LLM × instance)
0.110
10.0
Residual
0.478
43.7
Table 10: Variance decomposition of overall explanation ratings from a random-intercepts model with participant and explanation-instance effects.
Category
Feature
Description
Sentiment
sentiment_polarity
TextBlob polarity score ∈[−1,1] (negative to positive)
sentiment_subjectivity
TextBlob subjectivity ∈[0,1] (objective to subjective)
vader_compound
VADER composite sentiment score ∈[−1,1]
vader_pos
VADER positive sentiment proportion
vader_neg
VADER negative sentiment proportion
vader_neu
VADER neutral sentiment proportion
Table 11: Linguistic feature categories used in the textual analysis.
Figure 4: Cohen’s d effect sizes comparing linguistic features in high-rated (top quartile) and low-rated (bottom quartile) explanations across quality metrics. Positive values (red) indicate features more prevalent in high-rated explanations; negative values (blue) indicate the reverse.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Feature
Description
fico
FICO credit score at origination
dti
Original debt-to-income ratio
ltv
Original loan-to-value ratio
mi_pct
Mortgage insurance percentage
orig_upb
Original unpaid principal balance
loan_term
Original term of the loan in months
Appendix
Table A.1 : Features used for model training and analysis.
Component
Concrete instantiation
1) Predicted outcome
Prediction: Loan 31780 default probability =0.8712 . Actual outcome: default.
2) Top- k SHAP feature attributions with descriptions
Evidence (top- k SHAP; top-3 shown): • cnt_borr (SHAP:+0.4362): pushed toward default • fico (SHAP:–0.2609): pushed toward no default • dti (SHAP:+0.0799): pushed toward default (Full top- k list is passed to the LLM; only top-3 shown here.) Feature descriptions: cnt_borr = number of borrowers; fico = credit score; dti = debt-to-income ratio.
3–4) Instructional reasoning + guided subgoals
Instructions (follow in order): 1. Explain the predicted outcome in plain language. 2. Use the SHAP ordering; explain each listed factor: what it is, whether it increases/decreases risk, and why it matters to lenders. 3. Combine factors into a short overall justification. Constraints: borrower-facing language; do not mention SHAP/models/algorithms; reference only evidence-block factors; neutral tone; short sentences.
5) Borrower-level counterfactual prompt
Counterfactual guidance: Provide two realistic improvement steps tied to the risk-increasing factors (positive SHAP values). Use cautious phrasing (e.g., “If you were able to…, this would likely improve…”).
6) Output format constraint
Format: No bullet points in the final answer. End with one sentence summarising the main drivers.
Appendix
Table C.1 : Tabular-based prompt example instantiating Table 3 of the main text.
Component
Concrete instantiation
1) Predicted outcome
Prediction: Loan 2885 no-default probability =0.9489 . Actual outcome: no default.
2) Network connectivity analysis
Evidence (GNNExplainer summary; top-3 shown): • Top node features (importance scores): – fico (0.6967) – if_corr (0.6736) – ltv (0.3349) • Top edge connections: – Geographic area connection (Loan 2733 → 2885, cluster size ∼ 75): impact 0.2380 – Area-provider cluster connection (Loan 2622 → 2885, cluster size ∼ 9): impact 0.2314 – Same loan provider connection (Loan 2493 → 2885, cluster size ∼ 145): impact 0.2266 (Full lists are passed to the LLM; only top-3 shown here.)
3–4) Instructional reasoning + guided subgoals
Instructions (follow in order): 1. Explain the predicted outcome in plain language. 2. Describe the strongest connections using the edge list and cluster sizes. 3. Explain (at a high level) why geographic/provider groupings may matter, without adding unverifiable claims. Constraints: focus on relational/context evidence; mention at least two connection types; do not introduce attributes not shown in the evidence block; neutral tone.
5) Network-level counterfactual prompt
Counterfactual guidance: Provide one cautious “what-if” statement on how different relational patterns (e.g., fewer links to a given cluster) could change confidence; avoid implying control over other borrowers.
6) Output format constraint
Format: Include at least one sentence referencing a connection type and a cluster size from the evidence list. End with a one-sentence summary.
Appendix
Table C.2 : Network-based prompt example instantiating Table 3 of the main text.
Component
Concrete instantiation
1) Predicted outcome
Prediction: Loan 1007 default probability =0.9132 . Actual outcome: default.
2) Top- k SHAP feature descriptions
Evidence (tabular SHAP; top-3 shown): • cnt_borr (SHAP= +0.26 ): increases risk • dti (SHAP= +0.21 ): increases risk • fico (SHAP= −0.18 ): reduces risk (Full top- k list is passed to the LLM; only top-3 shown here.)
3) GNNExplainer subgraph descriptions
Evidence (network summary; compact subset shown): • Edge: Same provider connection (Loan 5021 → 1007, cluster size ∼ 12) • Edge: Area-provider cluster (Loan 2893 → 1007, cluster size ∼ 8) • Node: Loan 5021 (defaulted) identified as influential neighbour (Full subgraph/edge lists are passed to the LLM; only a compact subset shown here.)
4) Combined instructional reasoning linking both modalities
Instructions (follow in order): 1. Explain the predicted outcome overall. 2. Explain the strongest tabular drivers first (SHAP ordering), then the strongest network drivers (edge types + cluster sizes). 3. Include one explicit linking sentence describing how borrower-level and relational signals combine.
5) Combined borrower- and network-level counterfactual prompt
Counterfactual guidance: Provide two improvement suggestions: (i) one tied to a positive-SHAP tabular factor; (ii) one stated cautiously about relational context (how different clustering patterns could affect confidence).
6) Disambiguation hint for conflicting contributions
Disambiguation rule: If tabular and network evidence point in different directions, acknowledge both and state that the assessment balances the signals, without adding new evidence.
Appendix
Table C.3 : Bimodal prompt example instantiating Table 3 of the main text.
Pipeline
LLM
Perplexity
Gemma 3
19.38 ± 0.67
Tabular-based
DeepSeek R1
15.93 ± 0.36
Gemini 2.5
33.20 ± 1.21
Gemma 3
21.09 ± 0.59
Network-based
DeepSeek R1
20.06 ± 0.54
Gemini 2.5
28.96 ± 0.72
Appendix
Table D.1: Perplexity scores across pipeline–LLM configurations. Boldface indicates the lowest mean perplexity within each pipeline.
Effect
F
p
η2
Pipeline
18.34
1.6×10−8
0.010
LLM
1302.39
<10−262
0.687
Pipeline × LLM
68.85
1.1×10−50
0.073
Appendix
Table D.2: Two-way ANOVA results for perplexity with factors Pipeline and LLM.
Pipeline
LLM
UND
TRU
INS
SAT
CON
CVN
COM
USB
Gemma 3
3.89 ± 0.07
2.11 ± 0.06
2.15 ± 0.07
1.99 ± 0.02
2.62 ± 0.10
2.11 ± 0.06
2.70 ± 0.09
1.42 ± 0.10
Tabular-based
DeepSeek R1
3.79 ± 0.08
2.38 ± 0.10
2.34 ± 0.09
2.14 ± 0.07
2.75 ± 0.09
2.38 ± 0.10
2.95 ± 0.07
1.79 ± 0.08
Gemini 2.5
3.99 ± 0.08
2.80 ± 0.08
2.78 ± 0.09
2.58 ± 0.10
2.96 ± 0.08
2.79 ± 0.08
3.29 ± 0.12
1.86 ± 0.07
Gemma 3
3.62 ± 0.10
1.97 ± 0.03
1.98 ± 0.03
1.96 ± 0.04
2.15 ± 0.08
1.97 ± 0.03
2.78 ± 0.08
1.12 ± 0.06
Network-based
DeepSeek R1
2.98 ± 0.03
1.77 ± 0.08
1.81 ± 0.08
1.57 ± 0.10
1.91 ± 0.06
1.77 ± 0.08
2.02 ± 0.04
1.00 ± 0.00
Gemini 2.5
2.80 ± 0.08
1.40 ± 0.10
1.55 ± 0.10
1.29 ± 0.09
1.78 ± 0.08
1.41 ± 0.10
1.90 ± 0.07
1.00 ± 0.00
Appendix
Table D.3: Simulated CRP ratings across pipeline–LLM configurations. Boldface indicates the highest mean per column.
Pipeline
LLM
UND
TRU
INS
SAT
CON
CVN
COM
USB
Gemma 3
4.43 ± 0.11
3.44 ± 0.10
3.89 ± 0.06
3.45 ± 0.10
3.43 ± 0.10
4.13 ± 0.10
4.43 ± 0.11
3.38 ± 0.10
Tabular-based
DeepSeek R1
4.00 ± 0.08
3.20 ± 0.10
3.85 ± 0.08
3.36 ± 0.10
3.19 ± 0.10
3.91 ± 0.08
4.00 ± 0.08
3.10 ± 0.09
Gemini 2.5
4.39 ± 0.10
3.62 ± 0.10
4.19 ± 0.08
3.81 ± 0.08
3.62 ± 0.10
4.29 ± 0.11
4.39 ± 0.10
3.54 ± 0.10
Gemma 3
3.97 ± 0.07
2.97 ± 0.07
3.68 ± 0.11
3.02 ± 0.06
2.97 ± 0.07
3.86 ± 0.07
3.97 ± 0.07
2.97 ± 0.07
Network-based
DeepSeek R1
3.63 ± 0.10
2.64 ± 0.10
2.91 ± 0.06
2.65 ± 0.10
2.64 ± 0.10
3.02 ± 0.06
3.63 ± 0.10
2.41 ± 0.10
Gemini 2.5
3.09 ± 0.08
2.04 ± 0.06
2.45 ± 0.10
2.04 ± 0.06
2.04 ± 0.06
2.93 ± 0.06
3.09 ± 0.08
2.02 ± 0.06
Appendix
Table D.4: Simulated NCRP ratings across pipeline–LLM configurations. Boldface indicates the highest mean per column.
Metric
Fpipe
ppipe
ηpipe2
Fllm
pllm
ηllm2
Fint
pint
ηint2
UND
471.115
8.97×10−140
0.517
51.290
8.55×10−22
0.105
54.234
6.95×10−41
0.198
TRU
288.654
4.30×10−97
0.396
0.575
0.5628
0.001
52.728
7.66×10−40
0.194
INS
260.980
1.07×10−89
0.373
3.884
0.0209
0.009
38.091
2.03×10−29
0.148
SAT
243.487
7.18×10−85
0.357
0.930
0.3951
0.002
78.371
7.06×10−57
0.263
CON
445.581
2.41×10−134
0.503
0.512
0.5996
0.001
22.268
1.56×10−17
0.092
CVN
279.329
1.24×10−94
0.389
0.574
0.5632
0.001
50.728
1.89×10−38
0.188
Appendix
Table D.5: Two-way ANOVA results for simulated CRP ratings.
Metric
Fpipe
ppipe
ηpipe2
Fllm
pllm
ηllm2
Fint
pint
ηint2
UND
229.281
7.38×10−81
0.343
27.297
3.15×10−12
0.058
53.932
1.12×10−40
0.197
TRU
302.394
1.16×10−100
0.408
17.788
2.67×10−8
0.039
57.377
4.83×10−43
0.207
INS
685.194
4.47×10−180
0.609
54.104
6.93×10−23
0.110
123.595
1.06×10−83
0.260
SAT
503.172
2.23×10−146
0.534
16.600
8.39×10−8
0.036
72.315
5.50×10−53
0.248
CON
297.011
2.85×10−99
0.403
16.985
5.79×10−8
0.037
58.700
6.06×10−44
0.211
CVN
461.246
1.08×10−137
0.512
68.807
1.73×10−28
0.135
75.891
2.71×10−55
0.257
Appendix
Table D.6: Two-way ANOVA results for simulated NCRP ratings.
Figure F.1 : Demographic and background characteristics of survey participants.
Large language models (LLMs) have shown promise in translating model-based explanations into human-readable narratives. This study evaluates whether LLMs can serve as post-hoc explainability interfaces for credit risk models, focusing on their ability to preserve feature-importance rankings and generate autonomous explanations. Using a LendingClub dataset, we compare LLM outputs with SHAP and coefficient-based attributions on three major LLMs, including GPT-4-turbo, Claude-Sonnet-4.5, and Gemini-2.5-Flash. Results indicate that LLMs reliably reproduce reference rankings under controlled prompts but show limited alignment when generating explanations autonomously. These findings suggest that LLMs are best deployed as narrative interfaces rather than substitutes for formal attribution methods in credit risk governance.
Wenxi Geng, Dingyuan Liu, Liya Li +1
Citigroup, Irving, TX, United States · University of North Carolina at Chapel Hill, Chapel Hill, NC, United States · Georgia Institute of Technology, Atlanta, GA, United States +1
Credit scoring increasingly relies on models whose decision logic cannot be read off their parameters, in tension with supervisory expectations that adverse decisions be explainable. A common proposal closes that gap with a language model: compute feature attributions, hand them to an LLM, and let it write the rationale. We build such a system end to end and test whether the second half of the promise holds. The predictive component is a multi-scale stacking ensemble fusing four differently regularised gradient-boosting learners with a residual network through a neural meta-learner trained on out-of-fold predictions. On a public 32,581-application credit dataset it reaches test ROC-AUC 0.9539 (95% CI [0.9462, 0.9616]) and PR-AUC 0.9137, beating the best single model by Delta-AUC = 0.0143 (p = 0.016 under a conservative independence assumption). Our central finding is asymmetric. The ranking gain is real but operationally small: at the F1-optimal threshold the ensemble avoids only six additional missed defaults out of 1,422 against a tuned random forest, cutting cost-weighted loss by under 2%. The narrative layer fails in a way prompt engineering alone does not fix. In an audited case the model named three factors as risk-increasing that the supplied attributions scored as risk-reducing, omitted the dominant driver, and introduced a feature never given to it. We trace this to properties we measure rather than assume: SHAP and LIME agree on which features matter (overlap@10 = 0.80) but not on their order (tau = 0.43, p = 0.18), and the attribution sign for the model's most sensitive input is near a coin flip across applicants (modal-sign share 0.53). Calibration (ECS = 0.117) and perturbation stability (DPD = 0.078) both fall short of our own thresholds. Constrained prompting is necessary but not sufficient: grounding must be verified after generation, not assumed.
Gregorius Reynaldi Pratama, Kuo-Kun Tseng
School of Computer Science and Technology Harbin Institute of Technology, Shenzhen Shenzhen, China
Large language models (LLMs) are increasingly deployed not only to make decisions but to explain them. While AI decision fairness has been studied extensively, the fairness of AI explanations (whether LLMs justify decisions with equal quality, depth, tone, and linguistic sophistication across demographic groups) has received little attention. This paper introduces the Explanation Fairness Taxonomy (EFT), a framework comprising five formally defined, operationalizable dimensions: Verbosity Disparity, Sentiment Disparity, Epistemic Hedging Disparity, Decision-Linked Explanation Disparity, and Lexical Complexity Disparity. The taxonomy is instantiated in a controlled empirical study across 80 prompt templates, four consequential decision domains (hiring, medical triage, credit assessment, legal judgment), and five LLMs: GPT-4.1, Claude Sonnet, LLaMA 3.3 70B, GPT-OSS 120B, and Qwen3 32B. Two novel black-box metrics are introduced: the Hedging Density Score (HDS) and the Explanation Faithfulness Proxy (EFP), a heuristic indicator of decision-linked explanation variation. Across up to 400 prompt pairs, all eight EFT metrics show statistically significant disparities (Cohen's d ranging from small to large, all p_BH < 10^(-62)). Model choice is strongly associated with disparity magnitude: Qwen3 32B exhibits verbosity disparities 5.9x larger than LLaMA 3.3 70B. Two prompting-based mitigations show significant reductions in EFP disparity (78-95%) but no significant effect on stylistic dimensions, consistent with the hypothesis that stylistic explanation inequalities are encoded in pre-training distributions and are not resolvable through deployment-level instruction alone. A reproducible measurement framework is offered for explanation-level fairness auditing, with implications for AI regulation and deployment practice.