PaperlayerView exact revision on arXiv

Source-linked AI summary

Multi-Dimensional Behavioral Evaluation of Agentic Stock Prediction Systems Using Large Language Model Judges with Closed-Loop Reinforcement Learning Feedback

Mohammad Al Ridhawi, Mahtab Haj Ali, Hussein Al Osman

arXiv:2605.05739v3cs.LGcs.AIcs.CLq-fin.CP

TL;DR

Output-only evaluation cannot diagnose the intermediate decisions behind an agent’s result, and behavioral assessment of temporal, stochastic decisions remains underdeveloped. This paper evaluates logged decisions across six dimensions and feeds dimension-specific diagnostics into reinforcement learning; its intervention was associated with lower squared-error loss in the reported test results.

  • Problem

    Output-level evaluation cannot localize which intermediate decision was responsible for an error or establish whether a correct result came from sound decisions.

    Source: §I · p. 1

  • Method

    The framework scores logged behavioral traces across six domain-specific dimensions, validates dimension specificity with perturbations, and routes diagnostics into targeted reinforcement-learning reward changes.

    Source: §I · p. 2

  • Results

    DM = −7.83 (p < 0.001), with the negative sign indicating lower squared-error loss post-intervention in the fixed-effects panel pool.

    Source: H. Statistical Validation · p. 11

  • Takeaways & Limitations

    Per-dimension diagnostics can modify the existing controller’s reward so it addresses identified weaknesses without architectural changes.

    Source: §I · p. 2

    Temperature-zero judging removes stochastic variation that could reveal uncertainty on borderline cases.

    Source: D. Limitations · p. 15

Abstract

from arXiv · show

Agentic artificial intelligence systems produce outputs through sequences of interdependent autonomous decisions, yet standard evaluation assesses outputs alone and cannot diagnose the underlying process. We develop a behavioral evaluation methodology that complements output-level testing by scoring the intermediate decision process itself. Behavioral traces logged at each autonomous decision point are grouped into five-day episodes and scored along six domain-specific dimensions (regime detection, routing, adaptation, risk calibration, strategy coherence, error recovery) by an ensemble of three large language model (LLM) judges. A perturbation procedure that corrupts one dimension while leaving the other five intact confirms dimension specificity; cross-model agreement reaches Krippendorff's alpha = 0.85. The composite behavioral score correlates at Spearman rho = 0.72 with realized 20-day Sharpe ratio. Closing the loop, the framework converts deficient per-dimension scores into a credit-assigned penalty added to the Soft Actor-Critic reward. Three fine-tuning cycles, confined to validation data, reduce one-day MAPE from 0.61% to 0.54% (11.5% relative; p<0.001, d=0.31) on the held-out 2017 to 2025 test period, significant under Diebold-Mariano and localized by Giacomini-White to the high-volatility regime. The methodology is application-agnostic and applies to any agentic system whose intermediate decisions can be logged.

I. INTRODUCTION

Output-only evaluation cannot localize the quality of interdependent agent decisions, motivating a behavioral evaluation framework that complements conventional forecasting tests. The paper develops and demonstrates that framework on an adaptive stock-prediction system.

I. INTRODUCTION

  • Output-only benchmarks cannot establish whether an agent’s intermediate decisions were sound, and LLM-based behavioral evaluation for temporal, stochastic decisions remains underdeveloped.The paper frames the gap as especially relevant when outcomes and decision quality can diverge.Source: §I · p. 1

A. Case-Study System and Behavioral Context

  • The case study evaluates an adaptive regime-aware stock predictor whose daily forecasts result from a four-stage detect, route, predict, and adjust decision chain.Its architecture is treated as the demonstration system, not as the paper’s contribution.Source: §I · p. 2, A. Case-Study System and Behavioral… · p. 2

II. MATERIALS AND METHODS

  • The framework combines trace formalization, six domain-specific dimensions, mechanically isolated perturbations, and credit-assigned reward changes, alongside validation against outcomes and forecast tests.It supplements rather than replaces output-level evaluation.Source: §I · p. 2, §II · p. 2

B. Behavioral Trace Formalization

  • Each daily behavioral trace records market context, regime detection, routing, controller actions, forecasts, and rolling performance for evaluation across episodes.The trace includes reconstruction error and threshold alongside the regime label, allowing judges to assess classification margins and decision coherence.Source: B. Behavioral Trace Formalization · p. 2

C. Evaluation Episodes

The framework evaluates agent behavior in five-day episodes using six dimensions that cover both individual architectural decisions and emergent properties. Validation pilots favor six dimensions for balancing distinct failure modes, judge agreement, and predictive validity.

C. Evaluation Episodes

  • Five-day episodes capture the temporal context needed to assess adaptation, error recovery, and strategy coherence, with validation pilots favoring this window over three-, seven-, and ten-day alternatives.Episodes do not overlap to reduce dependence between evaluated units.Source: C. Evaluation Episodes · p. 3

1) Architectural decision dimensions:

  • The architectural dimensions assess regime classification, pathway routing near threshold boundaries, and the timeliness and proportionality of SAC parameter adjustments.Regime errors can propagate downstream through routing and prediction.Source: 1) Architectural decision dimensions · p. 3

2) Emergent behavioral dimensions:

  • Emergent dimensions assess risk posture against volatility, consistency across the decision chain, and recovery after prediction errors.Error recovery requires timely adjustments and subsequent measurable improvement; persistent errors and overcorrections lower its score.Source: 2) Emergent behavioral dimensions · pp. 3–4

D. Evaluation Dimensions

  • The six-dimension scheme shows limited inter-dimension correlation across 200 episodes; coarser variants conflate failure modes, while eight dimensions lower agreement without improving predictive validity.Mean pairwise Spearman correlation was 0.26; the highest was 0.41 between regime detection and routing.Source: 2) Emergent behavioral dimensions · p. 4

E. Scoring Rubric

The rubric scores six behavioral dimensions on anchored ordinal criteria, while an LLM-judge ensemble produces dimension-level evaluations that can be translated into calibrated SAC reward feedback.

E. Scoring Rubric

  • Scores use a 1–5 integer scale; scores 2–4 interpolate between the rubric’s endpoint anchors, and a 1–10 scale yielded lower agreement without added diagnostic information.Table II reports the anchor descriptions for scores 1 and 5.Source: E. Scoring Rubric · p. 4, Table II · p. 4
  • Rubric anchors reference observable trace indicators rather than abstract quality descriptors, and inter-judge agreement is measured with Krippendorff’s alpha.The rubric follows the structured-criteria principle attributed to.Source: E. Scoring Rubric · p. 4

F. LLM Judge Ensemble and Evaluation Pipeline

  • Each judge receives the system description, a five-trace JSON episode, the six-dimension rubric, and a structured output schema for scores, justifications, and optional failure labels.The prompt asks judges to assess market conditions and decisions against information available at the time.Source: F. LLM Judge Ensemble and Evaluation… · p. 4
  • Three independent judges from different model families score each episode at temperature zero, and cross-family aggregation reduces reliance on any single model’s idiosyncrasies.Structured reasoning asks judges to track decision chains across days.Source: F. LLM Judge Ensemble and Evaluation… · p. 5
  • Per-dimension consensus scores are the primary evaluation and training units; the composite is reported for cross-episode comparison but not used as a training signal.The composite would discard targeted feedback.Source: F. LLM Judge Ensemble and Evaluation… · p. 5

1) Reward modification:

  • The reward modification subtracts a weighted linear-hinge penalty for scores below θ = 3, with λ = 0.15 scaling feedback and the hinge limiting destabilizing penalty magnitudes.Dimensions at or above θ receive no penalty.Source: 1) Reward modification · p. 5
  • The λ sweep found joint prediction and behavioral improvement at λ ∈ [0.10, 0.20], under-correction at λ ≤ 0.05, and destabilization at λ ≥ 0.25.At λ = 0.30, final metrics were worse than the pre-intervention baseline.Source: 1) Reward modification · p. 5
  • Dimension weights equal validation-set Spearman correlations with realized Sharpe ratios over the subsequent 20 trading days and remain fixed during closed-loop cycles.Fixing weights avoids circular optimization of behavior and its evaluation weights.Source: 1) Reward modification · p. 5

2) Credit assignment:

Credit assignment routes dimension-specific behavioral penalties to SAC action components, while periodic evaluation and a strict temporal split constrain when the feedback is applied and tested.

2) Credit assignment:

  • The SAC controller has two action components, Δτ and Δα, so misrouting a dimension’s penalty could create conflicting gradients and harm non-targeted behavior..Source: 2) Credit assignment · p. 5
  • Credit vectors assign RD and RC to Δτ, RT to Δα, and AD, SC, and ER to both action subspaces.The complete mapping is identified in Table IV.Source: 2) Credit assignment · p. 5
  • Only action components with non-zero credit-assignment entries receive each dimension’s penalty; ablations report that targeted routing preserves non-deficient dimensions.Categorical failure labels can override the default mapping for atypical cases such as a routing failure caused by an incorrect threshold.Source: 2) Credit assignment · pp. 5–6

3) Periodic evaluation protocol:

  • Each cycle collects 40 days of normal operation, evaluates eight five-day episodes with three judges, then activates the reward modification for 10 SAC fine-tuning epochs if any mean dimension score is below 3.The modification is deactivated after fine-tuning; three cycles brought all dimension means above the threshold.Source: 3) Periodic evaluation protocol · p. 6

H. Experimental Setup

  • Closed-loop adaptation, tuning, prompt analyses, and ablations are confined to validation data; SAC weights are frozen before final evaluation on the held-out 2017–2025 test set.The experiments use the same 20 equities and temporal partitioning as prior work [7],.Source: H. Experimental Setup · p. 6

I. Perturbation Design

The perturbation design tests whether judges identify isolated behavioral deficiencies, while the study reserves the test period for final comparisons and establishes the pre-feedback baseline.

I. Perturbation Design

  • Six perturbations were applied independently to 60 baseline episodes, each designed to corrupt one behavioral dimension while mechanically preserving the other five.The baselines were sampled across three market-regime strata.Source: §I · pp. 6–7
  • The six perturbations target regime detection, routing, adaptation, risk calibration, strategy coherence, and error recovery, producing 360 perturbed episodes plus 60 originals for 1,260 judge evaluations.The transformations include regime inversion, forced wrong-pathway routing, frozen SAC, removed VIX scaling, contradictory actions, and disabled recovery.Source: §I · p. 7

J. Statistical Validation Procedures

  • Test-period validation combines paired tests, Sharpe-ratio bootstrap intervals, Diebold–Mariano tests across four losses, benchmark comparisons, Giacomini–White tests, and model confidence sets.The held-out 2017–2025 test set receives no updates, reward changes, or hyperparameter adjustments.Source: J. Statistical Validation Procedures · p. 7

A. Baseline System Performance

  • Before feedback, AE-NodeFormer + SAC achieved 0.61% MAPE and 71% directional accuracy on validation episodes, a 24% MAPE reduction versus the single-pathway NodeFormer-BERT baseline.The full baseline comparisons and multi-horizon results are reported in [7].Source: A. Baseline System Performance · p. 7

B. Perturbation-Based Validation

Perturbations produced targeted score drops, cross-model agreement was reliable across dimensions, and behavioral scores varied with market volatility in dimension-specific ways.

B. Perturbation-Based Validation

  • Every perturbation caused a significant targeted score drop (p < 0.001), ranging from −1.6 for disabled recovery to −2.4 for regime inversion.Regime inversion produced the largest drop; disabled recovery’s effect was attenuated when episodes lacked significant errors.Source: B. Perturbation-Based Validation · pp. 7–8
  • Mean off-target shift was 0.32, smaller than targeted drops; contradictory actions had the largest shift (−0.5), while frozen SAC and disabled recovery had the smallest (−0.2).Some downstream shifts are expected because later decisions depend on earlier decisions; the observed pattern retains targeted sensitivity while recognizing those effects.Source: B. Perturbation-Based Validation · p. 8

C. Cross-Model Agreement

  • Krippendorff’s α ranged from 0.74 for risk calibration to 0.85 for regime detection, with all six dimensions above the stated 0.667 threshold for meaningful reliability.Regime detection, routing, and strategy coherence exceeded 0.800; differences in risk-tolerance assumptions contributed to lower agreement on risk calibration.Source: C. Cross-Model Agreement · p. 8
  • GPT 5.4 and Claude 4.6 had pairwise kappa values of 0.71–0.83, while their agreement with Gemini 3.1 Pro was 0.65–0.79; averaging yielded more stable composite scores than any single judge.The judges differed in both directions on borderline lags and recovery overshoots.Source: C. Cross-Model Agreement · p. 8

D. Per-Regime Score Analysis

  • Risk calibration and adaptation responsiveness each fell 1.3 points from low to high volatility; risk calibration scored 3.4 versus 2.1, and routing declined by 1.2 points.The regime-stratified table reports mean dimension scores on a 1–5 scale.Source: Table VIII · p. 8, D. Per-Regime Score Analysis · p. 8
  • Error recovery scored 3.1 in high-volatility episodes versus 2.8 in medium volatility, while regime detection remained at 3.3 in high volatility.The authors attribute the recovery difference to clearer error signals and more opportunities to demonstrate correction in high-volatility periods.Source: D. Per-Regime Score Analysis · p. 8
  • The largest closed-loop gains were expected in high-volatility episodes because behavioral weaknesses were concentrated in particular dimensions and conditions.The authors state that this prediction was confirmed in Section III-G.Source: D. Per-Regime Score Analysis · p. 8

E. Predictive Validity

Behavioral scores track subsequent trading outcomes, and qualitative examples show how judges assess decision sequences. Three closed-loop cycles raised deficient validation-period scores, while the frozen controller improved held-out test metrics.

E. Predictive Validity

  • Targeted perturbations degraded their intended dimensions far more than other dimensions, supporting dimension specificity.Source: Figure 3 · p. 9
  • ρ = 0.72: the composite behavioral score correlated most strongly with realized 20-day Sharpe ratio.Individual dimensions had distinct predictive profiles: regime detection correlated most strongly with inverse MAPE (ρ = 0.69), risk calibration with inverse maximum drawdown (ρ = 0.64), and strategy coherence with returns (ρ = 0.57).Source: E. Predictive Validity · p. 9, Table IX · p. 9

F. Qualitative Analysis of LLM Judgments

  • A regime-label inversion was identified by all three judges, while routing was only moderately affected because its intermediate pathway weight remained reasonable.The episode illustrates that a perturbed decision dimension need not imply an equally poor score on another dimension.Source: F. Qualitative Analysis of LLM Judgments · p. 10
  • Judges grounded ratings in trace values and agreed on objective dimensions, while risk calibration and error recovery drew differing interpretations of acceptable response and proportionality.For example, judges differed on whether a one-day lag in adjusting risk posture was acceptable and whether a small correction overshot.Source: F. Qualitative Analysis of LLM Judgments · pp. 9–10

G. Closed-Loop Improvement

  • After three cycles, risk calibration rose from 2.1 to 3.5, routing from 2.6 to 3.7, and adaptation from 2.4 to 3.6, exceeding the intervention threshold of 3.Risk calibration adjustments became proactive, routing used intermediate pathway weights near the threshold, and adaptation corrected earlier in error sequences.Source: G. Closed-Loop Improvement · pp. 10–11
  • 0.61% to 0.54%: held-out one-day MAPE fell 11.5% after three cycles, with gains larger in high-volatility episodes (17.3%) than low-volatility episodes (7.8%).The five-day and twenty-day MAPE reductions were 10.8% (1.02% to 0.91%) and 8.6% (1.52% to 1.39%), respectively.Source: G. Closed-Loop Improvement · p. 11
  • Cycle gains showed diminishing returns; dimensions already above threshold stayed stable, with mean changes below 0.2, consistent with targeted penalties avoiding degradation.Routing improved fastest, risk calibration required all three cycles to exceed threshold, and adaptation's largest gain came in cycle two.Source: G. Closed-Loop Improvement · p. 11

H. Statistical Validation

The authors test the closed-loop system on held-out data against its pre-intervention configuration and statistical benchmarks. Multiple predictive-accuracy tests report significant gains, including a larger advantage in high volatility.

a) Paired significance and Sharpe bootstrap

  • 0.61% to 0.54%: held-out one-day MAPE reduction was significant (p < 0.001, d = 0.31), alongside directional accuracy rising from 71% to 74%.The Sharpe-ratio improvement was 18%, with a 95% bootstrap confidence interval of [8.2%, 27.4%].Source: H. Statistical Validation · p. 11

b) Diebold-Mariano test of equal predictive accuracy

  • DM = −7.83 (p < 0.001): the pooled test found lower post-intervention squared-error loss, and 18 of 20 per-stock tests remained significant after Bonferroni correction.All 20 per-stock tests were significant at the uncorrected 5% level.Source: H. Statistical Validation · pp. 11–12

c) Comparison against statistical benchmarks

  • The post-intervention system achieved roughly 60% lower one-day MAPE than each statistical benchmark, with all three Diebold-Mariano comparisons significant (p < 10^-6).The pre-intervention agentic system already outperformed the benchmarks; closed-loop training added a further significant improvement.Source: H. Statistical Validation · p. 12, Table XI · p. 12

d) Conditional predictive ability

  • High-volatility conditional advantage was about three times the low-volatility advantage, and the conditional predictive ability test yielded Wald = 86.41 (p < 10^-9).Post-intervention loss was lower in both regimes.Source: H. Statistical Validation · p. 12

e) Model confidence set across closed-loop cycles

  • The 90% model confidence set contained cycles 3 and 2, while the 75% set contained only cycle 3; the λ = 0 ablation was rejected against cycle 3 at 99%.This supports the need for the LLM penalty rather than generic additional training.Source: H. Statistical Validation · p. 12

f) Robustness across loss functions

  • Across squared-error, MAE, MAPE, and QLIKE losses, pooled Diebold-Mariano statistics ranged from −7.04 to −7.83, all p < 0.001.The authors report that the narrow spread indicates the improvement is not specific to squared-error loss.Source: H. Statistical Validation · p. 12

I. Framework Ablations

Framework ablations examine the judge ensemble, credit assignment, and prompt phrasing. Results favor ensemble variance reduction and targeted penalties, while prompt robustness is supported only within a bounded check.

1) Number of judges

  • The three-judge ensemble's per-episode score variance was 0.29, 29% to 37% lower than any single judge, producing a more stable SAC reward signal.All single-judge variants improved over the pre-intervention baseline, but the ensemble had better outcomes; its MAPE gap over single judges was 0.01 to 0.04.Source: 1) Number of judges · p. 12

2) Credit assignment strategy

  • Targeted credit assignment improved deficient dimensions by 1.3 points versus 0.9 for uniform assignment, while preserving non-deficient dimensions (mean change +0.1).Uniform penalties reduced non-deficient scores by 0.4 points on average, including degraded regime detection from a routing-targeted penalty.Source: 2) Credit assignment strategy · p. 13, Table XIII · p. 13

3) Prompt sensitivity

  • Prompt variants differed by a mean absolute 0.14 points per dimension and had ICC = 0.91, with no systematic bias reported.This bounded check covered 60 episodes, one judge, and only ordering and lexical simplification variants.Source: 3) Prompt sensitivity · p. 13

IV. DISCUSSION

The results support selective, reliable behavioral assessment and show that its scores can guide targeted optimization. The framework's broader use is conditional on structured traces and domain-specific design, while empirical magnitudes remain tied to the demonstrated case study.

A. Interpretation of Findings

  • Each dimension-specific perturbation lowered its targeted score by at least 1.6 points while shifting other dimensions by only 0.32 points on average.Cross-model agreement ranged from Krippendorff’s α = 0.74 to 0.85.Source: A. Interpretation of Findings · p. 13
  • The composite behavioral score correlated with realized 20-day Sharpe ratio at ρ = 0.72, exceeding each individual dimension’s correlation.Regime detection predicted MAPE, risk calibration predicted drawdown, and strategy coherence predicted returns.Source: A. Interpretation of Findings · p. 13
  • 11.5% one-day MAPE reduction and a 3-point directional-accuracy gain followed closed-loop correction, with gains concentrated in initially deficient dimensions and high-volatility episodes.Non-targeted dimensions remained stable, and the Hansen MCS rejected the same-step SAC training without an LLM penalty at the 99% level.Source: A. Interpretation of Findings · p. 13
  • Uniform penalty distribution degraded non-targeted dimensions by 0.4 points on average, while the λ sweep identified an effective range of 0.10 to 0.20.The multi-judge ensemble reduced per-episode score variance by 29 to 37%.Source: A. Interpretation of Findings · p. 13

B. Implications for Applied AI and Agentic-System Evaluation

  • Direct process assessment can support diagnosis and optimization when decisions are logged structurally, dimensions are domain-specific, and perturbations validate dimension specificity.The resulting scores identify which intermediate decision was unsound when a system failed, a question aggregate metrics cannot answer.Source: B. Implications for Applied AI and… · pp. 13–14
  • The framework extends LLM judging from static outputs to temporal decision sequences and routes evaluation feedback to specific controller action components.The paper presents this as an extension of the RLHF principle,.Source: B. Implications for Applied AI and… · p. 14

C. Generalizability beyond the Demonstration Domain

  • The core methodology transfers in principle, but new domains require redesigned dimensions, rubric anchors, and action-subspace mappings.Empirical magnitudes should be re-estimated: they reflect one architecture, 20 S&P 500 equities, and the 2017 to 2025 test window.Source: C. Generalizability beyond the… · p. 14

D. Limitations

The paper identifies limits in judge expertise, evaluation stability, and the scope of its empirical validation. It proposes broader validation, judge distillation, and uncertainty estimation as future work, while noting data-access and reproducibility constraints.

D. Limitations

  • LLM judges may miss subtle financial signals absent from trace fields, and a direct comparison with expert human evaluators remains outside this study’s scope.The judges are generalist models whose domain knowledge comes from pre-training.Source: D. Limitations · p. 14
  • The closed-loop method retains a Goodhart-type risk of optimizing for LLM approval rather than task performance, especially over many cycles or on behavior unlike judge pre-training.The authors report no such effect in this study and identify periodic corrective feedback as a safeguard.Source: D. Limitations · p. 14
  • The framework is validated on one architecture and one case study; empirical magnitudes should be re-estimated on distinct systems and broader samples.Evaluation cost scales linearly with episodes and judges, potentially requiring validated judge distillation for real-time or higher-frequency use.Source: D. Limitations · p. 14
  • Provider-side model updates can make future live re-queries diverge from archived evaluations; complete drift across all three judges would require re-validation.The archived scores and rationales support contemporaneous replay.Source: D. Limitations · pp. 14–15
  • Temperature-zero evaluations are deterministic but cannot reveal model uncertainty on borderline cases.The authors propose temperature sampling to estimate episode-level uncertainty and potentially soften penalties or flag cases for human review.Source: D. Limitations · p. 15, E. Future Work · p. 15

E. Future Work

  • Future work proposes validation on distinct systems and domains, distillation into a faster surrogate, and temperature-sampled uncertainty estimates.The surrogate would need to preserve dimension specificity and predictive validity.Source: E. Future Work · p. 15

DISCLOSURE OF INTEREST / FUNDING

DATA AVAILABILITY STATEMENT

  • The daily market snapshot is deposited at Zenodo, but X-derived sentiment data are not redistributable; framework code is available from the corresponding author on reasonable academic request.Reproducing the sentiment series requires independent collection under the platform terms in force at access time.Source: DATA AVAILABILITY STATEMENT · p. 15

APPENDIX

The appendix documents the judges’ fixed prompt and scoring instructions, then illustrates how daily traces are serialized for episode-level evaluation. The prompt asks judges to assess six behavioral dimensions using trace-specific evidence.

A.1 System message

  • All three judges receive a fixed system message, episode-specific trace serialization, and constrained output schema; only the trace serialization varies across episodes.The prompt specifies an expert evaluator for the stock prediction system and provides five consecutive daily traces per episode.Source: APPENDIX · p. 16, §A.1 · p. 16
  • Judges score six dimensions on an integer 1–5 scale, with rubric anchors and instructions to ground each score in specific trace observations.The prompt defines interpretation rules for regimes, routing, and timely, proportional, stable adaptation, and requests a four-step assessment process.Source: §A.1 · p. 16

A.2 Behavioral-trace serialization

  • Daily traces are serialized as JSON records and presented in chronological order across the five-day episode.The appendix includes a representative high-volatility record.Source: §A.2 · p. 16

A.3 Output schema and failure-label vocabulary

The output schema returns dimension scores and justifications, with optional failure labels for low-scoring dimensions. Those labels map failures to SAC action subspaces for penalty routing.

A.3 Output schema and failure-label vocabulary

  • The judge returns an integer score from 1 to 5 and a trace-referencing justification for each of six dimensions.Source: §A.3 · p. 16
  • A failure label is recorded only for a dimension scoring below 3 and selects the action component receiving the penalty through credit-assignment routing.The 12-label vocabulary and default SAC action-subspace mapping are listed in Table A1.Source: §A.3 · p. 16, Table A1 · p. 16

Related papers

Loading 2605.05739v3…