The research, conducted by Mohammadreza Rezaei, Sridhar Sreemulnath Iyer, and Omid Fatahi Valilai, tackles a gap that has long frustrated both researchers and practitioners. Multi-criteria decision-making methods such as TOPSIS can rank suppliers rigorously, but only after someone has formalised the criteria, weights, constraints, and evidence. Large language models, meanwhile, excel at reading natural-language procurement requests and extracting information from unstructured documents, yet they cannot be trusted to perform arithmetic or to rank suppliers autonomously. Existing approaches, the authors argue, address fragments of the problem while leaving the complete path from a spoken or written sourcing request to a traceable, defensible supplier decision largely uncharted.

The framework’s central design principle is a strict division of labour. Language model components handle the semantic work: interpreting what a procurement request actually means, formulating candidate evaluation scenarios, extracting claims from supplier documents, and explaining completed results in plain language. Deterministic services handle everything numerical: constructing weight vectors, applying constraints and evidence policies, calculating multi-criteria scores, resolving ties, and fixing the final supplier order. Crucially, the language models never touch the numbers. When a manager asks to place greater emphasis on geographic fit, the model proposes only a direction and a bounded strength category; a deterministic controller then converts that advice into a valid numerical weight vector through a constrained projection that enforces unit sums and policy-defined bounds.

Human authority is organised through four checkpoints. The data owner approves the structured supplier table and its quality record. The procurement analyst approves the scenario, its criteria, constraints, and the numerical weights. The evidence reviewer confirms whether extracted claims genuinely support their cited passages and what analytical role each claim may play. Finally, the decision owner reviews the complete package, including rankings, comparisons, explanations, and caveats, before any action is taken. Every consequential action generates an append-only audit event recording who acted, in what role, when, on what object, and why, creating a traceable chain from the original request to the final sourcing decision.

The empirical evaluation is unusually comprehensive for this field. Using more than 22,000 real procurement cases drawn from the European Union’s Tenders Electronic Daily database, the team first compared deterministic ranking methods. TOPSIS, which ranks suppliers by their distance from ideal and anti-ideal solutions after vector normalisation and weighting, outperformed frequency-based ranking, weighted-sum models, and VIKOR. But the more striking finding concerned the supplier representation itself. A generic four-feature baseline placed the historically awarded supplier in the top ten of only 5.259 percent of cases. When the researchers added procurement-specific feature groups capturing category fit, geographic relevance, and prior buyer relationships, that figure rose to 31.642 percent, and the median rank of the reference supplier improved from 914 to just 34.

The scenario experiments revealed how the framework responds when sourcing priorities change. Under a continuity profile that increases the weight on prior buyer relationships, the top-ten overlap with the balanced baseline remained high at 0.9369. Under a diversification profile that reduces reliance on existing relationships, overlap fell to 0.8304, yet suppliers entering the top ten had stronger geographic fit than those exiting in 96.984 percent of turnover cases. In other words, the rankings moved in traceable, explainable ways that followed the approved priorities, allowing procurement professionals to compare competing strategies and see exactly which suppliers enter or leave the shortlist under each.

The language model experiments were equally revealing. In intent formulation, the model correctly identified the scenario in 146 of 150 procurement requests and made the right proceed-or-clarify decision in all but one, with constraint and evidence-request extraction achieving F1 scores of 0.990 and 0.996 respectively. Evidence extraction proved robust: across 40 supplier evidence packets containing 160 expected claims, the module achieved a mean claim F1 of 0.928 and grounded 99.5 percent of its quotations in the actual source text, with unsupported quotations appearing in only 0.5 percent of cases. Yet the study is candid about weaknesses. Temporal-status accuracy fell to 0.834 and evidence-role accuracy to 0.748, meaning a claim can be quoted perfectly while still being assigned the wrong analytical meaning. Under conflicting evidence, performance degraded further, with hallucination rising to 4.3 percent.

Perhaps the most instructive experiment examined what happens when extracted qualitative evidence feeds into a deterministic ranking. In a controlled benchmark combining profit, lead time, and ESG evidence, all six material compliance violations were detected and eligibility accuracy was perfect. However, all twelve suppliers with mixed sustainability evidence, those with an emissions-reduction plan whose achieved reductions remained unverified, were mapped to an overly favourable certified state. The consequence was subtle: shortlists remained largely intact, with top-three overlap between 0.889 and 1.000, but the first-ranked supplier changed in many cases. The lesson is clear and important. High grounding and perfect violation detection can coexist with incorrect semantic qualification of evidence, which is precisely why the framework routes consequential evidence through human review before it can affect eligibility or scores.

Explanation fidelity completed the picture. In 24 controlled cases, the explanation module preserved the exact deterministic supplier order, used only valid supplier and evidence identifiers, and correctly escalated every case involving constraints, close score margins, negative evidence, or missing data. Its main weakness was over-inclusiveness: explanations sometimes mentioned all available feature groups rather than isolating those that genuinely drove the ranking. The authors treat this as a review item rather than a failure, presenting the narrative alongside the fixed score tables so the decision owner can inspect both.

The broader significance of this work lies in its refusal to choose between automation and augmentation. Rather than asking whether language models can replace procurement professionals, the framework specifies exactly which tasks each participant may perform and where approval is required. It also supports adaptive supplier reassessment: when market conditions, regulations, or supplier circumstances change, a revised scenario passes through the same validated, approved, deterministic pathway rather than someone quietly editing an existing ranking. The authors acknowledge that organisational deployment, including review effort and decision-owner behaviour, still requires field evaluation, and that benchmarking across multiple model families remains future work. But as a blueprint for bringing generative AI into high-stakes sourcing decisions without surrendering accountability, the study offers something rare: a complete, tested architecture in which the flexibility of language meets the rigour of mathematics, and where the final word always belongs to a human being.

Subject of Research: A governed human-in-the-loop framework integrating large language models with deterministic multi-criteria supplier evaluation

Article Title: A governed framework combining large language models and human oversight for supplier selection

Article References: Rezaei, M., Iyer, S. S., & Fatahi Valilai, O. (2026). A governed framework combining large language models and human oversight for supplier selection. Machine Learning with Applications, 26, Article 101028. https://doi.org/10.1016/j.mlwa.2026.101028

Image Credits: AI Generated

DOI: 10.1016/j.mlwa.2026.101028

Keywords: large language models, supplier selection, procurement, multi-criteria decision-making, TOPSIS, human-in-the-loop, supply chain management, explainable AI, evidence extraction, AI governance, sustainable sourcing, auditability

The starting point of the paper is a familiar tension in modern policing. Public-safety agencies already juggle fragmented, time-sensitive information streams—radio calls, video feeds, sensor alerts, crowd movement, and human reports—yet most of the technology available to them remains retrospective. Footage is reviewed after an incident, logs are examined once a response has been triggered, and pattern analysis is applied only after historical trends have formed. Recent surveys of agentic artificial intelligence show the field moving toward adaptive reasoning over complex goals, and affective-computing research has matured in estimating probabilistic signals related to arousal, stress, vocal intensity, gesture, and group movement. Drones add mobility, altitude, and rapid redeployment, making them attractive sensing platforms. Together, these capabilities make anticipatory public-safety support technically conceivable.

But the authors argue that the shift from detecting incidents after escalation to sensing conditions that may precede escalation cannot be judged by technical performance alone. A raised voice, a compressed crowd, a sudden gesture, or a cluster of bystanders can accompany conflict, yet the same cues may reflect celebration, unrelated stress, disability-related behavior, cultural expressiveness, or simply the ordinary density of urban life. When an aerial system interprets such cues, the error is not confined to a model output: it may redirect police attention, change how officers approach a scene, or make members of the public feel watched, classified, and pre-judged. Because false positives can disproportionately burden communities already subject to heavier surveillance, the paper treats governance as an architectural requirement rather than an afterthought.

At the heart of SkySentience is the Pre-Incident Escalation Index, or PEI, an interpretable advisory score that aggregates four bounded, time-indexed channels: visual agitation indicators such as abrupt gestures and repeated encroachment into personal space; non-lexical acoustic markers such as rising vocal energy and speech overlap, with no attempt to recognize words, intent, or identity; crowd-dynamics measures such as localized compression, slowing or reversing pedestrian flow, and clustering; and a contextual prior built only from documented, non-individualized variables such as venue type, scheduled events, and time-dependent crowding baselines. The index is computed as a weighted linear combination, PEI equals alpha times the visual score plus beta times the acoustic score plus gamma times the crowd score plus delta times the contextual prior. The weights are deliberately left as calibration targets rather than universal constants, because their relative informativeness varies across venues and because the weighting choice is itself a governance decision that should be documented, version-controlled, and reviewed like any other policy parameter.

Because escalation is dynamic, the framework also tracks the short-term change in the index, delta PEI, so that a moderate score rising quickly can attract more attention than a higher but stable score in a benign setting. Every modality must also emit a quality or confidence value reflecting sensor conditions and model reliability, and raw scores should be calibrated on held-out, deployment-relevant data using auditable methods such as isotonic regression or Platt scaling, with expected calibration error and reliability diagrams reported. Missingness is handled conservatively: a degraded channel is never silently imputed, and as a reference rule at least two independent sensing modalities of acceptable quality must be available before any PEI-based advisory is produced; otherwise the system abstains and reports insufficient evidence. The contextual prior does not count as an independent sensor modality, and protected characteristics, inferred identity, and raw historical enforcement frequencies are excluded from the prior because they could reintroduce historical policing patterns into an ostensibly behavior-based score.

To make the design concrete, the paper walks through a hypothetical observation at a busy transit hub in which the visual channel reports 0.60, the acoustic channel 0.48, crowd dynamics 0.70, and the contextual prior 0.20. With documented weights of 0.30, 0.25, 0.30, and 0.15, the PEI works out to 0.54, and against a previous value of 0.43 the trend is plus 0.11. Weighted by quality values from each channel, the aggregate evidence reliability comes to 0.79—a figure the authors stress is an indicator of evidence quality, not a 79 percent probability that escalation will occur. The system would then report the triplet of score, trend, and confidence, and, if validated thresholds were met, issue a plain-language advisory such as a note that the score is moderate and increasing, driven primarily by crowd compression and visual movement cues, recommending continued observation or human review. The authors emphasize that every number in the example is illustrative, not empirical.

The decision logic built on top of the index is intentionally narrow. When the score is low and stable, the drone continues passive observation. When the score is moderate or rising quickly, the system may recommend light-touch, reversible responses such as repositioning for a clearer view or notifying an officer with a short explanation. Even when risk remains elevated and corroborated across modalities, the system prioritizes officer awareness rather than autonomous action, and it is never permitted to stop, identify, pursue, confront, or otherwise act coercively toward members of the public. Language models, if used at all, are restricted to summarizing and explaining policy-checked outputs; they do not control the drone, select enforcement actions, or generate unverified factual claims. The architecture separates perception, advisory reasoning, human authorization, and accountability into distinct layers, so that every action affecting a person’s liberty, safety, or legal exposure requires an explicit human decision.

The governance requirements are equally specific. The authors call for tamper-evident, time-stamped, access-controlled audit logs recording each PEI computation, the contributing modality scores, confidence estimates, recommendations, officer responses, and rationales for acceptance, modification, override, or inaction. They propose an institutional division of labor in which the deploying agency documents deployment criteria and configurations, an independent oversight function outside the chain of command audits logs, override patterns, and subgroup error rates, and a separate public-facing channel handles redress requests from people who believe they were observed or flagged. Proportionality ties intervention intensity to evidentiary strength, confidence, and persistence, while explainability must serve two audiences: officers need concise operational explanations of what drove a score, and the public needs to know what the system observes, what it does not infer, how data are retained or deleted, and how to challenge an assessment. The framework is also jurisdiction-contingent: where law prohibits emotion inference in policing, any channel constituting such inference must be disabled or the deployment must not proceed.

Two illustrative scenarios show how the framework behaves when it is right and when it is wrong. In the first, rising voices, prolonged eye contact, and slowing pedestrian flow at a transit platform jointly lift the index, prompting only a respectful vantage point and a notification to a nearby officer, who retains full discretion to observe, approach, or stand down. In the second, a group celebrating a cultural occasion sings loudly and clusters tightly, pushing the score across a moderate threshold even though the behavior is benign; the system may only reposition or notify, the officer recognizes the celebration and records an override, and that override feeds later recalibration and bias review. The authors argue that the difference between a harmless false positive and a harmful one lies precisely in this governance layer: without bounded autonomy, override logging, and public redress, the same score could produce unnecessary police attention or opaque records.

The paper closes with a candid reckoning. The authors acknowledge the strongest objection—that affect-inferring drone systems should not exist in public policing at all—and they distinguish the justificatory question of whether a community should permit such a system from the architectural question of what must be true of it if deployed. Safeguards, they note, can reduce harm only if a system is deployed; they do not prove deployment is justified, and that prior question belongs to legal, democratic, and community processes. A staged validation plan follows: first testing the index on public benchmark data with subgroup-disaggregated performance, then using simulation and replay studies to examine officer-facing explanations, alert burden, and override patterns, and finally evaluating whether audit records, public explanations, retention policies, and community review mechanisms actually function in practice. A system that passes the technical stage but fails the institutional one, the authors conclude, should not be considered successful—and satisfying every requirement would still not, by itself, establish that any jurisdiction should deploy affect-related drone decision support.

Subject of Research: A conceptual governance-centered framework for accountable, human-supervised UAV decision support in public safety using a Pre-Incident Escalation Index.

Article Title: The SkySentience conceptual framework for accountable drone decision support in public safety

Article References: Majumdar, S., & Awasthi, A. (2026). The SkySentience conceptual framework for accountable drone decision support in public safety. Discover Informatics, 1(1), Article 16. https://doi.org/10.1007/s44564-026-00018-x

Image Credits: AI Generated

DOI: 10.1007/s44564-026-00018-x

Keywords: SkySentience, public-safety informatics, drones, UAV surveillance, agentic AI, Pre-Incident Escalation Index, affect inference, algorithmic accountability, human-AI collaboration, auditability, responsible AI, policing technology

The core idea is deceptively simple. In conventional federated averaging, known as FedAvg, the server weights client updates largely by data volume, with little regard for whether an update genuinely improves the model for everyone else. Khan’s framework instead asks other clients to evaluate each clipped update on disjoint certificate records, producing a point-gain estimate that reflects how much the update reduces error on data the contributing client never saw. These estimates become certificates: compact, auditable artifacts that trace a verifiable route from raw verifier evidence to the influence a client ultimately exerts on the aggregated model. Because the certificates are computed from held-out records and clipped updates, they resist both accidental miscalibration and deliberate manipulation.

The evidence for the approach is organized into three empirical stages, each designed to test a different layer of the claim. Stage I retains coefficient-level benchmarks across four domains, including public high-throughput chemistry spreadsheets, published voltage-controlled magnetic anisotropy source data, the Wisconsin Diagnostic Breast Cancer dataset and the Golub leukemia transcriptomic dataset. In the high-throughput chemistry setting, the certificate-weighted method delivered a 1.52 percent reduction in root-mean-squared error relative to federated averaging, a modest but consistent signal that utility-aware weighting can extract real gains even in small, heterogeneous scientific datasets.

Stage II moved to locked, confirmatory image experiments using Fashion-MNIST and CIFAR-10 with twenty non-identically distributed clients, forty rounds of training, and ten confirmatory random seeds. The non-IID design matters because real federated deployments rarely enjoy balanced data: each client’s distribution differs in ways that can poison naive averaging. Across all four dataset-scenario cells, point-gain weighting improved the negative log-likelihood, a strict probabilistic measure of prediction quality, over standard FedAvg. The study also examined a two-phase regime in which the server first pretrains a federated model and clients then fine-tune locally for personalization. Federated pretraining followed by local fine-tuning beat Local-only prediction in every image-dataset seed and regime, reinforcing the practical case for personalization pipelines built on trustworthy aggregation.

Stage III provided the most demanding test: an independent, frozen-configuration replication on the Human Activity Recognition Using Smartphones dataset from the UCI Machine Learning Repository, in which natural subject identifiers served as client partitions. The experiment ran 110 rounds with forty new random seeds, none of which were tuned after the configuration was locked. To stress the system under realistic data-quality threats, the researchers introduced fitting-label corruption, a scenario in which the labels a client uses to fit its update no longer reflect ground truth. Under this corruption, contextual marginal weighting reduced the participating-subject negative log-likelihood from 0.2199 under FedAvg and 0.1768 under an equal-access PointGainFedAvg control to 0.1730.

Those gains were not statistical noise. The paired improvements of 0.0468 over FedAvg and 0.00375 over the PointGainFedAvg control carried Holm-adjusted p-values of 3.64 times ten to the minus twelve and 0.00251 respectively, comfortably below conventional significance thresholds even after conservative correction for multiple comparisons. Crucially, natural performance remained within prespecified noninferiority margins, meaning the certificate machinery did not sacrifice accuracy on uncorrupted data in exchange for robustness under attack. That combination, significant gains under corruption without degradation under normal conditions, is precisely the profile regulators and industry operators demand from reliability mechanisms.

Perhaps the most striking result concerns auditability itself. The marginal certificate reproduced on a disjoint audit split with a Spearman correlation of 0.980, indicating that the utility scores computed during training faithfully predict utility measured on entirely separate records. Furthermore, as the researchers increased the level of label corruption, the certificate value declined even after controlling for certificate size and entropy, demonstrating that the score tracks genuine update quality rather than superficial statistics of the update. In other words, the certificate behaves like an instrument: it responds to the signal it claims to measure, and it can be re-verified by an independent auditor using data the original evaluation never touched.

The implications extend well beyond benchmark datasets. Federated learning now underpins multi-institutional medical imaging collaborations, fraud detection in electronic payment streams, condition monitoring in the Industrial Internet of Things, and privacy-preserving analytics across smartphone fleets. In each of these settings, a single corrupted or adversarial client can silently degrade a shared model, and existing Byzantine-robust aggregation schemes typically reject outliers without explaining why. Update-utility certificates invert that logic: instead of merely filtering bad actors, they generate positive, auditable evidence of value, giving every participant a defensible account of why each client’s contribution was weighted as it was. For regulated industries, that audit trail could prove as important as the accuracy gains themselves, since it aligns federated aggregation with emerging accountability and data-governance requirements.

The study also situates itself within the broader trajectory of personalized federated learning research. Prior approaches, from SCAFFOLD’s controlled averaging to FedBN’s local batch normalization and variational Bayesian personalization, have attacked the heterogeneity problem from the optimization and modeling sides. Khan’s contribution addresses the trust side, supplying the reliability metadata that personalized pipelines need before they can safely aggregate. The work draws on a long lineage of multisensor data fusion, robust statistics and differential privacy, and it deliberately uses only public secondary datasets, from chemistry reaction yields to smartphone accelerometer readings, analyzed in de-identified form. The author reports no competing interests and no specific funding, and the experiments relied exclusively on data whose creators are publicly acknowledged.

Taken together, the three stages establish a replicated pattern rather than a one-off result: point-gain improvements confirmed across seeds and datasets, an aggregate-context benefit on natural client partitions under label-quality degradation, and an auditable route from verifier evidence to client influence. As federated deployments scale into domains where a wrong prediction carries clinical, financial or safety consequences, mechanisms that make the value of every update verifiable may shift from a research curiosity to a baseline requirement. The certificate framework suggests that the next generation of federated systems will not merely average their clients’ work; they will be able to prove, record by record, why that averaging was justified.

Subject of Research: Auditable update-utility certificates for improving aggregation and accountability in personalized federated learning

Article Title: Auditable update-utility certificates for personalized federated learning

Article References: Khan, K. (2026). Auditable update-utility certificates for personalized federated learning. International Journal of Data Science and Analytics, 22(1), Article 295. https://doi.org/10.1007/s41060-026-01264-w

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01264-w

Keywords: federated learning, personalized federated learning, update utility, reliability certificates, auditability, FedAvg, label corruption, negative log-likelihood, human activity recognition, data privacy, robust aggregation, machine learning