A Semantic Integrity Risk Score is proposed: a learned, generator-agnostic measure of the degree to which a record violates the functional and logical dependencies mined from trusted data, which is positioned as a practical, proactive complement to statistical monitoring rather than a replacement for it.
Abstract
The generation of synthetic tabular data by unconstrained generative models has become a routine component of modern, data-centric machine-learning pipelines. Suéch generators reproduce the statistical structure of real data, yet statistical fidelity does not guarantee logical validity. A generator may emit records that are plausible with respect to every marginal and correlation of the training data while violating the inter-attribute dependencies that make a record meaningful—for instance, an individual simultaneously never-married and a husband. This failure mode, termed here semantic poisoning, is shown to pass undetected through the distributional monitoring on which pipelines ordinarily rely. Three contributions are made. First, a Semantic Integrity Risk Score (SIRS) is proposed: a learned, generator-agnostic measure of the degree to which a record violates the functional and logical dependencies mined from trusted data. Second, a Semantic Firewall applies SIRS at the point of data ingress to route each record to a pass, review, or quarantine decision. Third, an empirical study on real census data across three structurally distinct generators demonstrates that the proposed signal recovers injected semantic violations with an F1 of 0.86–0.95, whereas marginal drift detection remains near zero; that it generalizes to violations of dependencies withheld from any explicit rule set; and that it prevents downstream harm which aggregate-accuracy monitoring cannot perceive, at a bounded screening cost of approximately thirty milliseconds per thousand records. Semantic screening is thereby positioned as a practical, proactive complement to statistical monitoring rather than a replacement for it.
This work shows that a malicious developer can pair a benign-looking wrapper with crafted metadata to deterministically alter post-generation behavior without modifying model weights, training data, or inference backend, and introduces TIF-BAH, a lightweight middleware defense that verifies wrapper integrity and records behavioral attestations during inference.
Nokimul Hasan Arif, Qian Lou, Meng Zheng· 0 citations
A reproducible, license-aware knowledge-distillation recipe addressing the constraint of deploying a safety layer for large language models on commodity hardware by partitioning the corpus into seven safety categories aligned to a public hazard taxonomy.
Edson Rodrigues da Cruz Filho, Paulo Ricardo Ferreira Neves, P. H. Falsetti et al.· 0 citations
This work provides a mathematically grounded, highly efficient diagnostic tool to uncover human label failures, sanitize evaluation benchmarks, and ensure the integrity of LLM alignment data.
Yunting Song, Matthew Watson, Peter Grabowski et al.· 0 citations
Agentic systems act, so a defect in the evidence they retrieve becomes a wrong action with a currency cost. The most dangerous enterprise defects are metadata-borne: a stale price or a superseded record, perfectly well-formed in the payload and betrayed only by freshness, lineage, or provenance. Such a defect never enters the agent's context, and an agent cannot doubt data it cannot see. On a priced replenishment benchmark, a competent agent silently converts an injected metadata-borne defect into a costly action about 60% of the time, with zero data-quality flags and behavioral doubt markers at chance (AUC<= 0.50). Across four model tiers spanning roughly 15x in inference price, the rate stays flat: capability does not buy skepticism. A metadata-aware pre-action gate with downstream-only remediation recovers the loss fully on the signals its predicates cover and not at all on those they miss. A model-free oracle derived from the task's decision geometry tracks the measured rates with MAE 0.015 (Pearson r = 0.876, interval coverage 15/16 cells), giving the flat ladder an analytical form. Evidence integrity is a systems axis distinct from model capability; mitigation depends on enforcement placement and predicate coverage. Code, frozen results, and a deterministic analysis pipeline: https://github.com/besanson/dqSarc
This work proposes treating long-term agent memory as a distinct data model -- with its own write semantics (encoding, separation, consolidation, provenance) and read semantics (cue-driven activation across a linked memory graph) and presents FluctlightDB, an embedded engine that implements this contract via experience() and activate().
Enterprise data platforms lose substantial value to poor data quality, and database migrations routinely overrun, because integrity validation and query translation are treated as disconnected problems. We present DataFlow AI, a deployment-oriented framework coupling a stacked anomaly-detection ensemble with an audit-grade, cross-dialect SQL transpiler, evaluating both halves under a single reproducibility contract. The detector fuses four heterogeneous base signals—rule-based constraints, robust per-group statistics, an Isolation Forest, and a Local Outlier Factor—through a meta-learner trained on five-fold out-of-fold predictions, replacing fixed-weight late fusion. Treating the meta-learner as an experimental variable, we find a gradient-boosted stacker exceeds a linear one by up to 0.163 mean $F_{1}$ across ten seeds under a nested threshold (Holm-corrected $p\lt 0.01$ ), because agreement among base detectors carries conditional structure no linear combiner can express. We evaluate end-to-end on three real public corpora—SEC EDGAR financial-statement values, the New York City FY2024 payroll, and the UCI credit-default corpus—injecting fifteen reproducible anomaly families into each at a realistic prevalence near five percent. The stacked ensemble attains an $F_{1}$ of 0.601 on SEC, 0.572 on payroll, and 0.761 on credit default, outperforming ECOD, COPOD, HBOS, KNN, LODA, and Deep SVDD on every corpus—a margin owed to domain-specific rule signals those unsupervised baselines cannot use, not to superior learning machinery—and we report per-family recall so the two families it essentially fails to detect are visible rather than averaged away. For migration, we benchmark transpilation on 115 curated queries across five dialects, giving 575 source–target pairs. Parse and transpile success both exceed 0.99, while a stricter abstract-syntax-tree (AST) footprint-equivalence test yields 0.807 overall, degrading from 0.921 on easy queries to 0.738 on medium and 0.742 on hard. That gap is the substantive finding: surface translation is solved by mature parsers, and a construct-level analysis attributes the residual loss to LATERAL, PIVOT, JSON, and window features—though every executable pair reproduces the source result set exactly, so structural and semantic drift must not be conflated. Finally, we test the framework’s own premise against ground-truth, corrupted, and gated tables: quarantine by deletion degrades downstream accuracy sixfold and biases aggregates over columns it never inspected; quarantine-then-repair improves it fivefold. The contribution is therefore not a new accuracy record but a joint, deployment-aware evaluation framework for data-quality validation and SQL migration that measures their downstream interaction: where coupling helps, where it is inert, and where its most natural implementation does harm. Because the anomalies are injected rather than field-discovered, the reported $F_{1}$ values are benchmark behaviour, not field performance. All code, anomaly masks, and SHA-256-verified datasets reproduce end-to-end from fixed seeds.
Upasana Bhaumik, M. Berlin· IEEE Access· 0 citations
A new method, called CW-Net, translates the reasoning process of an autonomous vehicle’s AI system into understandable concepts that explain its behavior.