fallback strategy design

This work addresses a critical limitation of existing conformal prediction methods, which guarantee only marginal coverage and fail to characterize key operational metrics—such as decision frequency, error exposure, and rejection rate—and their inherent trade-offs in real-world deployment. To overcome this, the authors propose an operational certification framework that goes beyond coverage by introducing a calibration-audit two-stage mechanism to quantify and guarantee the statistical properties of system behavior under finite-sample settings. Key innovations include Small-Sample Beta Correction (SSBC) for finite-sample coverage guarantees, reusable confidence envelopes for operational metrics, and the revelation of geometric couplings and trade-off boundaries among these metrics under conformal partitioning. The framework successfully generates auditable operational configuration menus on Tox21 and AquaSolDB benchmarks, explicitly delineating performance boundaries and uncertainties across different calibration strategies.

In high-stakes environments, agent exploration is often terminated due to violations of safety constraints, while excessive conservatism hinders performance gains. This work proposes a conformal calibration method grounded in a safe reference policy that dynamically modulates the aggressiveness of a new policy according to a user-specified risk tolerance. Without requiring assumptions about the correctness of the model class or fine-tuned hyperparameters, the approach provides, for the first time, finite-sample safety guarantees for non-monotonic bounded constraint functions. By overcoming the limitations of traditional conservative optimization and existing conformal methods, it enables “safe-on-deployment” exploration while simultaneously improving policy performance, as demonstrated in tasks ranging from natural language question answering to biomolecular engineering.

This study addresses generalization risks and governance challenges in trustworthy AI by proposing a novel conformal prediction paradigm tailored for AI governance. Unlike conventional approaches that solely ensure statistical calibration, our method extends conformal prediction to bias identification and fairness governance, establishing an interpretable and auditable set-valued prediction framework. Specifically, we generate marginally valid prediction sets via calibrated data and integrate bias detection and mitigation mechanisms. Theoretically, our framework guarantees statistical validity of prediction sets under standard assumptions. Empirically, it significantly improves uncertainty quantification calibration across multiple tasks and datasets, robustly detects and mitigates model bias, and enhances AI system trustworthiness and accountability in high-stakes applications—such as risk assessment and regulatory compliance—thereby supporting socially responsible AI deployment.

To address inference failure in high-risk real-time fraud detection under open banking—caused by overreliance on external data—this paper proposes the first hierarchical fallback paradigm tailored for financial real-time machine learning inference. Our method embeds fault tolerance across the entire inference pipeline via (i) fine-grained fault classification, (ii) multi-level policy routing, (iii) lightweight backup model ensembles, and (iv) coordinated scheduling with Open Banking APIs, enabling dynamic degradation and semantically consistent fallback decisions. Evaluated on a production anti-fraud system, it achieves 99.98% service availability under extreme load, reduces inference latency volatility by 76%, and maintains ≥92% consistency in risk assessment. This work is the first to systematically define, implement, and empirically validate a layered fallback architecture for financial real-time ML inference, establishing a scalable, fault-tolerant framework for high-reliability online machine learning systems.

This work addresses a critical limitation in conventional execution-signal-based governance mechanisms, which can only assess actions locally and thus fail to detect runtime deviations from the global behavior space \( \mathcal{A}_0 \) established during admission. To overcome this, we propose the Invariant Monitoring Layer (IML), which leverages direct access to the generative model of \( \mathcal{A}_0 \) to achieve bounded-delay detection of behavioral drift within structural blind spots of existing enforcement mechanisms. We first establish an information-theoretic non-identifiability theorem, proving the fundamental impossibility of such monitoring under prior approaches, then introduce the IML architecture—integrating formal verification, σ-algebraic analysis, and trajectory-level behavioral modeling—to circumvent this barrier and provide provable upper bounds on detection latency. Empirical evaluation across four scenarios, including n8n webhook pipelines and LangGraph StateGraph agents, demonstrates that while traditional mechanisms issued no alerts, IML successfully detected all drift types within 9–258 steps.

This work addresses the challenge of conformal change-point localization and root cause analysis under contaminated observational data—such as outliers, sensor failures, or adversarial perturbations—by proposing weighted variants of CONCH and CROC (W-CONCH/W-CROC). These methods incorporate a weighting mechanism grounded in second-order uncertainty derived from Huber contamination models, evidential deep learning, or Bayesian inference, downweighting suspicious observations to significantly shrink prediction set sizes while preserving user-specified coverage guarantees. To the best of our knowledge, this is the first approach to integrate uncertainty-driven weighting into conformal change-point detection and root cause attribution. The weights are optimized via meta-learning with a differentiable surrogate objective, enabling applicability in nonparametric and contaminated settings. Experiments demonstrate consistent improvements in localization accuracy and practical utility across image and real-world benchmarks.

This work addresses the susceptibility of large language models to hallucinations in empirical reasoning—outputs lacking verifiable evidence or formal guarantees. The authors propose EG-VAR, an architecture that uniquely leverages the Lean 4 formal proof kernel as the sole trusted generator of claims. By integrating tool-certified axioms and a source elevation mechanism, EG-VAR ensures every output is bound to a kernel-verified chain of reasoning and tool invocation; otherwise, it abstains and provides a fully traceable audit trail. Evaluated on a TableBench subset, EG-VAR achieves perfect accuracy (120/120), substantially outperforming a 95% baseline. In counterfactual tests, it maintains 100% source fidelity—significantly higher than competing methods (80–90%)—and exhibits remarkably low semantic formalization error rates of 1.7% (Opus) and 3.3% (Sonnet).

This study addresses the challenge of providing certifiable runtime safety guarantees prior to tool invocation, focusing on three core issues: the representability of policy states, the observability of monitoring evidence, and the impact of interventions on future behavior. To this end, we propose the first formal theoretical framework for runtime safety-executable boundaries, distinguishing among static policy executability, statistical calibration under exogenous legal constraints, and closed-loop intervention effects. Building upon finitely controlled models, we develop a method for closed-loop safety certification that integrates register model identification, Neyman–Pearson hypothesis testing, conformal calibration, and occupancy planning. Empirical validation through static diagnosis, model enumeration, representation rewriting, and closed-loop re-execution experiments demonstrates the efficacy of our approach and exposes the fundamental limitations of static calibration under representation attacks.

In multi-institutional medical federated learning, standard conformal risk control (CRC) struggles to guarantee individual coverage for underperforming hospitals, leading to uncontrolled false negative rates. This work proposes a federated CRC protocol based on risk curve shrinkage: each participant uploads only its empirical risk curve, and the server applies shrinkage regularization to generate personalized thresholds for every hospital, ensuring both coverage validity and prediction set efficiency without exchanging any patient-level data. The study quantifies, for the first time on real-world multicenter brain tumor data (FeTS-2022), the coverage failure of federated CRC and introduces a hyperparameter to smoothly balance worst-case coverage–efficiency trade-offs. Experiments show the method reduces the proportion of institutions violating coverage guarantees from 40% to 13.5%, with only a 2.0× increase in prediction set size—dramatically lower than the 83× inflation observed with local baselines.

This work addresses the limitations of existing SMT-based program verification tools, which suffer from insufficient expressiveness and low solver trustworthiness. To overcome these challenges, the authors propose FLEX—the first end-to-end foundational Constrained Horn Clause (CHC) solver implemented entirely within Lean. FLEX encodes CHCs as propositions verifiable by Lean’s trusted kernel and integrates metaprogramming tactics with Floyd-Hoare semantics to enable composable strategies for verification condition generation and solving. By leveraging Lean’s expressive logic, this approach transcends the representational constraints of SMT solvers and supports correctness proofs for arbitrary functions within Lean’s rich ecosystem. Evaluated on the FLUX benchmark, FLEX automatically solves 95.7% of CHC problems and successfully verifies multiple low-level Rust libraries.