Minxuan Hu

Streaming monitoring systems are deployed as alerting systems, not as score generators alone. In these settings, evaluation must answer not only whether incidents receive high scores, but whether the realized alert stream changes under corruption, responds in time to matter, and remains workload-safe under nonstationarity. We present a reproducible, deployment-oriented framework for evaluating streaming alerting systems that treats the realized score-to-warning-to-episode pipeline as the evaluated object. The framework combines paired clean/corrupted replay for counterfactual alert impact, matched-burden event timeliness, explicit feasible-versus-infeasible operating regions under episode-budget and time-in-warning (TIW) constraints, and regime-sliced compliance audits. For adaptive policies, the framework preserves counterfactual parity by learning the clean-replay threshold schedule once and replaying it unchanged on corrupted streams at identical timestamps. Across a primary financial-services monitoring panel and two telemetry replications, feasibility emerges as a first-class result: operating points that appear attractive under score-centric or unconstrained views can become non-deployable once workload constraints are enforced. In the primary Consumer Financial Protection Bureau (CFPB) setting, median budget-only matched-burden coverage is 0.844, but median joint coverage after enforcing TIW falls to 0.025, showing that TIW is the binding safety constraint. Online recalibration preserves the narrow surviving feasible region that static calibration removes. Results on NASA SMAP A-1 and D-15 further show that although the best detector family varies by panel, framework-level conclusions transport across domains.

The deployment of autonomous AI agents in derivatives markets has widened a practical gap between static model calibration and realized hedging outcomes. We introduce two reinforcement learning frameworks, a novel Replication Learning of Option Pricing (RLOP) approach and an adaptive extension of Q-learner in Black-Scholes (QLBS), that prioritize shortfall probability and align learning objectives with downside sensitive hedging. Using listed SPY and XOP options, we evaluate models using realized path delta hedging outcome distributions, shortfall probability, and tail risk measures such as Expected Shortfall. Empirically, RLOP reduces shortfall frequency in most slices and shows the clearest tail-risk improvements in stress, while implied volatility fit often favors parametric models yet poorly predicts after-cost hedging performance. This friction-aware RL framework supports a practical approach to autonomous derivatives risk management as AI-augmented trading systems scale.

Learned spatial 3×3 convolutions dominate vision backbones, but their effective operators are implicit, and common proxies (parameter count, multiply-adds) can mispredict realized throughput and energy under modern compilation. We introduce ChebGate, a drop-in spatial block that makes the operator explicit by freezing a depthwise local operator and generating higher-order interactions through a gated Chebyshev recurrence; a lightweight controller outputs per-sample coefficients over polynomial orders, followed by learned 1×1 channel mixing. Instantiated as ChebResNet in a Wide-ResNet-style backbone, ChebGate exposes order allocation that enables mechanism analysis and gate-only interventions without changing weights. On CIFAR-10, clustering late-stage allocation into a compact program codebook enables substitutions that improve calibration with negligible accuracy loss (Top-1 97.02%→96.95%; expected calibration error (ECE) 0.1080→0.0094); matching logit norms largely removes the gain, indicating temperature-like logit-scale control. Separately, benchmarking four numerically equivalent realizations under strict FP32 checks shows throughput varies 0.47×–1.31× despite identical parameters and multiply-adds, with corresponding energy shifts under a fixed stack. ChebGate demonstrates that operator-explicit modules with program diagnostics and equivalence-checked benchmarking support both mechanistic interpretability and deployment accountability.