We study retrieval over catalogs of structured metadata, where each record is a small schema whose fields answer different kinds of query. Embedding a record with a text encoder first serializes its fields into a string, which forces a choice of field order. We show this choice silently controls retrieval quality once the encoder is fine-tuned. A standard fine-tune loses 7.4 nDCG@10 points when the index is rebuilt under a different field order, because it reads absolute position instead of the field labels. We propose permutationinvariant fine-tuning (PI-FT), which serializes each training record under a freshly sampled field order with random field dropout, so meaning binds to the labels rather than to position. The change is about two lines in the data loader; it costs negligible in-distribution accuracy and reduces the order-change penalty to 0.2 points. We study this in the discovery of development statistics, a catalog of nearly 10,000 indicators that should be searchable in many languages by a model small enough to self-host. DEVDATABENCH is a fully LLM-generated benchmark of grounded, facet-targeted queries across 15 languages, covering every indicator for both training and evaluation. A finetuned 118M-parameter encoder, runnable on a CPU, outperforms every zero-shot baseline including text-embedding-3-large (0.707 vs. 0.556 nDCG@10), with the largest gains on the low-resource languages those users speak. We release the benchmark, pipeline, and models, and the training framework applies to any labeled-field catalog. [page::0][page::1]
CLQT reframes closed-loop trading evaluation as a diagnostic instrument rather than a simple leaderboard, introducing a five-stage loop (gather, synthesize, construct a mandate-aware portfolio, execute with cost discipline, reflect) and a tamper-evident audit trail that enables reconstruction of every metric from the decision trail. It yields a five-axis capability scorecard (APM-CS): Coherence, Acuity, Composure, Discipline, Reliability, calibrated against held-out judges and ablations, and validated through backtest and a live broker track. The contribution is a durable map of agent competencies and limitations—across both backtest and live—rather than a single, once-off ranking. This diagnostic framework is designed to scale with more models, regimes, and iterations, aiming to inform growth and improvement of AI-driven portfolio-management agents [page::0][page::1].
This paper develops step-weighted distribution and demonstrates how to construct comonotonic and moment-matching approximations for sums of lognormal random variables. It shows that the proposed approximations achieve performance comparable to classical comonotonic methods in most cases, while offering improved accuracy in the right tail of the distribution, which is crucial for risk measures. A new step-weighting theory is introduced, including definitions, properties, and risk measures (VaR, TVaR, stop-loss) for step-weighted distributions, with further results on their behavior under comonotonic summation. The core contribution is a framework that yields comonotonic sums that match moments up to a given order, enabling tractable computation of distortion risk measures and tail risks for sums of continuous random variables, including sums of lognormals [page::1]
This paper develops a path-first theory using signatures as universal coordinates for deterministic paths, rough paths, jumps, and path-valued random variables. It introduces the geometricity defect principle, characterizes Itô vs Marcus lifts, and delivers finite-dimensional Hawkes closures for affine and scalar cases, along with a directional cross-area statistic. The results link shuffle, Marcus–Itô, and signature kernels to the geometry of the free nilpotent group, and establish identifiability and large-deviation implications for path-law representations. [page::0]
<paper_abstract>This paper studies how a risk holder should combine self-protection and self-insurance when market insurance is absent. In a Bernoulli loss model, self-protection reduces the residual loss probability, while self-insurance reduces the residual loss severity. The risk holder evaluates residual risk using either Value-at-Risk or Tail Value-at-Risk and incurs a joint risk-reduction cost that allows technological interaction between the two activities. We show that Value-at-Risk leads to a threshold-driven solution that the optimal strategy is either no risk reduction, pure self-protection, or pure self-insurance. By contrast, Tail Value-at-Risk creates a direct interaction between residual frequency and residual severity, making the problem non-convex even in the Bernoulli setting. We solve it using an isoquant geometry method based on the marginal-balance curves for self-protection and self-insurance. The analysis identifies when optimal strategies lie on boundaries, extreme constrained candidates, touching components, or crossing components, and shows how the confidence level and the cost technology determine whether self-protection and self-insurance behave as substitutes or complements. [page::1][page::4][page::12][page::18]
本研究将 AI 治理问题从单纯预测转向对权威分配的动态优化，提出一个以贝叶斯为基础的治理感知POMDP框架，用以在证据质量、不确定性和机构目标演变时，持续动态调整 AI 推荐的决策权限。通过合成压力测试、预测精度验证和历史市场回放，验证该治理策略在不同 AI 质量情境下的鲁棒性、可解释性与可操作性，且在与五种基准治理机制对比中展现出更强的一般化能力和稳健性。 [page::0]
